a framework for domain-driven development of personal health informatics technologies

A FRAMEWORK FORDOMAIN-DRIVEN DEVELOPMENT OF

PERSONAL HEALTH INFORMATICSTECHNOLOGIES

A Dissertation

Presented to the Faculty of the Graduate School

of Cornell University

in Partial Fulfillment of the Requirements for the Degree of

Doctor of Philosophy

by

Elizabeth Lindley Murnane

January 2017

A FRAMEWORK FOR

DOMAIN-DRIVEN DEVELOPMENT OF

PERSONAL HEALTH INFORMATICS TECHNOLOGIES

Elizabeth Lindley Murnane, Ph.D.

Cornell University 2017

This dissertation advances a vision of Personal Health Informatics (PHI),

a class of tools that can leverage personal data to support health self-

management. Today, a powerful combination of factors is coming together

that can facilitate the creation of these technologies and amplify their bene-

fits. Namely, the world is awash in data, software and sensors continue to cap-

ture more, increasingly capable algorithms are helping humans make sense of

it all, and ubiquitous devices (that people are keen to use to manage their well-

ness) can deliver this information via individually-tailored, insight-enabling,

personally-empowering, health-enhancing feedback.

A central argument of this dissertation is that domain knowledge can help

drive PHI development in order to fully capitalize on the potential of these tech-

nologies. A central contribution of this dissertation is a framework for engaging

in domain-driven development. In specifying this reusable development pat-

tern, I provide guidance on moving through stages of domain inquiry, domain-

driven health assessment, and domain-aware intervention design.

To begin, I describe what domain knowledge encompasses, why it is valu-

able, and how to synthesize insights from diverse sources in order to gain an ap-

preciation of the role technology can play in a given context. I then explain how

this understanding can inform research goals, strategies for assessing significant

health determinants, and implications for designing effective interventions.

To demonstrate this process in practice, I present my own research as a case

study on developing domain-driven technology that supports healthy sleep,

daily performance, and emotional wellbeing. Overall, I argue that a domain-

driven approach that foregrounds a deep understanding of a targeted aspect

of health, together with a compassion for the lived experiences of users, will

produce technological solutions that better meet individual needs and promote

more positive outcomes.

BIOGRAPHICAL SKETCH

Broadly speaking, Elizabeth’s research interests lie in human-computer interac-

tion, personal informatics, recommender systems, social computing, and per-

sonalization. She is particularly compelled by applications in the domains of

personal information management, civic innovation, and health.

Elizabeth received her Bachelor of Science in Mathematics with Computer

Science in 2007 from the Massachusetts Institute of Technology (MIT), where

her undergraduate research in the Computer Science and Artificial Intelligence

Laboratory (CSAIL) focused on information visualization and conversational

agents. Upon graduating, Elizabeth co-founded and spent four years as the

lead engineer of Architexa, a CSAIL spinoff that built interactive visualization

tools to help software developers make sense of and share important aspects of

source code.

Since 2011, Elizabeth has been a PhD student in Information Science at Cor-

nell University. In her graduate studies, she has continued to explore her over-

arching research goal: designing and deploying technologies that are aware of

idiosyncratic user needs and that support people in managing various aspects

of their daily lives. Much of Elizabeth’s current work (out of which this disser-

tation grew) is on developing novel mediums for manually collecting personal

data, lightweight algorithms for passive sensing and user modeling, and inter-

faces that provide tailored feedback and experiences.

In pursuing these directions, Elizabeth collaborates with interdisciplinary

teams comprised of domain experts from critical theorists to legal scholars

to clinicians. Through these partnerships she hopes to continue pushing the

boundaries of how systems can enable more positive interactions with and

through technology, on individual, group, and societal levels.

iii

ACKNOWLEDGEMENTS

I would like to extend a huge thanks to everyone who supported me throughout

the course of my PhD and helped to make this dissertation happen.

I am especially appreciative of my fantastic committee who saw me through.

Dan Cosley has embodied what it means to be a true mentor, Geri Gay has been

an incredible advocate and often felt like a co-advisor, and Claire Cardie has

been a valuable source of interdisciplinary perspective.

Humbly, I realize that it would not be possible to acknowledge by name

every other individual who has had a positive impact on me over the years; so

though I give broad thanks, it is heartfelt. I am tremendously grateful for my

family as well as friends, colleagues, and collaborators from Cornell University

and beyond — I look forward to our continued connection. I am also indebted

to all of the open-minded participants willing to be involved in my research and

all the known or anonymous reviewers that helped shape its presentation.

Finally, I feel very fortunate to have received a Graduate Research Fellow-

ship from the National Science Foundation, whose funding afforded me the

flexibility to undertake a number of diverse projects, broaden my horizons, and

transform ideas into reality.

iv

TABLE OF CONTENTS

Biographical Sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

1 Introduction 11.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Dissertation Overview & Contributions . . . . . . . . . . . . . . . 7

2 Background 102.1 The Roles of Technology in Supporting Health . . . . . . . . . . . 10

2.1.1 The Roots of Modern Medicine . . . . . . . . . . . . . . . . 112.1.2 The Age of Behavior Change . . . . . . . . . . . . . . . . . 142.1.3 The Individual as the Nexus of Health Management . . . . 19

2.2 Personal Health Informatics (PHI) . . . . . . . . . . . . . . . . . . 212.2.1 Capturing Personal Data . . . . . . . . . . . . . . . . . . . . 252.2.2 Analyzing Data to Assess Health . . . . . . . . . . . . . . . 322.2.3 Delivering User-Facing Feedback . . . . . . . . . . . . . . . 34

2.3 Domain-Driven Personal Health Informatics . . . . . . . . . . . . 482.3.1 Defining Domain Knowledge . . . . . . . . . . . . . . . . . 482.3.2 The Value of Domain-Driven PHI . . . . . . . . . . . . . . 492.3.3 Examples of Domain-Driven PHI . . . . . . . . . . . . . . . 512.3.4 A Domain–Practice Gap in HCI . . . . . . . . . . . . . . . . 53

2.4 A Framework for Domain-Driven PHI Development . . . . . . . 57

3 Domain Inquiry 623.1 An Overview of the Inquiry Process . . . . . . . . . . . . . . . . . 633.2 Selecting an Application Area . . . . . . . . . . . . . . . . . . . . . 70

3.2.1 Sleep, Cognitive Performance, and Emotional Wellness . . 703.2.2 Opportunities for Technology . . . . . . . . . . . . . . . . . 743.2.3 Related Academic and Commercial Work . . . . . . . . . . 753.2.4 Connecting with Chronobiology . . . . . . . . . . . . . . . 81

3.3 Gathering Domain Knowledge . . . . . . . . . . . . . . . . . . . . 843.3.1 Chronobiology & Circadian Rhythms . . . . . . . . . . . . 853.3.2 Chronotype . . . . . . . . . . . . . . . . . . . . . . . . . . . 873.3.3 Circadian Disruption . . . . . . . . . . . . . . . . . . . . . . 883.3.4 Traditional Assessment Methods from Chronobiology . . 91

3.4 Informing PHI Development . . . . . . . . . . . . . . . . . . . . . 943.4.1 Defining Scope . . . . . . . . . . . . . . . . . . . . . . . . . 943.4.2 Soft Sensing for Health Assessment . . . . . . . . . . . . . 97

v

4 Domain-Driven Health Assessment 1004.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

4.1.1 Participants & Procedures . . . . . . . . . . . . . . . . . . . 1014.1.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.2 Experiment 1 Findings . . . . . . . . . . . . . . . . . . . . . . . . . 1114.2.1 Daily Rhythms in Sleep and Social Technology Use . . . . 1114.2.2 Leveraging Social Data for Sleep Sensing . . . . . . . . . . 1144.2.3 Sleep’s Links with Neurobehavioral Functioning . . . . . . 122

4.3 Experiment 2 Findings . . . . . . . . . . . . . . . . . . . . . . . . . 1294.3.1 Smartphone Application Usage Patterns . . . . . . . . . . 1294.3.2 Circadian Alertness Rhythms Reflected in App Use . . . . 1324.3.3 Connecting App Use and Alertness with Sleep . . . . . . . 138

4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1404.4.1 Limitations and Future Work . . . . . . . . . . . . . . . . . 1444.4.2 Implications for Domain-Driven Health Assessment . . . 147

5 Domain-Aware Intervention Design 1515.1 Chronobiology-Aware PHI . . . . . . . . . . . . . . . . . . . . . . . 1535.2 Chronobiology-Informed Sleep Support . . . . . . . . . . . . . . . 154

5.2.1 Personalizing Sleep Hygiene Recommendations . . . . . . 1555.2.2 Stabilizing Circadian Disruptions . . . . . . . . . . . . . . 158

5.3 Biologically-Friendly Productivity Technology . . . . . . . . . . . 1655.3.1 Self- and Social-Awareness . . . . . . . . . . . . . . . . . . 1665.3.2 Scheduling and Activity Management . . . . . . . . . . . . 1715.3.3 Performance-Predictive Systems . . . . . . . . . . . . . . . 175

5.4 Mental Health Management . . . . . . . . . . . . . . . . . . . . . . 1785.4.1 Participatory Design of MoodRhythm . . . . . . . . . . . . 1795.4.2 Iteratively Refining Design Guidelines . . . . . . . . . . . . 1825.4.3 Implications for Preemptive Interventions . . . . . . . . . 187

6 General Discussion and Conclusion 1906.1 Summary of Chapters and Contributions . . . . . . . . . . . . . . 1916.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

6.2.1 Technology as a Double-Edged Sword . . . . . . . . . . . . 1956.2.2 Manual and Passive Modes . . . . . . . . . . . . . . . . . . 2016.2.3 Avoiding Over-Personalization of Tailored Experiences . . 2046.2.4 A Dual (Not Duel) Domain & Data Driven Approach . . . 210

6.3 Opportunities for Future Work . . . . . . . . . . . . . . . . . . . . 2156.3.1 Moving From Personal to Collective Informatics . . . . . . 2156.3.2 Applying the Framework in Areas Beyond Health . . . . . 2186.3.3 Open Challenges of Data, Analysis, and Design . . . . . . 224

6.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 229

Bibliography 230

vi

LIST OF TABLES

4.1 Categories of applications used by participants in Experiment 2,with examples and amounts of applications and usage events. . . 109

4.2 Average sleep duration for each participant according to social-sensor data, screen on/off data, and ground truth sleep diarydata. (* denotes inferences that fall within 95% confidence inter-val based on diary self-reports, p < .01). . . . . . . . . . . . . . . . 114

4.3 Median values of CMC-based activity levels following nights ofAdequate vs. Inadequate sleep. Significant differences in medi-ans marked on variable name (**p < .001, ***p < .0001). . . . . . . 124

4.4 Median values of cognitive functioning variables followingnights of Adequate vs. Inadequate sleep. Significant differencesin medians marked on variable name (*p < .05, **p < .01). . . . . 126

4.5 Median values of sentiment expressed in Facebook posts follow-ing nights of Adequate vs. Inadequate sleep. Significant differ-ences in medians marked on variable name (***p < .0001). . . . . 127

4.6 Median values of usage features during low vs. high alert-ness. Significant differences in medians marked on variablename (∗p < .05). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

5.1 Guidelines for designing self-monitoring technologies to sup-port mental health management. . . . . . . . . . . . . . . . . . . . 186

5.2 Variations in technology use identified as characteristic of maniaand depression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

vii

LIST OF FIGURES

2.1 A framework for domain-driven PHI development, comprisedof stages for tapping domain knowledge, collecting and ana-lyzing personal data to assess health, and designing user-facingfeedback. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.1 Chronotypes of participants in Experiment 1. . . . . . . . . . . . 1044.2 Distribution of participant chronotypes in Experiment 2. . . . . . 1044.3 Daily trends in participants’ average CMC-based usage. . . . . . 1114.4 Average sleep duration on work days and free days reveal the

“scissors of sleep” phenomenon — a discrepant pattern of sleepon work days versus free days that is reversed for early and latechronotypes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

4.5 Socially-sensed average social jet lag (discrepancy between mid-sleep on free days and work days) across chronotypes. . . . . . . 118

4.6 Shifts in sleep midpoint across study phases. . . . . . . . . . . . . 1194.7 Sleep duration and sentiment the following day. . . . . . . . . . . 1274.8 Hourly smartphone app usage by category. . . . . . . . . . . . . . 1304.9 Use across the week of productivity and entertainment apps

shown with standard error. . . . . . . . . . . . . . . . . . . . . . . 1314.10 Percentage increase (positive y value) or decrease (negative y

value) in amount of usage by early types compared to usage bylate types of productivity and entertainment apps across the day. 133

4.11 Temporal trends in application use (Usage) and alertness per-formance (Performance) across internal body clock time (InT).Usage axis is proportion (normalized to [0,1] scale) of all an appcategory’s usage events that occurred in a given hour. Perfor-mance axis is percent deviation from individual baseline of alert-ness measured in a given hour. Internal time axis is numberof hours since biological midnight, and accompanying spectrumindicates periods of the biological day. . . . . . . . . . . . . . . . 135

5.1 To provide peripheral self-awareness, this live wallpaper’s colortransitions in real-time in accordance with the user’s alertnesslevels at that moment. By default, brighter, more yellow satu-ration corresponds to higher alertness while a faded blue-graycolor corresponds to lower alertness. . . . . . . . . . . . . . . . . 167

5.2 This chronobiology-aware calendar background scaffolds self-awareness of personal alertness levels, which are represented us-ing a customizable color scale. Visual indicators on events pro-vide an at-a-glance sense of whether scheduling aligns with per-sonal alertness at that time. . . . . . . . . . . . . . . . . . . . . . . 168

viii

5.3 The color of personal “beacons” transitions over the course of theday to display personal alertness levels in a way visible to otherpeople. Beacons could be objects placed in the environment (e.g.,a desk ornament, lamp, etc.) or wearables worn on the body(e.g., a necklace, pin, bracelet, ring, etc). . . . . . . . . . . . . . . . 169

5.4 This “family portrait” (or “roommate portrait”, etc.) interfacesyncs to personal devices and uses a shared digital display in thehome to deliver chronobiology information (e.g., performanceand sleep-wake patterns) about household members as a wayto coordinate scheduling and improve interpersonal awarenessand empathy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

5.5 Specifying performance-related parameters (e.g., alertness, asshown here) enables scheduling recommendations that alignwith personal rhythms. . . . . . . . . . . . . . . . . . . . . . . . . 171

5.6 This “Fix My Day” scheduling assistant analyzes the day’sevents in order to provide suggestions for rearrangements thatoptimize alignment with personal performance patterns. . . . . . 172

5.7 Smartphone and smartwatch clock widgets show glanceableviews of personalized activity recommendations. . . . . . . . . . 173

5.8 Storyboard of a playful chronobiology-aware activity recom-mender that provides in-the-moment feedback through an inter-active, haptic experience. . . . . . . . . . . . . . . . . . . . . . . . 174

5.9 This interactive sleep-decision-support tool provides feedbackabout the net gains or losses of staying awake (e.g., to study fora test) in terms of the predicted impact of specified sleep choiceson performance the next day. . . . . . . . . . . . . . . . . . . . . . 176

5.10 Screenshots of MoodRhythm. . . . . . . . . . . . . . . . . . . . . . 1815.11 Percentages of respondents (y axis) who report tracking speci-

fied indicators (x axis). . . . . . . . . . . . . . . . . . . . . . . . . . 1835.12 Percentages (y axis) of various methods (key) used to track spec-

ified indicator (x axis). . . . . . . . . . . . . . . . . . . . . . . . . . 1845.13 Example of one participant’s custom tracking setup that cap-

tures personally meaningful variables, assesses daily status, anddelivers messages of encouragement. Diary entries have beenblurred to protect the participant’s privacy. . . . . . . . . . . . . . 185

ix

CHAPTER 1

INTRODUCTION

The World Health Organization defines health as “a state of complete physical,

mental, and social wellbeing — and not merely the absence of disease or infir-

mity” [533]. To achieve the full breadth of this definition, health care is breaking

free from its traditional illness-centered, doctor-dependent model of visit–test–

treat and moving toward a person-centric vision of medicine that is more proac-

tive, personalized, and self-driven. Making the individual the nexus of her own

health management, with a focus on overall wellness and prevention, could be

pivotal in solving the root of contemporary public health challenges [475].

Technology can play a major role in this metamorphosis. Specifically, com-

puting advances open new avenues to better understand and shape human be-

haviors — which is critical, given that a person’s behavioral and lifestyle choices

currently provide the single greatest opportunity to improve health amidst

today’s staggering prevalence of chronic disease [449]. Indeed, researchers

within the Human-Computer Interaction (HCI) community and relevant sub-

areas (e.g., personal informatics, persuasive computing, mobile health) are also

identifying opportunities where interactive systems can play a positive, perhaps

transformative, role in addressing modern society’s health problems.

Still, further steps must be taken in making these possibilities a reality. Most

relevant to this dissertation is ensuring that tools are personally relevant, con-

textually appropriate, and lead to genuine health improvements. I argue this

goal can be made more achievable by deeply grounding development decisions

in knowledge from health-relevant domains.

1

1.1 Motivation

Today, lifestyle choices are the root of poor health for many people. These be-

havioral health determinants are linked to preventable, chronic diseases that

require extended management and come with a heavy price, in terms of both

lost quality of life as well as financial spending [465].

Chronic diseases are now the leading cause of sickness, disability, and death

worldwide — attributable to 68% of all deaths and 43% of the global burden

of disease in 2014 [535]. By 2020, these figures are expected to rise to 73% of

all deaths and 60% of the global burden of disease [531]. This is the case in

both “developed” and “developing” countries. For instance, chronic medical

conditions affect 46% of the United States population [13], and 82% of deaths

attributed to chronic diseases occur in low- and middle-income countries [535].

Apart from mortality, most chronic diseases also negatively impact a person’s

functional abilities, productivity, and overall quality of life [322].

These statistics about prevalence are not only distressing in and of them-

selves, with respect to the value of human life, but they also foreshadow an

unsustainable financial burden [534]. Already, approximately three-quarters of

all health care expenditure in the U.S. is on patients with one or more chronic

conditions, many of which are preventable [82]; and over the next twenty years,

such conditions are expected to cost more than $47 trillion globally [52] as their

prevalence continues to increase worldwide [439].

In the face of these sobering personal and societal costs, the health domain is

primed for a shift toward a more proactive and personalized model of care [96],

where innovative services empower individuals to better manage their lifestyle

2

choices, behaviors, and overall wellness [439]. Personal technology has the po-

tential to support this revitalization of health care solutions. Sensor, web, and

mobile technologies in particular could fundamentally change how we monitor

and try to positively influence behavior. A powerful combination of factors is

enabling this change.

First, personal technologies are becoming increasingly sophisticated, in

terms of both their data-capture features and interactive affordances. Sensors

now standard on most smartphones include the accelerometer, compass, GPS,

gyroscope, ambient light detector, proximity detector, dual microphones, and

dual cameras [274]. A variety of other personal devices are now arriving on

the scene as well (e.g., wearables), and their sensing capabilities are progressing

similarly. Such functionality permits broad-scale, naturalistic collection of per-

sonal health-relevant data in an extremely granular, unobtrusive, and affordable

way. The technical ability to observe behavior continuously and in context also

makes it possible to tailor health interventions to optimize their effectiveness

for an individual user; plus these technologies provide an interactive medium

through which health applications can deliver that information.

Second, recent years have seen a massive swell in personal technology pen-

etration, which is becoming increasingly accessible and affordable. The adop-

tion rates of mobile phones are particularly striking. In the United States, over

90% of people own cell phones and 72% own smartphones [401], and 80% of

adults globally are estimated to have a smartphone by 2020 [153]. Adoption is

even higher for young generations — 86% of 18–34 year olds in the U.S. own

a smartphone and are heavily habituated users [64]. Smartphone ownership

does decline with age, but that trend is changing over time; and other tradition-

3

ally underserved groups in the U.S., including racial minorities or those with

relatively low income and education levels, rely heavily on their smartphones

for numerous important life activities [461]. Thus while the arrival of the inter-

net was accompanied by a “digital divide” that would have hindered the reach

of health interventions to some individuals, mobile phones have been widely

adopted across socioeconomic and demographic groups. Mobile network cov-

erage is continuing to expand as well. Over 85% of the world’s population is

now covered by a commercial wireless signal (surpassing many other forms of

infrastructure such as paved roads, electricity, and wired Internet), and over

70% of today’s wireless subscribers live in low- and middle-income countries

[244]. Such penetration is highly encouraging, especially given that it provides

avenues to engage with traditionally difficult to reach populations [233].

Finally and importantly, not only has the rapid development and spread of

technology placed a powerful health management platform in many pockets,

but people are also receptive to using these systems for self-monitoring [431].

Individuals are increasingly using technology to measure and record a variety

of health-related items; for example, 7 in 10 U.S. adults now track at least one

health indicator (e.g., blood pressure, mood, etc.) for themselves or for a loved

one [168, 169]. There are now over a quarter of a million health apps avail-

able to smartphone users, and it is expected that over 1.7 billion people will

have downloaded health apps by 2017 [413]. Additionally, adoption of health-

oriented devices (e.g., wearables like Fitbit or Apple Watch, “smart” objects like

networked scales or mattresses, etc.) is quickly climbing [299]. In fact, market

research indicates that such personal health management technologies will sur-

pass $70 billion by 2024 [201]. This increasing engagement is also reflected by

the rise in movements like “lifelogging” or the “Quantified Self”, which refer

4

to an individual using technology for self-tracking various aspects of daily life,

often with a focus on health and wellness improvement. Indeed, scholars have

recognized a shift in mindset from “My health is the responsibility of my physi-

cian” toward “My health is my responsibility, and I have the tools to manage it”

[475]. Further, medical professionals, governments, and international organiza-

tions (e.g., National Council for Behavioral Health, World Health Organization,

United Nations) also express receptivity to technology-based treatment proto-

cols, in many cases even strongly advocating for deeper integration of personal

self-monitoring tools into existing public health care systems [244].

Altogether, this swift spread and rapid technical evolution of computing

technology, together with the growing engagement from individuals who want

to proactively improve their health and quality of life, creates unprecedented

opportunities for HCI researchers to develop robust, interactive, and tailored

health applications that are also scalable and cost-effective [416, 489]. Creat-

ing such technological solutions that empower users to monitor, manage, and

improve personally-significant aspects of their wellness is thus an area of im-

mediately compelling opportunity.

This dissertation examines promising HCI efforts already underway (e.g.,

behavioral intervention technology, persuasive computing, personal informat-

ics); and in doing so, I identify a gap between evidence and practice that in-

dicates more work is still needed to make the types of personal data collected,

the health metrics modeled from that data, and the behaviors targeted for inter-

vention more personally, contextually, and clinically relevant. The crux of this

dissertation then lies in “domain-driven” research strategies I provide to close

this gap.

5

Specifically, I set forth a domain-driven development framework: a vision

and plan for connecting the creation of personal health informatics technology

more tightly with domain knowledge — the concepts, theories, empirical find-

ings, user feedback, and any other forms of pertinent background information

that can inform development choices regarding data collection, health model-

ing, and behavioral feedback.

To demonstrate this framework in practice, I use a case study I have un-

dertaken, aimed at advancing a class of chronobiology-driven technologies for

supporting sleep, daily performance, and emotional wellness. Much of this

work has been published in prominent conferences (e.g., the ACM Conference

on Human Factors in Computing Systems — CHI, the ACM International Joint

Conference on Pervasive and Ubiquitous Computing — UbiComp, and the In-

ternational Conference on Human-Computer Interaction with Mobile Devices

and Services — MobileHCI) and journals (e.g., Human-Computer Interaction,

Computers in Human Behavior, Assessment, and the Journal of the American

Medical Informatics Association — JAMIA) or is in print, under review, or in

preparation for future submission.

6

1.2 Dissertation Overview & Contributions

This dissertation makes methodological, empirical, and human-computer inter-

action contributions, organized into the following six chapters.

Chapter 1 (“Introduction”) has provided the high level motivation behind

my dissertation research — an alarming incidence of chronic disease, opportu-

nities for personal technology to help address this well-recognized health crisis,

and a domain-driven strategy I propose to make health management technol-

ogy more efficacious.

Chapter 2 (“Background”) offers a view, from an HCI perspective, of

the current landscape of technologies working to pursue those opportunities.

Overviewing various areas of interrelated work, I provide definitions and a

common vocabulary used in the remainder of the dissertation. In its review of

extant work and prominent systems, Chapter 2 also explains in detail the value

in developing new generations of tools whose design choices are more deeply

grounded in domain knowledge — from the data a system senses automatically

or allows a user to manually log, to the health metrics it extracts from that data,

to the personally and clinically relevant feedback it provides. After constructing

a working definition of “domain knowledge”, this chapter concludes by spec-

ifying the components of a domain-driven development framework, which is

unpacked and demonstrated in the chapters that follow.

Chapter 3 (“Domain Inquiry”) explains the initial, foundation-building

steps in developing a domain-driven system. Specifically, this chapter describes

a process for selecting a compelling application area, identifying salient do-

mains from which to gather background knowledge, and using that under-

7

standing to inform subsequent research and development decisions. To demon-

strate this process in practice, I use my research on technology for supporting

sleep, cognitive performance, and emotional wellness as a case study. I first ex-

plain how I recognized these as areas ripe for technological solutions and what

limitations I saw in extant related work. I then describe how I identified chrono-

biology as a relevant domain, and I overview the background information I

gathered — knowledge I hope additionally serves as a standalone resource use-

ful for others interested in doing chronobiology-driven work. Finally, I apply

this knowledge in order to plan the subsequent PHI work, particularly regard-

ing the scope and modeling strategies for health assessment (i.e., what to assess,

for whom, and how).

Chapter 4 (“Domain-Driven Health Assessment”) contributes the meth-

ods and findings from two experiments I conducted to execute the domain-

informed analytic plan devised in the previous chapter. Specifically, the first

experiment explored how social sensor data can be used to detect sleep-related

behaviors and circadian disruptions, and it took preliminary steps toward ana-

lyzing the impact of inadequate sleep on cognition and mood. Digging deeper

into daily functioning, the second experiment then built on chronobiology about

cognitive performance rhythms in order to explore and interpret a number of

relationships among smartphone app use, alertness, sleep, and latent biologi-

cal traits. This chapter reports on the outputs of this research, along with its

implications for the general framework.

Chapter 5 (“Domain-Aware Intervention Design”) presents a cycle of

preparing, creating, and evaluating designs to deliver interventions and other

user-facing feedback about personal health. To demonstrate this process in

8

the context of this dissertation’s main case study, I apply knowledge gath-

ered during domain inquiry, findings from the two experiments presented in

the previous chapter, and interactions with users in order to devise a series of

chronobiology-aware design guidelines, mockups, prototypes, and full systems

for supporting sleep, activity scheduling, and mental health management.

Chapter 6 (“General Discussion and Conclusion”) concludes this disserta-

tion by synthesizing takeaways and pointing out possible future directions. In

particular, this chapter discusses strategies for balancing key tradeoffs, which I

encountered in my case study research but are applicable to PHI more broadly:

technology’s impacts on health as a double-edged sword, integrating man-

ual and passive health management activities, avoiding over-personalization

when tailoring experiences, and combining domain-driven with data-driven

approaches. In closing, I describe opportunities for future work, including

moving beyond the traditional single-user model to more collective styles of in-

formatics as well as identifying areas beyond health where my domain-driven

framework could generalize.

9

CHAPTER 2

BACKGROUND

This chapter reviews and critiques academic and industrial work on technology

aimed to support personal health. Specifically, after describing how modern

medicine evolved to become what it is today and the role personal technol-

ogy has played, I coin a new term with a tractable definition that I use in this

dissertation to refer to a class of systems that support personal management of

healthy behavior. I also outline requisite features of this technology and provide

examples of extant systems that possess these characteristics.

In doing so, I also motivate a core argument of this dissertation: that

such technologies will have more successful health outcomes if they are made

more personally, contextually, and clinically relevant — something that can be

achieved by using domain knowledge to inform development choices, from the

types of personal data collected, to the health metrics modeled from that data,

to the designs of delivered feedback and interventions.

My intention in the following sections is to provide a context for my pro-

posed domain-driven framework, rather than an exhaustive review. When

available, I point to other sources that provide more encyclopedic reviews of

the literature.

2.1 The Roles of Technology in Supporting Health

As explained in the previous chapter, technology provides a tremendous op-

portunity to address modern health care challenges, combat chronic disease,

10

and overall improve health and wellness on a broad scale. This opportunity

has been recognized across numerous fields with ties to health, so such efforts

unsurprisingly go by various names. In this section, I overview terminology,

definitions, and relationships among these related, sometimes interchangeable,

sometimes overlapping, sometimes distinct labels. For the sakes of scope and

relevance, I focus on concepts and technologies most prominent within HCI and

relevant to this dissertation (e.g., mobile health, personal informatics, quantified

self, and so on). It is actually difficult to find a comprehensive review in the liter-

ature that provides such an overview along with a history of how these concepts

originated and came to mean what they do today. I therefore find it useful to

look at these ideas through a historical lens, in terms of health care’s evolution

over the past three centuries.

2.1.1 The Roots of Modern Medicine

Modern medicine has its roots in the 19th century [76], which saw changes in

the conceptualization of disease and illness (e.g., from the discovery of bacte-

ria), an introduction and increased use of statistical methods (e.g., correlational

analysis and hypothesis testing), improved care practices (e.g., anesthesia, bet-

ter hygiene), and the emergence of new fields (e.g., psychiatry) for previously

undertreated conditions such as mental illness [126, 154].

In the 20th century, medicine became increasingly professionalized and saw

new formal mechanisms for regulating medical practice. Organized medical

research also gained tremendous momentum. In the decades just prior to the

start of the 20th century, the precursor to today’s National Institutes of Health

11

(NIH) was established, and government initiatives began to support the fund-

ing and supervision of scientific medical research [363]. Such professionaliza-

tion also led to breakthroughs in public health research and programs. As public

health became an increasing priority during the past century, population-level

epidemiologic studies became more feasible due to instituted changes such as

the deployment of periodic standardized health surveys (e.g., the U.S. National

Health Survey, established in 1956) [474]. Studies also moved beyond only mea-

suring disease prevalence to more controlled experimentation (e.g., randomized

clinical trials), beginning in 1948 with the first modern clinical trial of an an-

tibiotic drug for tuberculosis [498]. Indeed, pharmacology in the 20th century

became increasingly sophisticated, with more focus on the development and

use of drugs and medications (e.g., antibiotics, vaccines, psychiatric drugs, and

vitamins) [474].

Numerous advancements in medical technology were achieved as well, for

instance beginning with the dissemination of the X-ray machine to most hospi-

tals near the beginning of the 20th century [496]. The technologies that are most

relevant to this dissertation also have their foundations in the 20th century, dur-

ing which time technological development continued to rapidly progress.

Enabled by the production of 20th century telecommunication technologies,

telehealth (also known as virtual health care) refers to the use of telecommu-

nication services and information technology to deliver health care or health

information from a distance. Though telehealth is often considered synony-

mous and used interchangeably with telemedicine, they can be distinguished,

as telemedicine refers specifically to traditional clinical diagnosis and monitor-

ing delivered remotely (e.g., supporting a physician’s remote monitoring of a

12

patient or enabling transmission of medical images for diagnosis), while tele-

health has a broader scope that includes the distribution of a range of non-

clinical health services in addition to or aside from clinical services.

Telehealth is further encompassed by eHealth [378], a recently defined um-

brella term that describes the use of electronic technology or digital data to sup-

port health care, whether it be used remotely or locally [125]. Beyond telehealth

services, other eHealth technologies include electronic health records, clinical

decision support systems, and computerized physician instruction tools. The

use of such information technologies in health care is often referred to as health

informatics (also known as health care informatics or medical informatics),

which is concerned with the collection, storage, retrieval, management, and use

of health information by a patient’s care providers.

eHealth also includes mHealth applications. mHealth (for mobile health)

broadly refers to the use of mobile phones or other wireless devices to sup-

port health care [244]. A sizable portion of my research would be considered

mHealth, as I too leverage mobile technology (particularly, smartphones) to

assess health behaviors and deliver behavioral interventions. I tend to focus

on participants from the United States due to access, and mHealth is certainly

applicable and can have significant impacts in industrialized nations and high

income areas. However, given the substantial and widespread penetration of

mobile phones, a recent thrust of the mHealth field has focused on applica-

tions in developing, rural, or low-income areas, with the goal of improving

health care quality and access for underserved populations. Examples of typical

mHealth technologies include phones, mobile applications, personal digital as-

sistants (PDAs), and patient monitoring devices (e.g., network-enabled glucose

13

monitors or portable electrocardiogram devices); and recent years continue to

see the arrival of various “smart” mHealth technologies such as smartwatches,

smart eyewear, and smart scales.

Such technological developments and new treatments have helped global

life expectancy rise by over thirty years in the past century [531]. However,

as the prevalence of illnesses common in the early 20th century (e.g., measles,

rickets, typhoid fever) shrank, an incidence in chronic diseases (e.g., diabetes,

heart disease, cancer, and depression) increased [439, 535]. A greater focus in

the later half of the 20th century on treating such non-communicable diseases

paralleled the emergence of behavioral medicine, an interdisciplinary field con-

cerned with developing behavioral and biomedical knowledge and techniques

relevant to health in order to support illness diagnosis, treatment, and preven-

tion [252, 450].

2.1.2 The Age of Behavior Change

I therefore designate the 21st century as the age of behavior change. One of

our greatest present health challenges is dealing with mental health problems,

chronic conditions, and lifestyle diseases, which are linked with how people live

their lives and are the leading cause of sickness, disability, and death worldwide

— attributable to 68% of all deaths (and 82% of deaths in low- and middle-

income countries) [535], as noted in the previous chapter. Similarly, the top risk

factors for premature death (e.g., high body mass index, physical inactivity, un-

healthy food choices, smoking, and excessive alcohol consumption) all relate

to lifestyle choices. According to the U.S. Centers for Disease Control and Pre-

14

vention (CDC), eliminating just three risk factors — inactivity, poor diet, and

smoking — could prevent 80% of heart disease cases, 80% of type 2 diabetes

cases, and 40% of cancer cases [324].

Consequently, scholars are reaching a consensus that “the single greatest op-

portunity to improve health and reduce premature deaths lies in personal be-

havior” [449]. Indeed, while genetics, behavioral patterns, socioeconomic cir-

cumstances, environmental exposures, and the quality and efficacy of health

care all influence our health [439], research now shows that an individual’s be-

havioral choices (40%) combined with social (15%) and economic environments

(5%) contribute the most (altogether, 60%) to personal health status [449]. Ad-

dressing such behaviorally rooted issues therefore requires getting individuals

more directly involved in their own care.

Thus, in light of the rise in chronic conditions and the major role that lifestyle

choices play in condition development, progression, and outcome, two major

shifts are occurring within the health domain. One, it is increasingly moving

toward care models focused on management rather than treatment, with mon-

itoring as a daily activity rather than an occasional consultation. In addition,

approaches are increasingly focusing on individually-targeted behavioral inter-

ventions. Technology is therefore seen as a highly appealing mechanism for

delivering such interventions, as it can provide behavioral coaching that is con-

tinuously available and delivered directly to the patient. An additional benefit

of contemporary technology is that it can reach populations normally unable to

access care due to various financial and physical barriers [337, 338].

Based on established behavior change counseling principles known as the

“5As” (assess, advise, agree, assist, arrange) [182, 522], interactive behavior

15

change technologies (IBCTs) were one of the first technology-based approaches

to using hardware and software (e.g., DVDs, PDAs, emails, phone calls, and

patient-centered websites) to deliver the 5As to patients before, during, and af-

ter primary care visits [181]. For instance, IBCTs have been used to improve

diabetes self-care by delivering walking interventions, encouraging medication

adherence, and facilitating patient-to-patient peer support [398].

Over time, additional terminology has been introduced to refer to the use

of technology to support behavior change, and novel digital mediums continue

to be appropriated for the task. Today, the term behavioral intervention tech-

nology (BIT) is broadly used to refer to a range of modern modalities (e.g., mo-

bile phones, web 2.0, and wearable and environmental sensors) utilized to sup-

port users in changing behaviors related to physical health, mental health, and

overall wellbeing [337, 339]. Persuasive technology (or persuasive computing)

is similarly intended to change or maintain attitudes and behaviors [164]; and

while persuasive technologies are found in many domains (e.g., sustainability,

education, and activism), many today focus on applications to health, accord-

ing to previous reviews as well as my own examination of academic reposito-

ries and consumer application markets [356]. Within HCI, the term “persuasive

technology” is still common, but it is considered somewhat controversial due to

concerns about implied coercion [406]. Researchers therefore instead sometimes

simply use the phrase behavior change technologies [206].

As the landscape of these technologies for supporting healthy behavior con-

tinues to evolve, one notable characteristic is that usage is not only moving out-

side of clinical settings where health care has been traditionally delivered — but

clinical oversight itself is diminishing. This is in large part due to the increased

16

availability and uptake of direct-to-consumer products for self -managing health

(e.g., physical activity trackers, internet-connected scales, and bluetooth heart-

rate monitors), which typically provide mobile applications or online counter-

parts to provide guidance in adopting or maintaining healthy behaviors [99].

A primary component of these self-monitoring systems is the collection of

personal data, through self-tracking done either manually or automatically via

sensors (e.g., accelerometer-based step count sensing). The term personal in-

formatics (PI) was coined within the last decade to refer to technologies aimed

at helping users collect and reflect on personal information [284]. This work

conceptualized self-tracking as a five stage iterative process through which a

person will prepare what data to collect and how, collect that data, integrate and

organize collected data, reflect upon and interpret data, and determine how to

convert gained understanding into a plan for action. Subsequent work has ex-

panded this model in a number of ways, including to identify additional styles

of tracking (e.g., goal-driven and documentation-based activities [431]), stages

of tracking (e.g., a maintenance phase and lapsed tracking [149]), and ways to

accommodate clinician-patient collaborations when self-tracked data from com-

mercial tools is used as part of treatment [97].

While people use personal informatics tools for various reasons (e.g., to track

finances, document visited locations, out of curiosity, or to receive rewards on

social networking sites) [149], a majority of users are interested in capturing and

accessing health data. Similarly, while improving self-knowledge is nominally

the goal of PI, the value in exploring personal data stems for most individuals

from a desire to translate gleaned information into self-improvement strategies,

particularly with respect to health behavior change [476].

17

A recent rise in the practice of self-tracking is known as the Quantified Self

(QS) movement, which refers to self-monitoring any aspect of one’s life by cap-

turing physiological, behavioral, emotional, cognitive, or environmental infor-

mation, typically with the goal of improving or optimizing physical or mental

health. Compared to personal informatics users in general, the QS commu-

nity has been identified as an example of “extreme” self-trackers [94], aiming to

make the body a more knowable and hence “calculable and administrable ob-

ject” through QS activities [476]. Indeed, these “QSers” capture a vast range of

personal data, sometimes using highly invasive methods (e.g., brainwaves via

EEG, neurotransmitter measurement via urine and blood serum, video records

of one’s totality of experiences using lifelogging [135] apparatus, or, for one

super-self-tracker, even the intensity of every crying session experienced over a

period of more than one and a half years [519]).

Such extensive self-monitoring pertains to a person’s interest in conducting

a form of self-experimentation [94]. This practice has a long history in medicine

and psychology, where doctors have traditionally volunteered for ethical rea-

sons as the first subject in a human experiment that has unknown risks [10].

Very recently, HCI researchers have begun designing technology to support

what they similarly refer to as self-experimentation in self-tracking, for instance

to assist an individual with Irritable Bowel Syndrome in identifying foods that

trigger symptoms or to help a person determine whether timing her exercise for

the morning does result in more energy later in the day [241]. This work is mo-

tivated by the fact that people want to use personal informatics tools to answer

specific questions like these about their health, but current tools typically fail to

effectively support such diagnostic self-tracking [240]. For example, many tools

output graphs of raw data that users find difficult to interpret or act on [148],

18

and tools generally do not support personal experiments that have sufficient

methodological rigor [94].

Self-experimentation technologies essentially help a user self-administer a

controlled experiment; the tool creates an experiment scheduling, encourages

adherence to experimental conditions, and automatically runs statistical tests

from which a user can draw causal conclusions. The experiment follows a

single-subject design (also known as an n-of-1 study), which is sensitive to in-

dividual differences and where a person serves as his or her own control [286].

HCI’s interest in n-of-1 style personal informatics coincides with calls from

within the medical community to adopt models of personalized medicine (or

individualized or precision medicine) that focus on individual, rather than av-

erage, responses to particular treatments. Such an approach can be advanta-

geous compared to methodologies involving larger samples (e.g., randomized

controlled trials), which can lead to therapeutic solutions that are beneficial to

some patients but minimally effective or even detrimental for others [176]. For

example, some routinely used medications benefit as few as 1 in 50 individuals,

while other drugs have been found to be harmful for entire ethnic groups since

clinical trials are often biased toward white Western participants [447].

2.1.3 The Individual as the Nexus of Health Management

In recent decades, modern medicine has thus been marked by three key devel-

opments that are highly relevant to this dissertation and that, in a nutshell, have

drawn the health domain toward a more individual-centric paradigm focused

on wellness and prevention.

19

First, health care is increasingly focusing on behavior change, given the afore-

mentioned growing crisis of chronic diseases, which are often preventable and

manageable through lifestyle changes. At first, efforts to promote healthy be-

havior largely focused on public awareness campaigns to increase health liter-

acy or policy-making to decrease unhealthy behavior (e.g., smoking, package

labeling); but such generic methods are only modestly effective [439].

Second, recognizing the inadequacy of such one-size-fits-all approaches to

behavior change, health care has become more personalized. Over the years,

health care has become more patient-centered [508] to resemble more of a part-

nership between practitioners and patients aimed at ensuring patient prefer-

ences are incorporated into clinical decision-making and that patients have the

needed support to participate in their own care [219]. Similarly, personalized

medicine now more explicitly considers such individual differences in patient

needs in order to account for the fact that a given treatment may not affect ev-

eryone equally. While studies on groups of people have dominated medical

science over the last century, with the randomized controlled trial considered a

gold standard for evaluating an intervention’s efficacy, we are now seeing in-

creasing momentum behind n-of-1 study designs, where the sample is a single

participant: the patient requiring treatment.

Finally, technological developments have allowed health care to move out-

side of clinical settings and have supported behavior change that is self-driven.

While technology was quickly recruited as a way to more broadly and afford-

ably deliver health care directly to an individual, physicians were still the pre-

dominant point of care until very recently. Today, thanks in part to the advent

of novel devices and mediums (e.g., smartphones, wearables, and web 2.0 tech-

20

nologies), individuals can independently collect, analyze, and act upon data

representing health and wellness. Such activities can complement — or even

replace — interactions with health care professionals [433], who individuals no

longer need to depend upon to introduce the notion of behavior change nor

drive progress. Less reliance on doctors could additionally help buffer the neg-

ative impacts of anticipated physician shortages. Though, it is worth noting

that for some conditions, professional oversight may be an important aspect of

care; in such cases, the goal of technology would be to supplement and extend

physicians’ efforts rather than to bring them out of the loop entirely.

In exploring how to provide each person with the tools needed to support

everyday health-related behaviors, this dissertation champions this idea of per-

sonalized self-management, where care centers on the individual, who is the

focus of empowerment and the nexus of positive change.

2.2 Personal Health Informatics (PHI)

As just reviewed, a variety of terminology is used to refer to various related

types of technologies designed to support behavior change (e.g., “behavioral in-

tervention technology”, “persuasive technology”, “personal informatics”, and

so on). Some terms are used interchangeably, some terms are considered subsets

of other terms, and some terms share a portion but not all of their characteris-

tics with other terms — plus some terms do not have definitions that are even

unanimously agreed upon.

21

For the sake of consistency and mutual understanding with the reader, I

therefore find it useful to introduce a new term I will use throughout this dis-

sertation to refer to the technology my research is aimed at developing, Personal

Health Informatics, defined as follows.

Personal Health Informatics (PHI) refers to technology that sup-

ports personal management of healthy behavior. In essence, this

class of tools (1) facilitates collection of personal data, (2) enables

analysis of information to assess a targeted aspect of health, and (3)

provides feedback to help a person gain self-knowledge and poten-

tially change or maintain behavior accordingly. While this definition

could be applied to a range of application areas, this dissertation is

particularly focused on the development of PHI systems that pro-

mote healthy sleep, cognitive performance, and emotional wellness.

First, I include the word “informatics” in the term, as a primary goal is aiding

a user in capturing, interpreting, and acting upon personal data. Since this dis-

sertation is focused on individuals engaged in this process for reasons related

to “health”, I include that word in the term in order to distinguish PHI from the

broader definition of “personal informatics”, which as explained earlier, can en-

compass a wider range of personal data from finances to energy consumption.

Finally, I include the word “personal” because (as described at the end of the

previous subsection) the focus is on personal usage that places an individual at

the center of services that enhance health by promoting self-management.

Because my work comes from an HCI foundation, the characteristics of PHI

are inspired by and most closely resemble tools that grew out of the HCI field,

namely personal informatics (PI) tools (compared to, say, telemedicine tools).

22

The key differences are that PHI is scoped to health self-management, more ex-

plicitly designates the health assessment element (whereas PI traditionally fo-

cuses on users’ data collection and interpretation practices and less on what is

potentially happening under the hood), and aims to provide user-facing feed-

back that is more personalized based on those user models (whereas PI tends

to provide generic visualizations [21]). At the end of the day, a main reason for

introducing a new term in this dissertation is simply to have a tractable defini-

tion with a clear-cut set of characteristics, given that prior definitions are either

broader than necessary for this dissertation’s purposes or somewhat discrepant

depending on the source doing the defining.

Specifically, I consider (1), (2) and (3) as the fundamental components to

“qualify” a system as PHI:

1. Capturing Personal Data: By harnessing emerging mobile and ubiqui-

tous technologies (e.g., smartphones, sensors, Internet of Things, perva-

sive network coverage, etc.), rich datasets about personal behavior can be

collected in context, through manual self-report and automated sensing.

2. Analyzing Data to Assess Health: From this personal data, health metrics

can be computed, condition symptoms detected, and predictive models of

personal behavior built.

3. Delivering User-Facing Feedback: Given this model of an individual’s

health and contributing behaviors, tailored feedback can be delivered in

order to support personal self-management and, in turn, enhance overall

wellness.

23

Thus essentially, personal data is the system input, user-facing feedback is

the system output, and in the middle, analysis transforms data into information.

Though, these elements can be overlapping and iterative; that is, there are not

necessarily hard boundaries between them. For example, many passive sens-

ing systems intertwine data collection and data analytics through a continuous

loop in which user models retrain themselves based on fresh data that is sensed

in real-time. Or, a system can collect a user’s positive or negative reaction to a

piece of feedback — whether it be implicit (e.g., acting upon a behavioral recom-

mendation) or explicit (e.g., using an option to “like” a recommendation) [247]

— to help profile that person’s preferences. Through such an iterative process

that incorporates new data, refines models, and updates feedback, a system can

continue providing support appropriate to a user’s evolving needs over time.

Further, I find it useful to think about each of these aspects — capture, anal-

ysis, and feedback — in terms of the level of responsibility and agency placed

in the hands of the user versus the system. Other researchers have used la-

bels such as how “participatory” [453] or “cooperative” [379] a system is, in

characterizing this relationship between its autonomy compared to that of the

user. First, personal information can be collected manually, using automated

approaches, or through a hybrid combination that augments self-tracking with

passively sensed data. Then, analyzing that data can be left up to the individual,

who is responsible for exporting, organizing, and making sense of that informa-

tion; or, a system might help integrate data, compute health statistics, or apply

machine learning algorithms to pull out patterns that the user can then examine.

Finally, feedback may be presented in more descriptive ways that are open to

interpretation by the user; or, feedback may be more prescriptive in dispensing

explicit directives.

24

In the following subsections, I describe each of these components (capture,

analysis, feedback) in more detail. Throughout, I include examples of related

systems in order to demonstrate how these qualities can look in practice; I fo-

cus on work from the HCI literature, as that is the community to which this

dissertation primarily speaks.

2.2.1 Capturing Personal Data

To begin, any personal health informatics tool requires input data. This input

provides details about the user’s behaviors, environment, or any other personal

attributes relevant to the health outcome the tool is designed to support. This

data can be captured manually by a user, automatically by sensors, or through

some hybrid approach. This subsection overviews these ways PHI technology

captures data, providing examples and pointing out advantages or drawbacks.

Manual Data Collection

People have manually self-tracked health-related personal information long

before digital tools existed to explicitly support the activity. In the 1940s, clinical

research began using written diaries, in which people self-report symptoms and

health actions as they occur [9, 501]. While such pen-and-paper approaches are

familiar and easy to use for many people, they also face well-known limitations

including the risk of forgetfulness, retrospection errors, uncertain adherence,

and lack of response-time information [55]. Numerous researchers have worked

toward addressing these limitations, and over the past thirty years or so, much

attention has been specifically focused on how technology can help.

25

At first, studies used digital devices such as pagers [118], pre-programmed

wristwatches [291], or text messages [15] to deliver reminders to record informa-

tion, though the recording itself was still made on paper. This sort of prompted

self-report has traditionally been associated with experience sampling (ESM)

[114] and ecological momentary assessment (EMA) [470], which are methods

used to collect information about various aspects of daily life in the moments

they are being experienced. Such in-situ assessment can reduce retrospective

biases, help in identifying ecological factors that may be contributing to an in-

dividual’s health status, and reveal whether a person is employing acquired

health management skills (e.g., from clinical therapy) in everyday life [420].

More recently, technology advanced to a point where it could be used not

only to prompt users to record information but as the recording medium too,

instead of paper. Early electronic data collection methods included handheld

computers such as PDAs loaded with questionnaire programs [33, 452]. Today,

most HCI researchers working on improving the experience of manually track-

ing with technology focus on smartphone-based self-report and target activities

that are quite difficult to detect automatically such as pain levels [491], food

intake [44], or subjective wellbeing [318].

One thread of HCI work is particularly interested in developing more

“lightweight” self-tracking interfaces. Several allow information to be recorded

directly from a smartphone’s lockscreen so that a user does not have to launch

a full diary application or even unlock the phone. For example, the SleepTight

smartphone app allows a user to tap icons on the lock screen in order to jour-

nal sleep disruptors [93]. The Slide-to-QuantifySelf app also allows a user to

record health metrics (e.g., daily water intake) using lock screen widgets [490],

26

and the LogIn system similarly repurposed unlock interaction gestures to cap-

ture self-reports for sleepiness and mood [548]. In developing the MoodRhythm

smartphone app for managing bipolar disorder, colleagues and I used notifica-

tions to prompt momentary self-reports, and we made it possible for patients to

quickly record target variables directly from the notification panel [312].

The manual capture of data is associated with several benefits. Self-tracking

can empower users with a sense of agency [354], plus directly engaging with

data can foster self-awareness [45, 94]. The “obtrusiveness” of the practice is

precisely its main advantage, as this is what enhances mindfulness about behav-

ioral choices and can promote adherence to behavior change goals [260, 262].

However, manual self-tracking is associated with disadvantages as well.

Foremost, self-report can be burdensome [104] due to the time and effort it re-

quires. This is a particular challenge if a technology is intended for long-term

use (e.g., to manage a chronic health condition). By reducing the amount of in-

formation a user needs to report, briefer assessments can overcome this burden

somewhat; but the shortness of such instruments can weaken their validity [73].

Burden can also degrade the quality of data collected if it translates into inad-

herence (i.e., missing data), which can also result from basic forgetfulness. This

issue is compounded if it leads to subsequent backfilling of missing data, which

is common [471] and typically inaccurate due to retrospective recall biases [112]

or even fabrication [55].

Data inaccuracy can also occur in cases where a person’s capacity for reliable

self-assessment is compromised, such as when suffering from sleep deprivation

[139] or if experiencing symptoms of certain mental health conditions including

bipolar disorder [190]. Relatedly, the act of self-assessment can be impeded in

27

certain contexts. For instance, self-report is simply not possible or could be dan-

gerous in some situations (e.g., during sleep or while driving); or performing

self-tracking may be difficult in some social settings, especially if the individual

is reporting about a potentially stigmatic health condition.

Next, while increased self-awareness can induce positive reactivity that

leads to desirable behavioral changes, negative reactivity can also result. For ex-

ample, in the context of self-reporting pain, some evidence finds that recalling

pain and coping strategies can lead to positive outcomes such as an increased

sense of control over the pain [204]. However, other findings suggest that re-

peated self-assessment of a potentially traumatic situation can actually draw

one’s attention to and foreground negative perceptions, thereby worsening the

lived experience of that condition [258].

Finally, it can be infeasible for a person to capture the array and granularity

of data necessary for a system to produce a sufficiently comprehensive profile

about that individual, comprised of the multiple personal variables, behavioral

determinants, and other indicators needed to accurately model a health out-

come of interest [45]. Thus, interest arises in more system-driven approaches

to data collection that are either fully automated or that complement self-report

with passively captured information.

Automated Data Collection

With automated or “passive” data collection, physiological or behavioral

trace data is captured using sensors or from technology usage logs. These sen-

sors can be worn on the body, located in the environment, or embedded within

a personal device.

28

Systems designed to encourage physical activity have used a variety of

body-based sensors such as pedometers [106, 290], biometric sensors like elec-

trocardiogram (ECG) [124], or custom sensing setups [107] composed of sensors

to capture sound, temperature, light, and humidity among other inputs [95].

Many of today’s commercial wearable devices designed to support healthy be-

haviors (e.g., Fitbit, Jawbone UP, Microsoft Band) are essentially accelerometer-

based wristbands that passively monitor activity and sleep [410], though newer

models incorporate additional sensors, for instance to measure heart rate or gal-

vanic skin response. Such sensors have also been incorporated into clothes and

jewelry (e.g., to assess mood [494] or support remote patient monitoring [414]).

The main disadvantages associated with body-based passive sensing are the

discomfort of wearing the sensing device, the limited battery life, and the fact

that smaller form factors inherently constrain the sensors that can be hosted —

although battery and miniaturization advances are helping to address some of

these issues [410]. As with manual data collection, forgetfulness can be an issue

for passive strategies as well; for instance, a user may forget to wear the sensing

device.

As environment-based sensors do not face these challenges, researchers have

been exploring how instrumented homes can automatically capture health data.

One system automatically captures weight using a scale built into the toilet,

heart rate data using an ECG monitor in the tub, and body temperature from

a bed sensor [376, 479]. Others have placed sensors to automatically collect

health metrics into furniture like chairs [195] or mattresses [257], onto home

appliances, or into cars [261]. The notion of an “internet of things” connected

smarthome would further extend such capability and connectivity to numerous

29

other objects within one’s living space, though this vision still remains largely

unrealized.

Smartphone-based sensing, on the other hand, has not only garnered

tremendous interest recently from researchers in HCI but has made substan-

tial progress, with mobile sensing emerging as a field in its own right [274].

The mobile phone has rapidly evolved into a powerful computing platform,

with a variety of sensors now standard in consumer smartphones for automat-

ically capturing motion (e.g., accelerometers, gravity sensors, gyroscopes); lo-

cation (e.g., GPS, orientation sensors, magnetometers); and environmental data

(e.g., barometers, photometers, thermometers, cameras, microphones). Mobile

sensing is often applied to the area of health, with recent reviews overviewing

prominent health-oriented smartphone sensing systems [86, 255]. As part of this

dissertation’s main case study is focused on sleep, I provide a more comprehen-

sive review of sleep sensing systems in particular in Section 3.2.3.

A novel twist on sensing is “soft sensing”, which passively captures data

not from hardware based sensors but from software usage logs. That is, as a

byproduct of interacting with numerous technologies (e.g., computer programs,

smartphone apps, or web services), individuals generate a massive amount of

data (often referred to as “big data” — or, “small data” if referring to the data

generated by a single individual [152]). By leveraging such digital trails, re-

searchers taking a soft-sensing approach have had success in inferring a user’s

health-related behaviors, contextual or psychological states, or other personal

characteristics of interest [120].

30

Overall, automated sensing helps relieve user burdens by reducing both the

time and mental overhead associated with self-tracking, plus sensed data is of-

ten more accurate and granular than manually tracked data. Automatic sens-

ing can also capture informative quantitative signals that are relevant to health

but may be imperceptible to the person generating those signals [523]. How-

ever, sensors can be privacy invasive or (for wearable sensors) uncomfortable

to wear, and they can reduce personal awareness about collected data [283]. In

addition, automatic data collection can work well to acquire “objective” infor-

mation like heart rate or location; but as mentioned, it does not lend itself as well

to measurement of subjective experiences. Finally, practically speaking, while

capturing some types of data is now reliable (e.g., location, walking detection),

other types of data are still more elusive.

Hybrid approaches attempt to make the best of both worlds by employing

more than one capture mechanism. For example, UbiFit automatically infers

physical activity data about walking, running, and cycling; but the system al-

lows the user to add activities it cannot automatically track like yoga or swim-

ming [108]. The Somnometer sleep support system automatically tracks sleep

duration from phone data but obtains sleep quality information from the user

[455]. Commercial health trackers like the Fitbit and Jawbone UP similarly use

accelerometers to automatically track activity and sleep data while also provid-

ing an interface where users can manually log additional information such as

mood, meals, caffeine, and water intake. Such information is notoriously diffi-

cult to automatically detect; and a number of self-monitoring technologies use

this hybrid, “semi-automated” approach to manually collect data that cannot

be otherwise captured. In Section 6.2.2, I suggest other potential strategies for

integrating manual and passive data capture methods.

31

2.2.2 Analyzing Data to Assess Health

Once a user’s personal data has been collected, it can be analyzed to derive rel-

evant health metrics and user profiles. This process is sometimes referred to as

data “mining”. The aforementioned model of personal informatics use would

consider this process as a part of “integration”: preparing and transforming the

data into information that can be reflected and acted upon, which is especially

necessary if the format of collected data is different from the format required

for reflection [284]. This process can again be driven more by the user or by the

technology.

Manual Analysis

Manual analysis involves an individual examining his or her data directly,

for instance, by sifting through daily weight measurements to pull out overall

trends or to compute correlations with other tracked personal data (e.g., phys-

ical activity, calorie intake, or numbers of hours slept). Quantified Selfers are

one group particularly known to engage directly with their raw data in order to

analyze it “by hand”. One recent study found that nearly half of its QS partici-

pants put their data into a spreadsheet tool like Excel or Google Spreadsheet that

could assist them in manually running simple statistics, and over one-third of

participants even programmed custom software to run such analyses [94]. More

anecdotally, on the Quantified Self community’s main blog and forum, one can

find numerous posts (e.g., [407]) that convey guidelines or personal experiences

about conducting manual analyses in order to identify patterns, discrepancies,

or health predictors from data like step counts, dietary intake, and medication.

32

Automated Analysis

In more fully system-driven analysis, technology automatically processes

tracked data on which it runs statistical analyses. For example, Health Mashups

[45, 488] automatically computes statistical correlations within various data

streams acquired through both manual tracking (e.g., of exercise, food, mood,

pain) and passive sensing (e.g., of calendar information, location coordinates,

weather, activity, and weight). The BeWell system [273, 275] automatically com-

putes a wellbeing score using data about activity, sleep, and social interaction.

Another approach is to use algorithms to infer health states using sensor

data. For example, Mobile Heart Health [346] applies an off-the-shelf algorithm

to a user’s real-time ECG data in order to derive a measure of stress. Given a suf-

ficient number of data points, technology can also generate a statistical model of

behavior. CenceMe [332] applies machine learning to mobile phone sensor data

(e.g., accelerometer, Bluetooth proximity to other devices, GPS, microphone) to

classify a user’s healthiness along with other attributes. As mentioned, a PHI

technology’s input, analysis, and output processes are often overlapping and it-

erative; these examples demonstrate passive sensing systems where continuous

data capture goes hand in hand with iterative analysis.

Somewhere in the middle of this manual–automated analysis spectrum,

some tools assist with converting raw data into a more human-inspectable for-

mat. For instance, a number of systems use machine learning or data mining al-

gorithms to do data aggregation or labeling (e.g., converting accelerometer data

into physical activity categories [108]). We could say a person analyzing these

pre-processed data would be doing so through a hybrid, system-supported

manner rather than through a fully manual or fully automated approach.

33

Essentially, the process of analysis serves to translate a repository of personal

data into metrics, statistics, and user models about an individual’s health. By

communicating such information through a user interface, a PHI system can

next support self-reflection, behavior change, and overall health management.

2.2.3 Delivering User-Facing Feedback

In addition to collecting and analyzing data, personal health informatics sys-

tems also aim to represent this information through legible feedback that con-

nects to a person’s real-life experiences and opportunities for behavior change.

This feedback can also be referred to as a health “intervention”.

Based on a literature review and my own design experiences, I have charted

this design space, identifying the following primary design dimensions of feed-

back supplied to a user: format, delivery medium, attentional demand, room for

interpretation, and level of personalization. Sometimes, dimensions can share

common borders or certain aspects can even overlap, often because some de-

sign choices tend to go hand in hand with each other. For example, audio is the

feedback format of a device that uses chime sounds to communicate step count

information, while the volume of that audio modulates the feedback’s attentional

demand. Further, this is not meant to be an exhaustive set of all the possible at-

tributes feedback can have — for instance, it can also be important to consider a

piece of feedback’s audience (e.g., private vs. public viewability), scope of input

(e.g., personal, family, or community level data), permanence (e.g., temporary

vs. archival), or explorability (e.g., static images vs. an interactive interface),

among a variety of other possible dimensions.

34

Still, I believe format, delivery medium, attentional demand, room for in-

terpretation, and level of personalization represent the key design levers to be

configured when deciding how information will be conveyed by a PHI tool.

Accordingly, this multi-dimensional design space can be used for prescriptive

purposes (e.g., when developing a new PHI technology, in order to identify de-

sign decisions to make along with trade-offs among those choices) or for more

descriptive purposes (e.g., to map an existing PHI technology onto the design

space in order to characterize its feedback techniques).

In this subsection, I describe each of these dimensions in more detail, in-

cluding pros and cons to different feedback strategies. In doing so, I provide

thoughts around when a designer might or might not want to pull certain levers,

depending on the goals of the system, the types of information it wants to com-

municate, and the kinds of interactions it wants to support. I also map existing

PHI tools onto these dimensions both as a way to review how extant technolo-

gies typically present feedback and also in order to illustrate examples of how a

system can actually embody a design dimension in practice.

Format

Feedback can be presented via any of the human senses: sight, hearing,

touch, smell, and taste. PHI systems often display information visually, for

instance using printed text; colored light; or charts, maps, or other visualiza-

tions. For example, the aforementioned Health Mashups system presents the

results of its statistical analyses using charts (e.g., bar, line) as well as feeds of

observed correlations expressed in natural language sentences (e.g., “You are

happier on days when you sleep more”) [45]. MyBehavior uses eating behav-

35

iors and physical activity data to generate healthy lifestyle suggestions that it

provides through lists of text-based recommendations about meals to have or

avoid and physical activity to perform (e.g., “Small walks each hour near Gar-

den Ave”) [408]. PHI systems running on devices with small screens can limit

the amount of visual information displayed; for instance, the Fitbit displays step

counts in a single number or uses LEDs that illuminate when activity or sleep

goals are met, while the companion app and website provide more detailed vi-

sual feedback in the form of charts and graphs. MoodLight similarly uses light

and color, (specifically, lamps with color-changing bulbs) to reflect an individ-

ual’s mood based on biosensor data [463].

Other systems use visual metaphors to communicate about aspects of a

user’s behaviors. For example, Gluballoon, a diabetes monitoring application

that runs on a wearable display, uses the metaphor of an animated hot air bal-

loon to illustrate a patient’s blood-glucose levels [137]. UbiFit’s phone wall-

paper displays a garden, where the number of flowers represents a person’s

amount of activities and butterflies represent attained goals [108]. BeWell simi-

larly uses a phone wallpaper (an animated wallpaper in this case, as the newer

BeWell system runs on smartphones), where the amount of fish, turtles, and fish

schools reflects the user’s wellbeing scores for activity, sleep, and social interac-

tion, respectively [273]. Along the same lines, Fish’n’Steps uses a kiosk display

to reflect a person’s physical activity levels with fish characters, where the size

and facial expressions of these fish map to goal progress and achievement [290].

Systems using metaphors are based on the premise that such representations

are more intuitive to understand and will therefore facilitate information being

more easily digested, interpreted, and acted upon.

36

PHI systems can also use audio as a format to communicate information. For

instance, the InShape footworn accelerometer system is designed to encourage

physical activity by playing pleasant chime sounds to encourage brisker walk-

ing [214]. Health apps often play a notification sound when a piece of feedback

is available or sometimes ring an alarm to remind users when it is time to log

data. Such notifications are often accompanied by tactile feedback as well (e.g.,

vibrations). More generally, such “haptic” formats use the sense of touch to

convey information. Finally, while the use of smell or taste to communicate

feedback is relatively uncommon, HCI researchers are beginning to investigate

such ideas. I can imagine a PHI system designed to promote healthy eating

might someday change the taste of a smart spoon in order to discourage a user

from eating a food it deems unhealthy.

When considering which format a piece of feedback should take, a number

of factors can be considered. First, it is important to judge whether a particular

format can adequately and sensitively convey the meaning held by the data it is

intended to represent. For instance, visualizations like charts and graphs grew

out of a scientific tradition, with these diagrams’ geometry often designed to

emphasize smooth trends. However, in my work on bipolar disorder, I found

that individuals managing that condition can experience a real disconnect be-

tween the erratic nature they associate with their mood fluctuations and the

smoothed patterns of traditional representations often used by PHI systems.

Additionally, skills, knowledge, values, culture, and various other personal

attributes can impact how an individual perceives the format of a given piece

of feedback [216]. For example, children might find the more metaphorical for-

mats more attractive and engaging [174] (e.g., a UbiFit garden of flowers or a

37

Fish’n’Steps tank of cartoon sea characters), while if a tool’s intended users are

adults, they may prefer more traditional charts. An individual with low vision

may have substantial difficulties with feedback that uses a small font, while

someone who is color blind may not get much value out of a PHI system that

uses color to communicate (e.g., MyBehavior’s red or green borders to indicate

whether a consumed food is healthy or MoodLight’s lamps that communicate

mood through colored light). Similarly, systems considering the use of audio

should consider the anticipated contexts of use, as sound might introduce user

concerns about privacy, stigma, or social etiquette.

Delivery Medium

The delivery medium refers to the device where the feedback is provided.

Mobile phones are now the most popular PHI feedback delivery medium, given

their ubiquity along with their capability to both capture data as well as output

feedback. Other common delivery mechanisms for PHI systems include web-

sites, wearables, public displays, or virtual reality. Homes or other buildings

may be able to deliver information in the future via a variety of walls, surfaces,

or other objects in one’s living or work spaces [222].

A main consideration when it comes to selecting a device for delivering feed-

back is optimizing the chance of information being received, especially if that

feedback is time or context sensitive. This makes mobile phones attractive due

to their portability and the tendency for users to keep their phones on them-

selves or nearby almost constantly, even during sleep [64]. Using mobile phones

also relieves a user from having to carry a separate device for health manage-

ment. Depending on the type of data a PHI technology is aimed at providing

38

feedback about, other types of devices may be more appealing to a designer,

however. For example, if feedback is based on physiological data (e.g., heart

rate, breathing, stress), then it makes sense to choose a feedback device that can

double as a data collection device (e.g., a wrist-worn wearable embedded with

sensors for capturing those physiological data). The intended contexts of use

can also help guide selection of delivery media (e.g., mobile devices would be

more appropriate for on-the-go usage, while a large display in the home might

be better suited for a PHI system aimed at helping parents monitor their chil-

dren’s growth patterns). Finally, practical issues of usability or affordability can

impact whether or not a device is well-suited to delivering feedback. For ex-

ample, while smartwatches were originally touted as having many of the same

benefits as smartphones when it comes to portability and easy access to infor-

mation, many of these devices received criticisms for being uncomfortable to

wear (e.g., too thick), hard to maintain (e.g., poor battery life), or impossible to

upgrade (e.g., too expensive). If a PHI system for managing a chronic condition

required frequent and long-term use of a device, then such issues would be a

major barrier to adherence.

Attentional Demand

Next, feedback can be provided using ambient, subtle cues or in more con-

spicuous ways. Ambient displays often have a focus on aesthetics and aim to

integrate well into the environment without being distracting, while overt feed-

back more directly demands that a person notices and engages with it [315]. The

phone wallpapers of UbiFit and BeWell take the ambient strategy; they require

low attentional demand and are appreciable just enough that they are noticeable

but not interruptive [216]. ShutEye similarly presents a glanceable timeline of

39

sleep hygiene recommendations on a user’s mobile wallpaper [36].

Other systems have been designed on the same premise of providing passive

cues about personal data, but information is displayed in more publicly-visible

mediums in one’s environment rather than on the private screens of personal

devices. Many have focused on communicating emotional data with the aim of

supporting awareness of personal or collective wellbeing. For example, Mood-

Light uses lamps to convey mood, as mentioned. Similarly, mood.cloud [451],

an interactive data-as-art installation, is composed of a touchscreen tablet that

collects mood data via the Photographic Affect Meter [400] along with a large

cloud-like sculpture made of LED light strands that change color to reflect the

mood of the last two dozen people who interacted with it. Other research has

pursued the similar idea of embedding an LED, whose color maps to an indi-

vidual’s bio-sensed emotion, into a variety of other delivery mediums such as

hanging lanterns and crystal charms [434].

On the other hand, some systems are more obtrusive in their feedback de-

livery in order to draw concentrated attention. A number of researchers have

used short message service (SMS) or text messages to deliver interventions or

tailored messages with behavioral suggestions (see [161] for a review), which

can be timed for delivery at points in the day when an individual is particularly

vulnerable or in need of information. More recently, PHI systems have begun

using smartphone push notifications to deliver alerts that contain reminders

to self-report data, behavioral recommendations, or updates about accomplish-

ments (e.g., an achieved step goal). Such technologies “push” feedback to the

user. “Pull” technologies would be much less demanding of attention, as they

rely on individuals proactively requesting or seeking information, for instance

40

by taking the initiative to visit a website containing visualizations about daily

physical activity progress.

Sometimes overt alerts are triggered when specific states are detected

through passive sensing. For example, when the MONARCA smartphone ap-

plication detects manic symptoms in an individual with bipolar disorder, a

screen of coping strategies is automatically launched [31]. Devices like the Fitbit

and Jawbone UP provide idleness alarms that vibrate the device or flash lights

to prompt the user to move around if stationary too long. Similarly, low lev-

els of physical activity trigger the PersuasiveSens system to send an SMS with

healthy eating or physical activity encouragements (e.g., “It is a beautiful day.

Go out and do brisk walking for 30 mins”) [85].

Such real-time performance feedback can be a powerful driver of positive

health behavior change [246]. The emerging area of “just-in-time” intervention

design is particularly interested in delivering personally-tailored, contextually-

aware, and well-timed feedback in a non-irritating way. A number of recent

systems have explored just-in-time prompting to motivate behavior change

(see [361] for a review). As just two examples, the EmoTree smartphone app

aims to help individuals avoid bouts of emotional eating [80], while the sensor-

triggered iHeal uses physiological data about stress to guide substance abuse in-

terventions [62]. Relatedly, other work has found that a person’s current affect,

stress, activity, location, and the time are important in predicting one’s cogni-

tive, physical, and social availability to attend to a delivered intervention [440].

In these ways, feedback can be designed to demand attention less or more,

depending on a system’s intentions in delivering it (e.g., enhancing peripheral

awareness versus preempting destructive behaviors with an urgent message).

41

Room for Interpretation

In mapping out these design dimensions, I observed that a tool’s attentional

demand often aligns with the degree to which that tool gives guidance for inter-

preting delivered feedback or leaves the interpretation up to the user. This room

for interpretation could also be thought about in terms of the prescriptiveness

versus open-endedness of the feedback.

On one end of the spectrum, feedback can leave little opening for interpre-

tation. For example, MyBehavior conveys dietary feedback with messages like

“Avoid large meal” written in red and with a picture of a stereotypically un-

healthy food. Little interpretation is needed (or afforded), as the directive is

clear: do not eat that. On the other hand, a system providing more descriptive,

open-ended feedback might present a bar chart of the user’s step counts across

the days of the week. Such a view would leave most of the sensemaking up to

that individual. Moving a bit more prescriptive, the system could add a dotted

“goal” line across a point on the step count axis (e.g., at 10,000 steps, a gener-

ally recommended daily target) or paint the bars of the days falling below that

number with a color associated with poor performance (e.g., red [346]). These

more explicit signals serve to guide the user toward insights the system wants

to ensure she gleans from the feedback (and, in turn, the behaviors the feedback

is intended to prescribe).

Other research has identified a similar dimension that it terms “control ver-

sus empowerment”, where a controlling technology would make decisions for

the user or automatically control the environment, while an empowering tech-

nology would present information to the user or help a user learn how to con-

trol the environment herself [222]. For example, the aforementioned EmoTree

42

smartphone app might deliver an empowering, just-in-time intervention mes-

sage to preempt emotional eating, while a more controlling technology in a fu-

turistic smart home might automatically lock the refrigerator door.

From a critical anthropological perspective, empowerment is considered the

more advantageous approach, as to mitigate concerns about manipulation, free

will, and exploitation [359]. Designing for flexible interpretation can also help

avoid users from feeling that a system is passing normative judgements about

them, which can feel insensitive and be demotivating. Designs that leave more

room for interpretation can also facilitate self-reflection, with open-ended feed-

back mediating the sensemaking experience between current self and envi-

sioned self. Such mindful introspection is valuable, as it helps a person become

more aware of her current status along with how her behavioral choices impact

health — self-driven insights that can help an individual learn how to make

more positive choices, even beyond the scope of one particular piece of feed-

back, including when technological guidance is unavailable.

On the other hand, a lack of guidance in interpreting data can sometimes be

troublesome for users — confusion and uncertainty about how to act on pro-

vided feedback can lead to frustrating experiences or potentially detrimental

misinterpretations. A designer can therefore make choices about this design di-

mension by considering characteristics of a PHI tool’s intended user or targeted

aspect of health. If it is possible and beneficial for a person to learn the roots of

health symptoms or engage in self-directed contemplation, it may be desirable

to deliver feedback that affords more interpretation; whereas, if the capacity for

self-reflection is compromised in a population of interest or if humans inher-

ently find it difficult to make sense of a given health condition, then a more

43

direct, close-ended form of feedback may be preferable. Further, as with all de-

sign dimensions I have presented in this section, a PHI system might consider

adapting the interpretability of its feedback over time or depending on context;

for example, a system might encourage self-reflection and personal growth in a

novice user, yet provide more explicit suggestions to a well-seasoned individ-

ual who has already come to understand the roots of her health condition and

simply needs to receive, say, fitness-related feedback that will optimize calorie

expenditure given the parameters of her current situation.

Level of Personalization

Personalized feedback is tailored to meet individual needs and expectations,

while generic feedback is designed for a more prototypical user. Sometimes, a

generic intervention may suffice. For instance, idiosyncratic differences might

have relatively little impact on some health outcomes; and generic feedback can

at least be more effective than no feedback, as found in a few studies on mo-

tivating feedback recipients to increase physical activity [71] or reduce weight

[229]. In some cases, personalization can even be problematic, for instance when

over-personalization produces filter bubbles or propagates discriminatory pat-

terns — issues I discuss in more detail in Section 6.2.3.

In most cases, however, interventions are more successful when they’re tai-

lored to accommodate individual differences, as I explain below. Furthermore,

generic feedback can actually be harmful in some contexts, especially when

dealing with vulnerable populations (e.g., patients with serious mental illness)

[136, 161, 337]. Therefore, I generally argue for the development of PHI systems

that tend more toward the personalized end of the spectrum.

44

Here, I discuss why individual differences matter when designing around

behavior change. In fact, a number of personally-variable factors can affect

whether an individual performs positively or negatively while pursuing a be-

havior change goal [27, 29, 296]. Similarly, the efficacy of technology-mediated

behavioral interventions can be impacted by individual differences in psycho-

logical, cognitive, demographic, or contextual attributes.

Commonly explored psychological differences typically revolve around per-

sonality since it is a well-established construct with well-validated instruments

available to measure it. Studies have found that personality helps determine a

person’s motivation and engagement [373, 374] as well as reaction to persua-

sion [235, 369], including persuasion delivered through health-related behavior

change technologies [198, 237].

Beyond personality, research shows that how much an individual is inclined

to engage in and enjoy cognitively complex activities — inclinations linked to

one’s ”persuadability” [236] — can also influence that person’s receptivity and

adherence when it comes to health and lifestyle related persuasive requests [77].

In addition, willingness to use eHealth technology (specifically, health educa-

tion and behavior change applications) [460] has been linked to behavioral risk

factors (e.g., depression) as well as demographic characteristics (e.g., age and so-

cioeconomic status) [482]. Studies also suggest that females are more receptive

to behavior change strategies in general but that there can be gender-differences

in the persuasiveness of particular strategies [381].

In addition, environmental or situational factors can also affect the efficacy

of behavior change tools [163, 164] — external factors that can vary across indi-

viduals or within the same individual over time (e.g., location, nearby people,

45

or various other contextual factors that may relate to a targeted aspect of health

[128, 443]). For example, the walkability of a person’s neighborhood can impact

adherence to physical activity interventions [253]. In addition, health behaviors

are largely conditioned by social context, and a network of social influences can

substantially aid or undermine various efforts at personal change [74]. These

environmental and social variables can also interact; for example, an exercise

support group of coworkers, together with the availability of worksite fitness

classes or equipment, can lead to increased physical activity [513].

Consequently, given that innate characteristics as well as extrinsic variables

can influence behavior and the efficacy of interventions [147, 374], it is consid-

ered important for such feedback to be personalized [46]. Personalization has

been defined as “a process that changes the functionality, interface, information

content or distinctiveness of a system to increase its personal relevance to an in-

dividual” [50]. Various studies indicate that personalized messages (e.g., based

on a person’s health goals, motives, perceived barriers, and readiness to change

behavior) improve adherence [136] and have a greater positive impact on health

behavior change than untailored or bulk messages [70, 71, 161, 437]. Further, re-

search shows that personalizing content and experiences can reduce cognitive

load, improve user satisfaction, strengthen the impact of persuasion, and pro-

mote continued use [156, 381, 492]. Given that individuals respond differently

to design cues [374], personalization also helps ensure a system’s feedback has

its intended effect.

Personalization can be implemented by providing customizable settings or

an adaptive interface [446]. A customizable system provides affordances (e.g.,

settings, interactive options) for a user to control this personalization. For ex-

46

ample, ShutEye allows users to specify custom bedtimes, which adjusts how the

system generates sleep hygiene feedback. Research associates customization

with positive outcomes (e.g., increased attachment, appreciation, and satisfac-

tion with technology [156]) as well as beneficial side-effects from the expression

of individuality [51, 473]. However, there are also disadvantages of customiza-

tion. Ironically, choices can be formulaic, shallow, and inadequate for satisfying

individuals’ idiosyncratic needs — needs into which users do not even neces-

sarily have good insight themselves [456]. At the same time, too many choices

can be overwhelming and demotivating [223], and striking the right balance

in supplying options can be challenging. An adaptive system drives personal-

ization more automatically by tailoring its functionality to a given user’s char-

acteristics or current contextual information, often by using machine learning

methods that “learn” from the user’s data. For example, MyBehavior learns

its user’s preferences in order to automatically tune recommendations about

healthy meals and exercise routes.

Altogether, these findings suggest one-size-fits-all approaches to PHI design

may not be the best approach, given individual differences can influence the

efficacy of an intervention for a given person [234] and personalized guidance

is linked to more successful behavior change outcomes [75]. Across individuals

— and even within individuals across changing circumstances — an array of

variables can differ. Such idiosyncrasy necessitates the development of tailored

tools that consider, even embrace, such differences in order to support diverse

personal needs. In turn, these personalized interfaces can minimize negative

user experiences while optimizing for motivation and positive outcomes.

47

2.3 Domain-Driven Personal Health Informatics

So how do we go about developing a personal health informatics system, comprised of

components for collecting data; analyzing that data to extract health-related variables;

and providing personally meaningful, understandable, and actionable feedback?

Central to this dissertation is the idea that domain knowledge can help drive

this PHI development process. In this section, I explain why. I also introduce the

notion of a domain-driven development framework, which supplies concrete

steps for incorporating domain knowledge into HCI practice. I describe this

framework in more detail in subsequent chapters, where I also demonstrate it

in action using my own research as a case study.

2.3.1 Defining Domain Knowledge

In essence, knowledge is an understanding of something. Organized bodies of

knowledge originated from the human need to make sense of the world around

us [285]. This understanding can be theoretical or practical in nature. That

is, knowledge can be generated through use of the scientific method in order

to produce theories and empirical evidence — known as “scientific theory” or

“formal theory”; while experiences, trial and error, or one’s own ideas about

a situation produces “situated knowledge” or “informal theory”. A domain is a

subject area that holds relevance to a given problem. Thus domain knowledge is

a body of understanding specific to that particular area — the part of the world

investigated by a specific discipline [304].

48

Domain knowledge can include concepts, definitions, descriptions, empiri-

cal evidence, formulas, algorithms, theories, taxonomies, guidelines, principles,

methods, and procedures [98]. In this way, domain knowledge provides a set

of abstractions together with a concrete vocabulary for describing, explaining,

and predicting phenomena. Domain knowledge about such information and

the skills for making sense of it can be acquired through education or experi-

ence, by learning, discovering, and perceiving. Someone who has such training

or expertise in a given domain is referred to as a domain expert.

In the context of PHI, domain knowledge typically refers to familiarity with

disciplines beyond computer science, information science, or other fields as-

sociated with HCI (i.e., training in solving problems that are typically not the

primary focus of these fields’ curricula). For example, HCI researchers devel-

oping behavior change technologies might turn to health professionals with

expertise in physical or psychological health such as clinicians, biomedical re-

searchers, nutritionists, exercise physiologists, behavioral scientists, and psy-

chologists [11]. In the context of PHI technology, envisioned users of a system

can also be considered a type of domain expert, in the sense that they possess

expertise in their own lived experiences and perceived needs associated with a

targeted health condition.

2.3.2 The Value of Domain-Driven PHI

To date, PHI research has focused primarily on behavioral theory as a source of

domain knowledge. This is a fitting choice, as behavioral theory provides a rep-

resentation of the causal processes involved in behavior change [329], though

49

other forms of domain knowledge can be a basis for interventions as well (e.g.,

disease etiology and epidemiology, theories from education and communica-

tion, etc). Advocates for incorporating such information into PHI development

argue that it can inform both system design and evaluation [206], and several

studies indicate that the application of behavioral theory is especially beneficial

in informing the content and timing of provided feedback [107, 416].

Theory-driven designs are indeed believed to be more effective at chang-

ing people’s health behaviors and attitudes than atheoretical ones [329, 518]. A

main reason is because interventions that build on existing theoretical knowl-

edge are better able to account for the factors central to behavior change, which

enables intervention efficacy to be optimized [327]. Systems informed by the-

ory are also more effective since they are better able to address the multiple

and often unintuitive personal barriers to behavior change [259]. In addition,

theory-based interventions allow the developer to avoid design choices based

on assumptions that may not only lack evidence but have even been invalidated

[329]. Further, theory-driven approaches that target specific data might better

manage and protect users’ privacy concerns.

Finally, using behavior change interventions based on theory is advanta-

geous because that theory then provides an explanation of why and how the in-

tervention works (or does not work). This understanding can in turn facilitate

the development of better theories — and subsequently better interventions,

again better theories, and so on [329].

Thus, when aiming to help a person change or maintain behavior, looking

to domain knowledge helps a researcher evaluate what is currently happening

with that individual (e.g., his or her health status, condition severity, behavioral

50

tendencies), why that may be happening (e.g., personal, environmental, or so-

cial factors), and what is likely to happen next — and how technology might

intervene accordingly to bring about positive change. Altogether, incorporat-

ing domain knowledge can therefore enable the development of better systems,

better theories, and ultimately, better health outcomes and broad impacts [206].

2.3.3 Examples of Domain-Driven PHI

As mentioned, most research in the PHI context has considered behavioral the-

ory as the primary source of domain knowledge. Researchers draw on this in-

formation mainly to inform interface design, determine target users, and guide

evaluation [206].

Prominent examples of theories commonly employed to inspire and inform

behavior change technology include goal-setting theory, the transtheoretical

model, and Fogg’s behavioral models for persuasive design. These are indeed

well-suited to the task. Goal-setting, the intentions behind goals, and goal-

directed activities have been theorized as integral to behavior change and the

attainment of said goals [28, 29, 188, 294]. Empirical research further confirms

goal-setting as a key behavior change technique [328]. The transtheoretical

model (TTM) [403] then provides a conceptual framework to evaluate a per-

son’s readiness to embark on a goal and monitor her progress through stages of

behavior change (precontemplation, contemplation, preparation, action, main-

tenance, and potentially relapse).

One of the earliest and perhaps most well-known examples of a PHI tech-

nology guided by such theories is the UbiFit system for monitoring and main-

51

taining physical activity [107]. Influenced by the TTM, UbiFit targets the con-

templation, preparation, and action stages of change [107, 108] and incorporates

goal-setting theory [295] into design choices. Fish’n’Steps [290] also drew inspi-

ration from the TTM and goal-setting theory in its similar implementation of a

personal display that reflects a user’s physical activity levels through fish char-

acters whose size and facial expressions map to goal progress and achievement.

Other PHI examples designed on premises of goal-setting include systems

to encourage physical activity [349, 350] including at the group or family level

[102] or in combination with other wellness activities such as diet management

and relaxation [177, 316]. Content analyses of health management apps provide

additional examples of recent tools implementing goal-setting techniques [302].

Some systems attempt to help users develop positive habits related to their

goals [386, 466] by incorporating psychological theory related to habit formation

[459, 530] or positive and negative reinforcement [326]. Other systems that sim-

ilarly incorporate reward and punishment schemes based on reinforcement the-

ory often use gamification elements to motivate user engagement [370], though

such approaches have met criticisms for shallowly applying those theories or

using methods that border on exploitation [53].

Lastly, Fogg has supplied a series of guidelines to promote behavior change

and corresponding design recommendations for persuasive technology [162,

164]. Fogg’s Behavior Model for Persuasive Design (FBM) [165] is comprised of

three elements: motivation (e.g., hope, fear, social acceptance or rejection), abil-

ity (e.g., time, money, effort), and triggers (e.g., inspiring videos, reminders),

which together determine whether a target behavior will be achieved. Per-

suasive technologies guided by the Fogg behavioral theories span multiple do-

52

mains related to health (e.g., physical activity [7]) and beyond (e.g., sustainabil-

ity [472] or social activism [411]).

Aside from these commonly leveraged frameworks, other theories from psy-

chology and behavioral science are sometimes utilized by HCI researchers too.

For instance, UbiFit also used Goffman’s theory of Presentation of Self in Ev-

eryday Life [185] to inform design choices about the types of control to give a

user over personal information [107]. However, as mentioned, most PHI sys-

tems focus on goal-setting, TTM, or the Fogg theories, if any theory is adopted

at all. This actually suggests an important open question about the extent to

which such systems’ developers have geniunely determined that these theories

are the most relevant for their work — or whether they are simply part of a

“theory cascade”, selecting familiar, in vogue, and commonly employed theo-

ries rather than seeking out alternative ones that might in fact be better suited

to their particular PHI projects.

2.3.4 A Domain–Practice Gap in HCI

Thus while some PHI technologies do draw on the science of behavior change,

reviews find that the majority of health applications do not incorporate salient

theory into designs [23, 206, 259, 368]. They also typically overlook other im-

portant sources of domain knowledge such as evidence-based clinical practice

guidelines, to which reviews find few PHI tools adhere [5, 302]. This means that

PHI work is often disconnected from valuable domain knowledge that could

help ensure support is being provided in a clinically, contextually, and person-

ally appropriate manner.

53

Such a domain-disconnected approach can constrain both how a system

analyses a user’s personal data as well as the interventions it provides. First,

consider analysis, the component of PHI technologies explained in section 2.2.2

that transforms personal data into health-related information (e.g., quantitative

health metrics or qualitative behavioral feedback that can be reflected or acted

upon). To compute such metrics or build models about a user’s health, PHI

systems many times use statistical or machine learning methods. However, in-

stead of defining variables or constructing features in a manner grounded in

domain knowledge (e.g., by operationalizing well-established theoretical con-

structs), many systems take a more data-driven, “black box” approach (i.e., cap-

ture an abundance of data, run a model, and let the patterns shake out).

One problem with this strategy is that it can produce noisy models [34]

that are inaccurate [546] and do not surface important underlying relationships

present in the data [196]. Researchers expressing skepticism about the reliability

of these domain-disconnected measurements argue that analytic outputs may

be incomplete or even misleading [300, 301]. Further, scholars have pointed out

other troubling human-centered aspects to data-driven user modeling, since it

reduces the user to a metric to be algorithmically optimized, relies on a narrow

set of assumptions regarding who the user is, and overall cannot fully accom-

modate the diverse, multifaceted nature of human identities [278, 320].

Therefore, these systems may not be modeling the factors most relevant to

the health outcomes of interest nor adequately supporting the idiosyncratic user

who will engage with the output of such PHI analyses. And though these

systems may be capable of identifying statistical patterns in data logs, with-

out a theoretical foundation, it is easy to misinterpret such observations or fail

54

to control for confounding factors underlying those results. Said another way,

domain-disconnected systems may simply have a veneer of robustness, where

success has been achieved in robustly fitting a model — but this display of va-

lidity holds little practical value if the appropriate health determinants are not

being targeted in the first place. Instead, using domain knowledge can improve

the relevance of variables analyzed, particularly to ensure that analysis does not

omit factors known to be central to behavior change (e.g., relevant psychologi-

cal processes, personal traits, contextual information, etc.) Ultimately, this can

improve the utility of models and the interventions they drive [196, 327].

For example, Health Mashups, as mentioned, extracts patterns from various

sensed and self-reported data streams to help a user answer questions like, “Do

I sleep better on nights after I work out?” [45]. However, daylight exposure —

something well studied in the chronobiology domain — can have a substantial

impact on sleep, even more so than exercise [60]. In this example, the system

would indeed be useful in helping an individual discover links between sleep

and exercise; but if that person were aiming to improve her sleep, it would

actually be preferable to target lifestyle changes related to daylight exposure,

which a domain-driven strategy would be equipped to handle.

In addition, a common data-driven rationalization is that the pathway to

better modeling is simply more data (e.g., in the above example, this would go

something like: if we had captured more and more data, including information

like daylight exposure, the model would have eventually been able to discover

its strong impact on sleep, even if a domain-informed strategy had not been fol-

lowed). However, this is not really a practical solution, as such exhaustive ap-

proaches are not particularly feasible for either humans or machines [292]. First,

55

it is unrealistic to expect an individual to manually capture such an expansive

amount of data. This task is difficult for technology as well, due to physical

and practical limits of computing. For instance, battery drain can plague even

“simple” continuous sensing toolkits [274, 515]. Similarly, algorithms for data

cleaning, aggregation, and compression are still not up to the task of processing

or storing the vast amount of data personal health devices can generate (e.g., a

clinical heart rate monitor produces approximately nine gigabytes of data in a

month) [337, 476]. Domain knowledge can help avoid unnecessary processing

that uses up finite computational resources, while ensuring important facets of

data are not overlooked.

Beyond the analysis phase, failing to incorporate domain knowledge into

the design of a PHI system’s interventions is similarly problematic, consider-

ing the benefits of domain-driven strategies overviewed earlier in Section 2.3.2.

But while theoretically-informed interventions are more successful at produc-

ing positive outcomes [328, 417], reviews and content analyses (conducted by

other researchers as well as myself) find that many system designers do not

draw on theory and instead rely on intuition or trial and error [373]. For ex-

ample, digital smoking cessation tools rarely follow established clinical prac-

tice guidelines for treating tobacco use and dependence, plus these tools are

typically one-size-fits-all and use game-ified elements to maintain engagement,

rather than providing self-help strategies that are tailored to the individual char-

acteristics (e.g., motivations, barriers, pace of goal progression) that behavioral

theory has established are key predictors of cessation outcomes [355]. Similarly,

PHI technologies aimed at supporting healthy sleep tend to encourage behav-

ioral modifications based on folk wisdom, rather than providing personalized

support that accounts for factors scientifically-backed as relevant to sleep [3].

56

Even in cases where domain knowledge does play a role, it is often used only

to explain behavior but not to change behavior (i.e., to inform interventions)

[329]; or, systems that claim to incorporate theory into delivered interventions

often only give theory a cursory mention and leave details ambiguous as to how

constructs were actually translated into design elements [259, 382]. As a result,

some researchers question whether these personal health informatics systems

have the ability to produce measurable, long-term behavior change [303, 317].

2.4 A Framework for Domain-Driven PHI Development

As just described, system designers taking a domain-disconnected approach

may not build in support to target the most theoretically, clinically, or person-

ally meaningful health determinants for assessment and intervention. A greater

awareness of domain knowledge can enhance the creation of PHI systems across

all phases of development, by informing choices regarding how to collect data,

model health, and provide feedback in ways that will help people more effec-

tively self-manage their wellness.

It is therefore vital that HCI’s PHI community adopt more domain-driven

approaches that base development decisions on foundational knowledge from

disciplines relevant to the aspects of health on which a given technology is fo-

cused. A main challenge in doing so, however, is that the field lacks an explicit

set of guidelines for how to actually go about domain-driven development.

57

In this dissertation, I provide a framework that specifies just such details.

In doing so, I pursue a long-term vision of end-to-end PHI systems that use

domain-driven approaches to capture and analyze a variety of pertinent signals

from personal data streams and that implement domain-aware design guide-

lines for providing tailored feedback about the self-manageable factors signifi-

cant to health.

Figure 2.1 illustrates this framework — the reusable development pattern

I have used to conduct the PHI research presented in this dissertation. Each

subsequent chapter more fully describes and demonstrates each of these frame-

work components, which at a high-level are as follows:

• DOMAIN INQUIRY. The first step in conducting domain-driven develop-

ment involves identifying an application area where personal health tech-

nology could have meaningful impacts, considering the feasibility and ap-

propriateness of such a technology-based solution, and determining the rel-

evant domain(s) from which knowledge can be drawn. Rich bodies of scien-

tific literature or empirical evidence may hold such knowledge. For exam-

ple, I draw upon years of chronobiology research to support my case study

research on sleep, daily performance, and emotional wellness. Domain in-

quiry can also involve engaging with domain experts or learning about the

lived experiences, extant practices, and perceived needs of users. Altogether,

gleaned domain knowledge can provide a deep understanding and sensitiv-

ity about the role of technology in a given health context, which can inform

subsequent development steps.

58

• DOMAIN-DRIVEN HEALTH ASSESSMENT. Knowledge gathered during

domain inquiry reveals valuable constructs to operationalize, suggests types

and sources of data likely to hold information relevant to those constructs,

helps guide analysis of that data once captured, and aids in interpreting ana-

lytic results — all activities undertaken in this phase of domain-driven health

assessment. Compared to more purely data-driven approaches, using do-

main knowledge to drive variable selection reduces the chance of emphasiz-

ing convenient items that might not be relevant and helps reduce the compu-

tational costs that come with calculating extraneous features. Domain-driven

approaches are also often more interpretable; knowing that a particular vari-

able is informative has some value, but understanding why it matters can

help modelers, designers, and theorists choose variables that are more likely

to be appropriate to their needs. I emphasize this in my case study research,

where I strive to present findings in a way that goes beyond describing what

was observed to get closer to why. Further, such insights can fuel subse-

quent design work by suggesting effective interventions that target signifi-

cant health determinants.

• DOMAIN-AWARE INTERVENTION DESIGN. An overarching goal of this

domain-driven framework is to support health self-management with end-

user tools. In this iterative phase of intervention design, guidelines are de-

termined regarding what feedback to present and how, mockups and pro-

totypes of various fidelities are built, and user models produced during the

framework’s phase of assessment are instantiated in personalized interfaces.

Throughout this cycle, participatory activities support the evaluation and

continued refinement of these domain-aware, user-centered designs.

59

PLAN

IDEATE

BUILD

REVIEW

DOMAIN-AWAREINTERVENTIONDESIGN

• Implications• Guidelines

• Mockups• Prototypes

• Evaluation• Feedback

• Gatherrequirements• Users,theory

DOMAININQUIRY

IDENTIFYDOMAIN

GATHERKNOWLEDGE

INFORMDEVELOPMENT

SELECTAPPLICATION

AREA

• Identifydatasources• Selectmethods• Definescope

• Identifyproblem• Setgoals• Assessmerit,feasibility

• Literature• Experts• Users

DOMAIN-DRIVENHEALTHASSESSMENT

COLLECTDATA

• Manualself-report• Passivesensing

ESTABLISHMEASURES

• Operationalizingconstructs• Definingvariables

MODELHEALTH

• Computemetrics,buildmodels• Interpretresults• Refinetheories

• Mapgoalstodomain– e.g.,(Behaviorchangeà behavioraltheory)(Biologicalprocessà chronobiology)(Mentalhealthà clinicalpsychology)

Figure 2.1: A framework for domain-driven PHI development, comprisedof stages for tapping domain knowledge, collecting and an-alyzing personal data to assess health, and designing user-facing feedback.

60

This framework provides a footing for an HCI researcher undertaking a

domain-driven PHI endeavor in any sector (e.g., academia, industry, govern-

ment). As such, these guidelines are intended to be a flexible “process guide”

rather than a binding set of cookbook practices. These strategies may not be

applicable literatim for every PHI solution, but my research presented in this

dissertation shows instances where they have worked successfully. The bulk of

my work is from a case study on bringing domain knowledge from the field

of chronobiology to the design of PHI technologies that support healthy sleep,

cognitive performance, and emotional wellness. This chronobiology-informed

perspective has allowed me to develop approaches to computationally assess

sleep-wake behaviors, disruptions, and daily alertness levels. In addition, I

have used this approach to guide design work for technologies that deliver per-

sonalized, biologically-aware feedback to improve sleep habits, productivity,

and mental health.

To conclude this chapter, I will reiterate that HCI researchers are increas-

ingly exploring ways technology can support personal health — work that has

made swift advances and imparted legitimate benefits to people striving to gain

personal insights, manage conditions, and improve overall wellbeing. How-

ever, this field is still young, and it continues to change with the regular arrival

of new devices, applications, and interfaces born in both academia and indus-

try. It also faces challenges related to the lack of explicit, well-established, and

domain-informed procedures for analyzing personal data or delivering effec-

tive interventions. I argue that taking domain-driven approaches can help to

realize the full potential of these tools for self-managing health, which would

lead to more successful experiences with technology and more positive health

outcomes overall.

61

CHAPTER 3

DOMAIN INQUIRY

The first step in developing a successful Personal Health Informatics (PHI) sys-

tem is gaining knowledge in the domains relevant to that endeavor. As defined

in Section 2.3.1, I characterize domains as areas of understanding that can be

brought to bear on PHI research, design, and development. Scientific theories,

methodologies, and empirical evidence could all be considered types of domain

knowledge, as could practitioner expertise and an understanding of problems

and extant solutions within that space. I refer to the practice of gaining this

understanding as domain inquiry. (See “Domain Inquiry” portion of Figure 2.1).

Domain inquiry is not a practice uniquely useful for creating PHI technol-

ogy. A software architect would not likely attempt to build, for instance, a dis-

tributed banking application or any other non-trivial system without first en-

suring enough familiarity with relevant business areas. Having an ability to

make good development choices — whether about banking products or effec-

tive health management tools — results from having sufficient understanding

of salient domains [372]. This chapter identifies and demonstrates practices to

attain that understanding.

I consider the following as core components of the domain inquiry process:

(a) selecting an application area and relevant domains, (b) gathering knowledge

from said domains, and (c) using this knowledge to plan subsequent stages of

PHI work. In this chapter’s introductory section (3.1), I describe each inquiry

component in turn, at a high level. While I provide a few brief examples for

illustrative purposes, I keep these initial descriptions at a more general, concep-

tual level to make it easier for a reader to find and scan a boiled down set of

62

inquiry guidelines. Then, to demonstrate each inquiry component in practice

(and as a way of giving the reader the specific, prerequisite background about

my case study work), each following section of this chapter (Sections 3.2, 3.3,

and 3.4) revisits the inquiry components and reports on the outputs of the in-

quiry process for this dissertation’s case study on sleep, cognitive performance,

and emotional wellness.

3.1 An Overview of the Inquiry Process

Selecting an Application Area

The initial phase of domain inquiry involves identifying a compelling prob-

lem area where a PHI solution would be valuable, followed by determining the

area or areas of expertise (i.e., domains) that are useful for informing the work

to develop that technology.

In selecting an application area, several factors can be important to con-

sider. Again, these principles are not necessarily exclusive to PHI — most fund-

ing agencies’ review guidelines would likely be along similar lines (e.g., the

National Science Foundation’s evaluation criteria for “Intellectual Merit” and

“Broader Impacts”). Rather, these are well-established motivations I feel are

important to articulate here to encourage their contemplation.

Perhaps foremost, it is valuable to think about the potential for the research

to deliver significant benefits on individual, group, and larger community lev-

els. In a nutshell: is the research worth doing? These benefits may easily align with

well-established societal goals (e.g., improving human health or quality of life,

63

preserving the environment, or empowering users especially from underserved

or vulnerable populations). Though, an issue does not have to be grandiose to

merit attention — modest research goals can certainly be worthwhile too.

Another element to consider is the potential of work in that area to meaning-

fully advance scientific knowledge and understanding about that topic. Such

understanding can in turn be practically applied or used to devise solutions to

problems (e.g., the current PHI problem being addressed or problems that are

the focus of future work in the area). A researcher may also want to keep in

mind whether an application area is well-suited to interdisciplinary study (i.e.,

that can transcend the boundaries of HCI and the target domains). For example,

the application of computing to health and medicine is a burgeoning, integra-

tive research area — one where future research activities could continue to cross

additional disciplinary boundaries, for instance, by connecting with geographic

information science to study various issues central to the spread of diseases.

When selecting an application area, it also makes practical and potentially

ethical sense to determine whether a technological solution is actually feasible

as well as appropriate in that context. As an example of infeasibility, a PHI ap-

proach dependent on a well-developed internet infrastructure would not be de-

ployable nor cost-effective in some rural areas [538]. As far as determining the

appropriateness of a technological approach, an example from my own work

on developing technology for managing bipolar disorder illustrates the impor-

tance of asking this question — and the difficulty that can come with answering

it. Namely, I found that patients can experience a number of positive outcomes

from using technology for condition management but that the same technology

can sometimes have an agitating effect on symptoms or even trigger a relapse —

64

a tension requiring careful navigation as a system designer [354] and discussed

further in Section 6.2.1. Such tradeoffs should be pondered at this point; and if

it is determined that the injection of technology in a context would come with

practical obstacles, ethical risks, or other drawbacks likely to outweigh potential

benefits, then perfectly legitimate outcomes of this stage can include seeking out

a different application area or exiting the inquiry process entirely (and perhaps

disseminating these discovered implications to not design technology here [38]).

However, if a PHI solution is deemed advantageous, we move on to identifying

salient sources of domain knowledge.

Identifying Salient Domains

After selecting an application area, we can next determine relevant domains

from which to draw background knowledge. Such information is highly valu-

able, as it provides a lens through which to study the topic of interest, increases

a researcher’s awareness of existing challenges and solutions in the space, and

helps scope and direct the work. This deeper understanding also suggests be-

fitting methods, guides potential analyses, and aids interpretation of findings.

Overall, domain knowledge thus helps a researcher more thoughtfully reflect

upon a problem and offers practical advantages as well.

One way to initially go about identifying pertinent domains is by contem-

plating the goals of the research. For example, the goal of the aforementioned

UbiFit system, a healthy lifestyle intervention technology, was to motivate indi-

viduals to change their behavior (specifically, do regular physical activity) and

sustain those changes [105]. UbiFit’s designers therefore looked to theories from

behavioral psychology that focus on behavior change [107]. Though a number

65

of such theories exist, the researchers identified two theories as particularly rel-

evant to their research goal: Locke and Latham’s goal-setting theory [295] and

Prochaska’s transtheoretical model of behavior change [402].

My own research on smoking cessation [355] provides another example.

In this case, I leveraged social media data to predict a smoker’s likelihood of

successfully quitting, and my ultimate goal was to develop a PHI system that

could use such predictive models to provide tailored behavior change support.

So, like UbiFit’s designers, I sought out behavior change theories. In my re-

view of the literature, the transtheoretical model (TTM) stood out, especially

because numerous studies over the years have tested it with smoking cessa-

tion interventions, with findings generally supporting the model and confirm-

ing that it can be used to predict smoking abstinence [129, 404], including for

a diverse set of smokers with various demographic backgrounds or smoking

behaviors [419, 500]. The choice worked well; the TTM helped guide my assess-

ment phase, providing me a conceptual framework with which to first evaluate

an individual’s readiness to embark on a cessation goal and then monitor her

progress through stages of behavior change. Given their apparent utility, these

social media based measures that captured meaningful aspects of TTM could

then be embedded into an envisioned tool for smoking-related behavior change.

A final example is provided by this dissertation’s primary case study, which

focuses on sleep, cognitive performance, and emotional wellness. These aspects

of health are essentially biological processes, so I looked to biology for domain

knowledge. More specifically, my reviews of the literature and consultations

with medical experts led to the branch of biology known as chronobiology, the

field of study concerned with the rhythms that guide biological functioning.

66

Gathering Domain Knowledge

Having identified a compelling application area and relevant domains that

can inform the PHI work, we move on to gaining a deep understanding of those

domains. Essentially, this involves eliciting knowledge from various sources

— for instance, through continued review of the literature and conferral with

domain experts (e.g., clinicians, chronobiologists, etc.) as well as by engaging

with envisioned users to ensure a system meets their needs and respects their

extant practices.

Such processes can be time-consuming and challenging. Identifying, con-

tacting, and meeting with experts who may be remotely located and available

in short supply can be particularly difficult. Still, such efforts are highly worth-

while to increase the chances that future steps move in the right directions and

that PHI outcomes will be efficacious (e.g., clinically relevant and personally

beneficial). And for many contexts, these efforts are not just advantageous but

imperative. Given PHI’s focus on health and especially for applications for vul-

nerable populations, ill-informed assessment and design work does not only

risk irrelevance or poor usability but can have lethal consequences, not to be

too dramatic. For example, delivering a sleep plan unsuited to an individual’s

biological profile can contribute to circadian disruption, which (as I will describe

later in this chapter) is associated with numerous negative health consequences;

or a misguided tool for managing a mental health condition like bipolar disor-

der could trigger a life-threatening relapse.

Overall, gathered domain knowledge can provide an informed worldview

from which to make thoughtful decisions during a PHI enterprise [372], includ-

ing to inform the work’s scope, modeling strategies, and design requirements.

67

Informing PHI Development

Having acquired background information and a strong foothold of under-

standing, we can devise an informed strategy for assessing and supporting

health. First, domain inquiry helps define the scope of this PHI work. Especially

if domain knowledge suggests that an area is associated with a high degree of

intra- or inter-individual variability, then it will be challenging to develop a

PHI solution that meets such manifold needs within and across people and sit-

uations. Instead, it may make more practical and impactful sense to initially

concentrate on a scoped subset of the population that is particularly at risk or

a select aspect of a condition that is a strong determinant of overall health. Re-

latedly, it is important to consider here whether a personalized PHI solution is

necessary or whether something more generic would suffice or even be prefer-

able for a given context. Knowledge about individual differences in a pertinent

domain can therefore further help in determining whether data can be analyzed

at an aggregate level or whether it is necessary to obtain per-person information

and develop individualized models.

Next, operationalizing constructs, defining variables, and selecting a

methodology can all benefit from a researcher’s awareness of domain knowl-

edge, which helps ensure salient concepts in an application area get appro-

priately translated into measurable factors. Again using my work on smok-

ing cessation as an example, domain knowledge from behavioral psychology

helped me realize the strong influence that an individual’s motives, mindsets,

and strategies could have on the outcome of a cessation goal [295, 402, 403].

Based on such information, I defined “Behavior Change Process” variables de-

signed to model a person’s cessation process and evaluate whether or not she

was exhibiting behavioral signals known to correlate with successful outcomes.

68

Analyzing these domain-driven variables produced striking results. I found

that individuals who relapsed were far more likely to: quit for more casual,

shallow, and unrealistic reasons; procrastinate before cessation (procrastination

is a known marker of low commitment and unsuccessful outcomes in the behav-

ior change literature [129, 402]); and choose a cold turkey strategy rather than

use more effective treatment methods that I discovered during domain inquiry.

Finally, domain knowledge helps in identifying data likely to hold informa-

tion relevant to these constructs and in guiding the capture of this data. As de-

scribed in Section 2.2.1, personal data can be collected in several ways — manu-

ally or passively, via hard or soft sensing, or through some combination thereof

— about a variety of physiological, psychological, behavioral, or contextual in-

formation. Domain knowledge can help in making choices about which strategy

to take and which data to acquire, depending on the goals of the PHI project. To

use my smoking cessation work as an example one last time, domain knowledge

helped me identify that the tweets and social network information of smokers

on Twitter were an attractive source of data for assessing key personal variables

from the domain literature (e.g., preparedness, emotional distress, temptation

exposure, social support, etc.) that a cessation intervention tool would want to

harness in predicting relapse and providing tailored behavior change support.

My strategy was also inspired by that of a growing body of recent research fo-

cused on conducting “natural experiments” [131] on social media, which have

demonstrated that the abundance of data on such sites creates a scenario that

resembles the environment of a traditional controlled experiment and can even

support causal discovery [380] during the stage of assessment.

69

In the following sections, I return to each component of domain inquiry (se-

lecting an application area and identifying salient domains, gathering knowl-

edge from those domains, and using that knowledge to plan strategies for as-

sessment and intervention) to demonstrate them in action using my case study

on sleep, performance, and emotional wellness. Furthermore, by reporting on

and documenting the outputs of my domain inquiry process in this context,

I hope that these sections will also supply constructive motivation and back-

ground information for chronobiology-driven PHI research and serve as a use-

ful knowledge resource for readers interested in doing future work in this area.

3.2 Selecting an Application Area

To begin, I explain this case study’s motivations to focus on sleep and its im-

pacts on physical, cognitive, and psychological functioning. Specifically, in the

subsections that follow, I explain why I believe this is a compelling problem

area that is ripe for novel technological solutions, I review related work already

underway, and I point out how domain knowledge can help to address short-

comings with these extant approaches.

3.2.1 Sleep, Cognitive Performance, and Emotional Wellness

Sleep plays a pivotal role in our overall health. It has a direct and substantial

impact on numerous aspects of our daily lives, from the functioning of our im-

mune system to our decision making abilities to our psychological wellbeing —

and even moderate sleep disturbance can have severe detrimental effects [277].

70

A large body of scientific evidence links poor sleep habits (e.g., too little,

too much, or erratic sleep [340]) to a range of negative consequences for physi-

cal health, including obesity and poor dietary habits [16], greater susceptibility

to illnesses like the common cold [100], and an increased risk of more serious,

chronic diseases such as diabetes [192, 540] and heart disease [22, 83]. Gener-

ally speaking, persistent sleep loss including from clinical sleep disorders (e.g.,

insomnia) is often comorbid with and indicative of numerous other debilitating

health conditions [155, 166, 306, 387].

Sleep also has well-known effects on cognitive performance [184]. Sufficient,

high-quality sleep increases productivity and work performance [103, 432]; en-

hances learning and problem solving [509]; and improves energy, alertness, and

reaction time [232, 428]. On the other hand, inadequate sleep can suppress brain

function, interfere with memory consolidation [308, 509], impede learning, and

hinder concentration [56, 132].

In addition to directly impairing performance, inadequate sleep is also as-

sociated with subsequent feelings of fatigue [268, 269]. Such reduction in alert-

ness and functional ability can pose a serious safety hazard, as it significantly

increases the risk of industrial and motor vehicle accidents [57, 132, 477]. In fact,

it is estimated that over a third of all road accident fatalities in the United States

result from driver fatigue [280], and statistics are similar for fatigue-related avi-

ation and marine accidents [367].

Beyond the physical and cognitive effects, insufficient or erratic sleep is a

major risk factor for developing various psychological problems, including de-

pression [167, 293], anxiety [180, 191], and stress [30, 334]. Sleep disturbance

may also trigger symptoms or even the onset of mood disorders like bipolar

71

disorder [189, 399] and schizophrenia [395]. In addition, poor sleep is one of the

top factors contributing to overall unhappiness [127].

Such mental health issues affect a significant portion of the world’s popu-

lation and can result in debilitating and life-threatening outcomes. As such,

mental health is becoming an increasingly pressing health care issue. Globally,

about 450 million people suffer from mental illnesses, and neuropsychiatric dis-

orders make up 4 of the 6 leading causes of years lived with disability [532] —

though, prevalence and impact may be even greater, as many cases go undiag-

nosed and the burden of psychiatric conditions is likely heavily underestimated

[298]. These problems are particularly acute for younger generations, for whom

inadequate sleep is widespread and associated with stress, suicidal thoughts

and actions, and depression [65, 208, 254, 481, 514, 529] — emotional problems

reported as so severe by a majority of students that they regularly impact daily

functioning [231]. Finally, affective health ties back to physical health, with neg-

ative affect linked to an increased risk for illness and mortality and positive

affect linked to lower morbidity and better health outcomes overall [101, 468].

An ample amount of quality sleep is thus essential for maintaining physi-

cal, cognitive, and emotional health. However, despite evidence establishing

the importance of sleep as well as public health organizations’ efforts to pro-

mote healthy sleep routines, chronic sleep deprivation is prevalent [227]. Sleep

pathologies and associated conditions resulting from poor sleep are considered

to be reaching epidemic levels, affecting millions of people around the world

[57, 306]. According to the U.S. Centers for Disease Control (CDC) and the Na-

tional Institutes of Health (NIH), sleep disorders affect 50–70 million people in

the U.S. alone, with many others likely not yet diagnosed [103].

72

According to the U.S. National Sleep Foundation’s latest report, insufficient

or poor sleep negatively impacts daily activities at least once a week for 45%

of Americans [366]. Duration, consistency, and quality of sleep could all be im-

proved. While it is recommended that adults get at least 7–9 hours of sleep per

night and younger individuals another hour or more beyond that [210], people

typically report getting an average of between 6.5–7.5 hours [365, 366]. In ad-

dition, maintaining a consistent sleep schedule can be as or more important for

health than sleep duration [448], yet the same studies find irregularity in sleep-

wake behaviors, with duration fluctuating an average of 40 minutes between

work and free days. Further, over a third of questioned individuals report that

their sleep quality is ”poor” or ”only fair” [366].

Practically speaking, the annual direct and indirect expenses of treating

sleep-related problems are estimated at $14 billion and $150 billion, respectively

[103, 512]. Mental health problems are associated with substantial financial costs

too — over $100 billion annually for mood disorders like anxiety and depression

[248]. The intangible costs of physical, cognitive, and emotional difficulties can

be far greater than any dollar amount, however, on both personal and societal

levels [347].

Poor sleep and its consequences have thus been identified as a crucial chal-

lenge that calls for researchers, including those from the HCI community, to

seek solutions in a multi-disciplinary effort [277].

73

3.2.2 Opportunities for Technology

I argue that technology can help address these challenges, for reasons similar to

those I presented throughout the introductory sections of this dissertation. Tech-

nologies (e.g., personal computing devices, the internet, and the digital trails

left from interacting with any of those systems) offer a tremendous opportunity

to study, monitor, and positively modify behaviors. In particular, technology

seems well-suited to helping people achieve better sleep, optimize cognitive

performance, and manage emotional wellness, for the following reasons.

First, the fact that personal technologies are carried and used in naturalistic

settings, throughout the course of one’s daily activities, and by a large and di-

verse population of people means that they provide a window through which

we can study phenomenona of interest in a broad, unobtrusive, and affordable

way. This ability to continuously analyze connections among experiences, be-

havioral patterns, and health indicators can help us gain a greater scientific un-

derstanding about these subjects — for example, by enabling long-term obser-

vation of how sleep quality relates to cognitive performance or by contributing

to our fundamental conceptualization of mood disorders [524].

Second, technology-based monitoring can also support assessment. Symp-

toms of poor sleep often go unnoticed, both because most people do not un-

dergo lab-based sleep tests that might help identify clinical disorders and be-

cause most people’s awareness is limited when it comes to their own sleep pat-

terns and quality [502] or what constitutes healthy sleep practices in the first

place [366]. Data from personal tracking tools could help doctors detect sleep

or mood disorders that often go undiagnosed [444] or help individuals iden-

tify lifestyle factors negatively impacting their sleep and subsequent daily func-

74

tioning [92]. Passive approaches might work especially well to promote such

outcomes while alleviating some of the burdens and biases of self-monitoring.

Finally, personal technology provides an accessible medium for delivering

feedback and interventions that can guide an individual to improve sleep habits

and manage cognitive performance and mental health. And importantly, poten-

tial users express a strong need and receptivity toward such tools [92].

It therefore seems compelling and appropriate to explore the use of technol-

ogy to measure and manage sleep along with related aspects of cognitive and

emotional health. One of the most comprehensive examinations of this topic to

date is similarly optimistic, outlining a number of design opportunities in this

space [92]. Numerous HCI researchers appear to agree (and perhaps thousands

of consumer app developers, based on my recent searches of the Android and

iPhone app stores [356]). Indeed, these areas of health have gained considerable

recent interest from the HCI community in both academia and industry.

3.2.3 Related Academic and Commercial Work

Reviews of sleep technologies identify five main platforms: mobile devices (e.g.,

smartphones and tablets), wearables sensors worn directly on the body (e.g.,

bracelets and smart clothing), environmental sensors (e.g., mattress sensors and

bedroom wall cameras), desktop or website platforms, and accessory appliances

(e.g., specialized alarm clocks, wake-up lights, and sound machines) [257]. Such

tools generally focus on measurement (i.e., tracking via manual input, auto-

matic sensing, or a combination) and/or intervention (i.e., providing feedback

about current sleep habits or recommendations for making improvements).

75

Some prominent and recent examples of research prototypes developed to

help users manually record sleep data with mobile phones include The Well-

ness Diary [318], Sleepful [277], SleepTight [93], and ENTRAIN [510]. Commer-

cial tools for manual sleep journaling exist as well, but fewer are available today

than just a few years ago, in favor of more passive tracking apps. For example,

apps like Tylenol SleepTracker or the once popular YawnLog are now defunct;

while in contrast, the passive sensing apps ElectricSleep, Sleep As Android, and

SleepCycle, which were among the first to attempt automatic sleep tracking us-

ing smartphone accelerometers, still have thousands of users on the Android

and iPhone app stores.

Academic research has similarly been shifting toward the development of

more automated and unobtrusive approaches. Much work centers attention on

automatic sleep measurement using various smartphone sensors. For instance,

the systems iSleep [200] and wakeNsmile [267] use a phone’s built-in micro-

phone to detect sounds (e.g., snoring, coughing) and body movement in order to

predict sleep phases, while ApneaApp [362] emits frequency-modulated sound

signals from a phone to detect sleep events through a sonar-like system. Sleep-

Miner [24], Best Effort Sleep (BES) [89], and Toss’n’Turn [333] have similarly

used ambient sound and light together with phone usage data such as screen

unlock events, battery status, app use, and communication logs to predict sleep

stage, quality, and duration. Such phone use data along with contextual in-

formation like time and location have also been used to predict bedtime, sleep

duration, and irregular sleep patterns [217] as well as sleep quality [225].

While not as prevalent as smartphones, wearable devices are seeing an in-

creasing rate of adoption as affordability and accuracy improve, and most wear-

76

ables now on the market provide automatic forms of sleep measurement. Wear-

able devices designed to manage diagnosed sleep disorders such as sleep apnea

have been around for years, but recent efforts have been made to design less in-

vasive tools; for instance, WatchPAT [521] asks users to attach a probe to a finger

during sleep so that the system can detect respiratory disturbance by monitor-

ing peripheral arterial tone (PAT). Regarding more general sleep monitoring,

early commercial wristbands such as the WakeMate and those by Lark Tech-

nologies used actimetry-based sensing, which is often used in clinical settings

[12], to measure nightly sleep duration and quality. The Zeo headband, another

early product for at-home sleep monitoring, provided an alternative wearable

form factor and was particularly well-received by Quantified Self enthusiasts;

however, the device was considered too cumbersome by the average consumer,

and the company is no longer in business. To explore other body-based sensors,

academic researchers have developed a prototype neck-cuff system for real-time

sleep monitoring [430]. But today, the best selling wearables for sleep tracking

are wrist-worn devices like the Fitbit, Jawbone, Apple Watch, and Microsoft

Band, which use accelerometers to measure movement and determine sleep on-

set and wake times as well as phases of light and heavy sleep.

A main reason sleep assessment has been moving away from self-tracking

and toward more automated methods is that manual tracking tools are often

burdensome since they depend on users explicitly indicating when they go to

bed and when they wake up day after day, and any inadherence hinders their

utility. Using smartphone or wearable sensors for passive tracking helps relieve

some burdens, but these approaches are not fully immune to such problems

either, as they often require users to keep the phone in bed during sleep, can

be intrusive or uncomfortable to wear, and face challenges to sensing accuracy

77

introduced by sleeping partners and pets or a user’s failure to properly charge

or configure the device.

Recent academic attention has therefore been placed on creating contactless

sleep assessment systems that use environmental sensors. For instance, using an

off-the-shelf Doppler radar sensor, DoppleSleep [409] monitors breathing rate,

heart rate, and body motion in order to classify sleep states at an accuracy com-

parable to more intrusive and expensive clinical sleep sensing techniques. The

Lullaby system [242] uses bedroom sensors and cameras to record temperature,

sound, light, motion, and pictures in order to help users identify environmental

factors responsible for interrupted sleep. Commercial contactless tools exist as

well. For example, a bedside device by ResMed uses sonar to monitor a person’s

breathing rate and associated sleep stage. Devices like Withings Aura, Beddit,

and Sleepace Reston are similar to other tracking technology in that they use

body motion as a way to assess sleep, but they use mattress-based rather than

body-worn sensors. Along the same lines, the sleep tracker Sense clips to the

user’s pillow.

Another approach has eliminated physical sensors entirely and instead fo-

cuses on leveraging technology usage patterns as a way to assess aspects of

sleep or daily behavior (i.e., soft sensing, as described in Section 2.2.1). By ap-

plying computational techniques to naturalistic data, such work aims to infer

health-related behaviors and mechanisms and in a manner that can be more

broadly, cheaply, and quickly deployed than with a physical device like those

described above [120]. Regarding sleep assessment, smartphone usage logs

about app launching, outgoing and incoming communication, and screen un-

locking have been used to predict sleep stages [333], duration [217], and qual-

78

ity [24] as mentioned earlier, though these data are often analyzed in conjunc-

tion with data from a phone’s hardware-based sensors (e.g., the GPS, battery,

and ambient light sensors). Beyond smartphone-based usage, social media data

(specifically, Twitter posts) have also been passively mined and used to detect

individuals suffering from insomnia [224].

Also relevant to this dissertation, given that parts of its case study focus

on cognitive performance and emotional wellness, are studies that connect as-

pects of cognition and mood with technology-mediated behaviors. Regarding

the former, higher levels of attention have been linked to the use of particular

types of computer and mobile apps (e.g., email, messaging, notification trays)

as well as certain usage behaviors (e.g., window switching and lapses in de-

vice use) [309, 396], while inattention has been associated with short bursts of

smartphone use [383]. Such usage behaviors along with demographic informa-

tion and contextual data (e.g., time, location, light levels) have also been used

to model boredom [309, 397] and proneness to boredom based on the types and

amount of smartphone app use [311].

Similarly, other studies suggest mood and mental health can also be assessed

using soft sensing, often using social media data in particular as a window

through which to evaluate individuals’ psychological characteristics and affec-

tive symptoms. For example, behavioral cues detected in Twitter data have

been used to identify individuals experiencing depression [388] and predict the

onset of depression [123], and similar analyses have been applied to Facebook

data in order to detect signs of postpartum depression in new mothers [122].

Beyond social media, aspects of internet use more broadly have also been used

as a marker of mental health problems. For example, excessive video viewing

79

and late-night use have been associated with symptoms of depression for col-

lege students, while frequent email checking has been identified as a signal of

high anxiety levels [263].

Personal sleep information — whether captured manually or passively and

via physical hardware or soft sensors — can then be used to help users gain

insights into their sleep-wake habits, guide them in taking corrective measures

to improve these behaviors, and ultimately help individuals achieve improved

wellness. Based on this premise, a number of systems have been designed to

provide such feedback.

Specifically, most of the aforementioned sleep tracking technologies also pro-

vide user-facing views into that data. For example, after syncing collected data

manually or automatically with a mobile device, products like the Fitbit, Jaw-

bone, and Microsoft Band allow users to browse sleep information related to

duration, onset and wake points, and entry and exit from sleep phases. These

interfaces also encourage users to reflect on how certain daily behaviors such as

caffeine consumption or water intake during the day may have impacted that

night’s sleep. Some apps provide additional information; for instance, the Fit-

bit calculates a measure of sleep “efficiency” based on sleep duration and the

amount of time a person takes to fall asleep after getting in bed, while the Jaw-

bone UP allows users to set sleep goals and provides progress reports. These

tools also sometimes incorporate social features, though sharable data is usually

limited to fitness or other activity-related information rather than sleep. A few

exceptions are BuddyClock [251], which infers sleeping state (awake, snoozing,

asleep) based on the status of a user’s alarm clock and allows a user to share that

information with a social network, as well as “Got Sleep?”, described below.

80

Finally, when sleep technology began gaining traction in the HCI commu-

nity a few years ago, Choe et al. [92] offered a forward-looking summary of

various other feedback-related design opportunities for technology to support

healthy sleep behaviors. The ShutEye smartphone app [36] realizes some of

those ideas through a glanceable wallpaper display that conveys how various

activities (e.g., drinking caffeine, exercising) will affect that night’s sleep if done

at the current time. Other ideas have been implemented in SleepTight [93],

which provides two forms of feedback: sleep summaries (visualizations of sleep

duration and quality over various time spans) and comparative relationships

among tracked data (e.g., the average sleep duration or the amount of caffeine

consumed on nights of good, neutral, and poor sleep) to help users identify

how various factors contribute to sleep quality. Similarly, the prototype app

“Got Sleep?” [438] provides natural language summaries about a user’s sleep,

suggests sleep targets, and provides features for sharing sleep scores on social

networks in an effort to motivate users through competition and playfulness.

3.2.4 Connecting with Chronobiology

These efforts toward sleep assessment and intervention are encouraging steps

toward supporting users in monitoring and improving their sleep-related be-

haviors as well as increasing our scientific knowledge surrounding people’s real

life sleep and daily activity. However, this work often makes incomplete as-

sumptions, interpretations, and design decisions, I argue, due mostly to a lack

of grounding in a highly relevant domain: chronobiology.

81

Within our bodies there are hundreds of biological clocks coordinated by

a “master clock” in our brain. These body clocks control oscillations in our

biological processes, which vary significantly, predictably, and idiosyncratically

throughout the day [79]. These fluctuations affect when we sleep and influence

nearly all other aspects of neurobehavioral functioning as well, from digestion

to concentration to mood [266].

However, the related work I just reviewed generally does not consider such

information when assessing sleep (e.g., it does not factor in the effect of light

exposure, a key factor in “setting” our biological clock). Similarly, PHI tech-

nologies aimed at improving cognitive performance are typically designed on

assumptions that our capabilities over the course of a day are steady or could

be made steady, when in reality, human biochemistry dictates that our perfor-

mance levels naturally rise and fall throughout the day [79]. To build effective

solutions in this space, researchers must account for the fluctuating nature of

performance and the behavioral, environmental, social, and biological factors

driving those fluctuations — something that current work rarely does, accord-

ing to chronobiologists [445].

While studies often report a time-of-day effect, the fact that they do not

take any biological factors into consideration hinders both the range of anal-

yses explored as well as the ability to offer deeper explanations as to why par-

ticular trends are observed. For example, the aforementioned studies connect-

ing aspects of cognitive performance with technology-mediated behaviors (e.g.,

[309, 397]) have identified consistent patterns of technology use and have asso-

ciated cognitive or psychological states such as boredom and inattention with

particular usage behaviors. However, such work does not provide satisfying

82

biological explanations of these patterns. Studies on technology use and sleep

similarly lack the domain grounding necessary to more holistically interpret ob-

servations in a way that bears in mind latent biological aspects. Research guided

by a theoretical understanding of the biology behind sleep and daily behaviors

would be more aware of the need to investigate biologically-rooted factors in

order to glean novel insights into idiosyncratic use behaviors.

Moreover, biological idiosyncrasy necessitates personalized approaches to

modeling and feedback. Biological rhythms display individual differences, with

body clock types falling on a spectrum from early to late types. This means that

the mindset behind the maxim “Early to bed and early to rise makes a man

healthy, wealthy, and wise” is erroneous for the majority of the world’s pop-

ulation [426] — yet it is a commonly held one, and it is often embedded into

design choices that encourage all users to adhere to schedules that are really

only appropriate for early types (e.g., the Jawbone UP’s “Early to Bed” goals

encourage users to adhere to earlier and earlier sleep times). As another exam-

ple, there are biologically-based individual differences in the effects caffeine has

on sleep [542], but current systems tend to present more generic feedback (e.g.,

“End caffeine consumption 8–14 hours before bedtime” [36]).

Overall, more chronobiology-driven PHI work is thus desirable for several

reasons. Studies with a restricted theoretical understanding of sleep and wak-

ing behavior provide a fragmented picture of both sleep and our broader daily

experiences; instead, a grounding in domain knowledge would enable more

holistic modeling of sleep and daily behaviors. Further, even though technol-

ogy has the potential to improve sleep, technology can also impair sleep if not

implemented properly [257]. In particular, research without biological under-

83

pinnings is unaware of the extent of individual differences in this context and,

in turn, the level of personalization necessary. In addition, interventions that

commonly target only sleep disturbances may merely be treating the symptoms

of misaligned biological clocks rather than helping to address the root causes or

work in tune with a user’s unique biological profile in the first place. A greater

awareness of our innate biological rhythms could therefore positively impact

how we design technology, enabling us to build more effective and personal-

ized user-facing tools for supporting sleep, performance, and overall wellbeing

on a broadly deployable scale.

3.3 Gathering Domain Knowledge

As explained at the beginning of this chapter, once a fertile application area and

relevant domains have been determined, knowledge from those domains can

be gathered from literature, experts, or any other identified sources that may

contribute information that can strengthen the theoretical and methodological

foundations of the research at hand.

In the following subsections, I present background on chronobiology, the

domain providing that foundation for this dissertation’s main case study. Be-

yond chronicling my application of the domain-driven framework, I hope this

section on gathered knowledge can serve as a standalone resource that is useful

for others taking on chronobiology-driven work.

84

3.3.1 Chronobiology & Circadian Rhythms

As mentioned, chronobiology is the field of study concerned with the rhythms

that guide biological functioning. Like that of nearly every terrestrial organism,

human physiology has adapted to the periodic changes in sunlight exposure

and temperature that occur as the Earth rotates around its axis approximately

every 24 hours. Over this course of a day, our biochemistry varies significantly,

causing regular changes in, for example, blood pressure, cortisol, and melatonin

levels [428]. These changes follow what are known as circadian rhythms, a term

that refers broadly to any self-sustaining diurnal biological cycle that keeps a

roughly 24 hour period (“circa”: about, “diem”: a day) [197].

These fluctuations affect when we sleep, eat, and have an impact on our

physical and mental performance and mood [266] — such as when we can swim

the fastest (in the late evening) [39], when we are most prone to heart attack (in

the morning) [348], when working memory has more capacity (generally in the

afternoon) [79], and when depressive symptoms worsen (early morning) [526].

Jurgen Aschoff, a co-founder of chronobiology and the first researcher to

investigate circadian rhythms in human beings, noted that “whatever physio-

logical variables we measure, we usually find that there is a maximum value at

one time of day and minimum value at another” [19]. His research introduced

a new basis for explaining these patterns, namely by identifying that genetic

determinants of behavioral rhythmicity (“clock genes”) are modulated by envi-

ronmental information in order to keep our body clocks running [20].

More specifically, clock genes interact with each other to generate oscilla-

tions in gene expression. This successive gene activation forms a cycle, with the

85

initial activation of a gene regulated by the last gene in the sequence, creating an

auto-regulatory feedback loop that takes about 24 hours [8]. Preciseness of this

rhythm is maintained by a process known as “entrainment”, whereby a group

of nerve cells in the brain use external information (predominantly sunlight) to

keep our body clocks synchronized with changes in our environment [421].

The biochemical processes responsible for sleep and wake activity are in-

fluenced by two mechanisms working against each other: the body’s circadian

oscillator promotes wakefulness throughout the day and determines the timing

of sleep, while its homeostatic system increases the need for sleep the longer a

person is awake and determines sleep duration, with sleep need abating during

sleep [59, 127]. Social factors such as relationships and work schedules further

affect our sleep patterns [422]. The timing and quality of sleep are thus influ-

enced by three complex factors: our circadian rhythms, our homeostatic sleep

drive, and a “social clock” based on social constraints [428].

Beyond sleep, biological clocks also influence our cognitive performance lev-

els, which naturally rise and fall throughout the day [79]. Alertness, attention,

reaction time, response inhibition, short-term and working memory, and higher

executive skills all follow rhythmic patterns [49]. In my case study research, I

focus on alertness in particular because it is considered a cornerstone of cog-

nitive performance [445], correlates with a number of cognitive functions [14],

and displays substantial variation over the course of a day [79]. In addition,

alertness deteriorates considerably after lost and interrupted sleep [377]. As

fatigue and the need for sleep accumulate while awake, an accompanying de-

crease occurs in cognitive ability and alertness [49]. These effects can become

severe. For shift workers, the increased chance of accidents and injury due to

86

fatigue is well established [418]. More generally, the impairment effects of fa-

tigue coupled with the endogenous decrements in cognitive functioning over

the day have been equated to alcohol intoxication [269]. Fatigue also hinders

meta-cognition and one’s ability to self-assess and recognize performance re-

ductions [139], which may lead people to rationalize the sacrifice of sleep and

disregard the well-studied negative impacts of sleep loss on performance, fur-

ther compounding fatigue-based performance losses [183].

3.3.2 Chronotype

Our body clocks thus produce temporal fluctuations across a range of biological

variables. In addition, these biological rhythms vary between individuals. That

is, humans display inter-individual differences in the phase and amplitude of

their circadian rhythms, from the timing of sleep-promoting hormone secretions

[426], to the duration of sleep necessary to support health both short and long

term [157, 266, 429], to the times when alertness peaks and dips each day [212].

A person’s chronotype reflects his or her unique circadian profile [109] that

manifests in such biological and behavioral differences. A common distinction

is made between early and late chronotypes (“early birds” and “night owls”)

— people whose biological clocks drive them to sleep and wake earlier or later,

respectively; however, chronotype is not binary but rather lies on a continuous

spectrum from extreme early to extreme late [426]. Chronotype is a phenotype,

meaning that it results from a person’s genetics interacting with features of her

environment like light exposure [428]. Research has found that 50% of chrono-

type features are heritable [505].

87

Demographic factors such as age, ethnicity, and gender might also influence

chronotype [426]. Children (preschool-aged as well as school-aged) are gener-

ally early chronotypes [457], transition to increasingly later types during ado-

lescence, and reach a maximum lateness around 20 years old. One’s chronotype

then begins shifting earlier once again; and in general, people over 60 years

old have an early chronotype. The shift to a later chronotype begins sooner

for females than males, which is in accordance with the general biological phe-

nomenon that females tend to mature earlier. This means that men are relatively

later chronotypes compared to females of same age for most of adulthood [426],

until their chronotypes coincide around age 50, the average age of menopause.

Chronotype can also be impacted by light exposure. Longer exposure to day-

light can shift individuals toward a later chronotype [428]; specifically, spending

more than two hours outside has been correlated with chronotype shifting an

hour later [421]. Chronotype can therefore also vary according to a person’s

geography, based on the variability of seasonal sunlight duration at a given lat-

itude and longitude [330].

3.3.3 Circadian Disruption

As described, our bodies’ circadian system plays a crucial role in synchroniz-

ing our internal processes with each other and with our external environments.

However, a number of factors, which I detail below, can disrupt an individual’s

circadian rhythms and, in turn, various aspects of functioning from sleep-wake

cycles to metabolism to cognitive ability to mood stability [429]. Unfortunately,

such problems stemming from circadian disruption affect daily life for millions

88

of people [347]. Chronic circadian disruption is linked to increased mortality

in animal studies; and while it is often confounded with sleep deprivation for

humans, research associates detrimental cognitive, behavioral, and physiologi-

cal consequences with even brief disruption [239]. Living out of sync with our

individual circadian rhythms can thus not only make us feel fatigued in the

morning or frustrated at work but can be legitimately harmful, with serious

long-term consequences for our health and wellbeing [266].

Foremost, while we are biological creatures, we are simultaneously social

beings. Our combination of biology and society is arguably what separates us

most from other animals, with our social structures dictating, restricting, and

altering our biological responses [478]. Unfortunately, these social constraints

often work against our innate rhythms. Every day, our internal circadian tim-

ings experience interference from externally determined social factors such as

work schedules and leisure engagements. As examples, standard work sched-

ules may require a late type (whose body clock wants her to fall asleep and

wake later) to use an alarm in order to rise during what is still the middle of

the night for her, biologically speaking; while an early type may stay up later

than she would naturally prefer due to evening social schedules shaped by late

types [426] but then be driven by her biological clock to wake up early, even if

the amount of sleep obtained was inadequate [425].

The result for many people is markedly different sleep and activity patterns

on work days versus free days [428]. Given that these demands manifest in

sleep and wake fluctuations comparable to jet lag, this discrepancy is referred to

as “social jet lag” since the causes are socially rooted [528]. Unlike the transient

misalignments of jet lag from travel, however, social jet lag can be chronic.

89

Earlier, I described that sleep-related problems affect millions of people and

are associated with billions of dollars in direct and indirect costs. What I did

not mention is that social jet lag is believed to be a major root of these sleep

pathologies [426]. In fact, a recent large-scale study found that more than 70% of

the population suffers from significant social jet lag, with individuals’ biological

and social clocks differing by more than one hour [424].

Social jet lag can additionally lead to a number of serious illnesses such as

cardiovascular disease, diabetes, obesity, and cancer [266]. Shift workers, who

often suffer from chronic and severe social jet lag, are more likely to experience

these illnesses compared to daytime workers [385, 469]. For young adults in-

cluding students, who also have an increased likelihood of circadian misalign-

ment [426], social jet lag can additionally increase the risk of using drugs and

alcohol [481, 528] and result in learning deficits [81].

Along these same lines, shift work and school schedules have been the most

commonly studied culprits of social jet lag [230, 377]; but in today’s world, so-

cial demands have also begun emanating from the increasingly widespread use

of digital technologies. I believe this digital connectivity may be bringing with it

additional social constraints that can further disrupt our individual body clocks

— an impression corroborated by some of my research that I will present in

Chapter 4. Potentially, the aforementioned growing prevalence of circadian dis-

ruption may therefore be partially explained by this increasing adoption of per-

sonal devices and information technologies that implant an ethos of constant

connectivity and expected availability.

Circadian disruption also has a strong association with mental health condi-

tions and neuropsychiatric illness, including anxiety, attention-deficit hyperac-

90

tivity disorder, bipolar disorder, depressive disorder, obsessive-compulsive dis-

order, and schizophrenia [202]. Recent studies indicate that circadian abnormal-

ities are not only linked to a number of psychiatric disorders but that disruption

may be directly responsible for disease etiology [271, 323], for instance, poten-

tially triggering the onset of schizophrenia in susceptible individuals [238].

My work in the area of mental health focuses primarily on bipolar disorder

(BD). One of the most prominent features of BD is its rhythmicity, including

mood episodes that cycle on an approximately regular basis [464]. Circadian

instability has been identified as a contributing factor behind the development

of BD [43], and compelling evidence establishes a link between circadian distur-

bances and BD symptoms [282], including the onset of relapse after remission

[43, 171, 399]. Other mood disturbances including major depressive disorder,

seasonal affective disorder, and sundowning (a psychological condition associ-

ated with increased anxiety and agitation in patients with dementia) are also

associated with negative changes in circadian rhythm functioning [520].

3.3.4 Traditional Assessment Methods from Chronobiology

A number of methods exist to measure circadian rhythms and disruptions. Bi-

ological markers are the most accurate, but they are also the most invasive. For

example, core body temperature is considered a robust biomarker of circadian

rhythms and circadian dysregulation [116, 499], with measurement via rectal

probes being the most accurate and widely used method in the scientific liter-

ature [345]. While consistent efforts have been made to perform less invasive

assessment through wearable devices that measure oral and skin temperature

91

[371], such approaches are less reliable [539]. Hormones in the human body are

also used as circadian biomarkers. Two of the most well studied are melatonin,

which is measured from blood, saliva, and urine [335], and cortisol [256], which

is measured in the same manner [325].

Next, given that sleep is both a reflector and modulator of our latent cir-

cadian rhythms, tracking sleep-wake patterns is useful in determining circa-

dian patterns and disruptions. The gold standard for sleep monitoring is

polysomnography (PSG); however, the required setup, controlled environment,

and specialized equipment makes PSG infeasible for longitudinal or in-situ

tracking. Instead, a wide array of studies use actigraphy, which measures body

movement through the use of a wearable sensor. A number of studies have

found sleep patterns inferred from actigraphy to be reliable and consistent with

PSG [12]. However, while actigraphy is less invasive than procedures associated

with biomarker measurement and is more practical than PSG, it still requires a

participant to wear a specialized device all day and night for the duration of the

study period, which typically lasts at least 7 days but preferably spans 14 days

or longer [384]. This condition may be less problematic for laboratory or field

studies of a short duration, but using actigraphy to track circadian rhythms over

an extended period of time and across a large population is still difficult due to

device burden and wear-compliance.

The use of biophysiological assessments such as those mentioned above are

therefore mostly limited to small laboratory studies given their invasive nature.

For more broad scale investigations, manual self-report via survey or diary in-

struments can be a more suitable approach for capturing sleep and wake pat-

terns and the underlying circadian rhythms. One of the most prominent survey-

92

based instruments for assessing behavioral manifestations of circadian rhythms

is the Munich Chronotype Questionnaire (MCTQ) [428]. To measure individ-

ual chronotype, the MCTQ includes questions related to sleep-wake behaviors

(e.g., timing, preferences) as well as daily activities (e.g., light exposure, lifestyle

details) for both work and free days. The use of the MCTQ to assess chrono-

type has been clinically validated in controlled settings against biomarkers and

actigraphy data [427]. Another self-report method commonly used for sleep as-

sessment is sleep logging via diaries, which a number of studies have utilized

to determine sleep onset, offset, awakenings, and duration; and they are often

applied as part of diagnosing and treating sleep disorders and circadian rhythm

abnormalities [547]. Comparison of journaled sleep logs with actigraphy-based

estimation of sleep behaviors generally shows reasonable agreement [297]; how-

ever, diarying faces limitations associated with self-report in general, including

non-adherence, inconsistent completion, and potentially unreliable subjective

and retrospective recall.

A number of techniques thus exist for assessing circadian rhythms and dis-

ruptions. However, the methods used in laboratory studies (e.g., that require

specialized equipment or regular blood samples) are not scalable for adminis-

tration to a large population. Subjective reports and surveys are more broadly

deployable, but these methods are not well-suited for continuous monitoring

over longitudinal periods and often fail to capture subtle details and instan-

taneous changes regarding the relationship between the circadian system, in-

dividual sleep patterns, and environmental effects. The ability to answer fun-

damental questions about sleep and circadian rhythms in real-world settings

therefore depends on developing new approaches to detect and infer behav-

ioral traits of circadian biomarkers in a low-cost, reliable, and scalable manner.

93

As a result, chronobiologists have pointed out the need for broad, in-situ data

collection methods that can record real-time data for large populations span-

ning various time zones and geographical locations [423]. In my PHI work, I

have attempted to answer this call by developing passive sensing approaches

to assess circadian rhythms and attendant aspects of health.

3.4 Informing PHI Development

In the previous section, I reviewed the knowledge that I gathered from the

domain of chronobiology. To briefly summarize, individuals have idiosyn-

cratic circadian rhythms, which fluctuate in substantial and predictable ways.

These biologically-rooted rhythms impact nearly every aspect of our function-

ing, from nightly sleep patterns to hour-by-hour cognitive performance to long-

term physical and mental wellness. In addition, disruption to these rhythms can

have a substantial negative impact on these same aspects of our overall health.

Overall, this understanding of the biology behind the rhythms that guide our

lives motivates PHI approaches to deeply consider this information when as-

sessing sleep and activity. In this section, I show how I made such consider-

ations in my chronobiology-driven case study, using gathered knowledge to

determine what to assess, for whom, and how.

3.4.1 Defining Scope

As described earlier, it can be desirable for the sakes of practicality and impact

to work on scoped solutions to particular sub-realms of a given health area,

94

rather than attempting to create a catchall PHI system. I have taken this scoped

approach in my case study. As described, circadian rhythms govern nearly all

aspects of our biological functioning including our sleep patterns, cognitive per-

formance, mental health, digestion, sensitivity to pain, athletic performance,

and much more [266]. Someday, I hope to see chronobiology-driven PHI sys-

tems that support all of these elements. However, my work to date has not

attempted a realization of this vision in one fell swoop. Rather, I have incremen-

tally focused on a few select facets of our chronobiology: sleep-related circadian

disruption, alertness performance, and disordered mood. These areas provide

breadth across physical, behavioral, cognitive, and psychological aspects of our

wellness while remaining tractable.

In the context of sleep, I focused on sleep-related circadian disruption be-

cause I was motivated by the significant negative effects disruption can have on

overall health, as described in Section 3.3.3. In the context of cognitive perfor-

mance, where our cognitive processes are complex and multilayered, I focused

on alertness because it is a primary construct in the human performance sys-

tem and correlates with a number of other cognitive functions, as described in

Section 3.3.1.

Next, domain knowledge about the degree of inter- as well as intra-

individual differences associated with the phenomena being studied can also

help in scoping. Because circadian rhythms vary from person to person and

considering that chronotype can be influenced by environmental factors and

changes over the course of a single person’s lifetime, different populations or

even the same person over time can have drastically variable requirements from

a chronobiology-driven tool. Such variability not only motivates personalized

95

approaches, but it also highlights the difficulty in achieving a PHI solution that

adequately meets such a diverse set of needs in full — a challenge we can ad-

dress by scoping and tailoring initial efforts to a specific group who stands to

benefit the most from PHI technology.

For my work on assessing sleep and alertness, I made the decision to initially

focus on a student population. This choice was not arbitrary or made out of

convenience — rather, this group is compelling to spotlight for several reasons.

Studies find these individuals suffer from chronic lost and interrupted sleep,

which can lead to poorer academic performance, increased stress, and mental

health problems [481]. At the same time, individuals of this age group are hav-

ing to manage an increased risk for developing anxiety, depression, and other

mental and emotional health problems due to academic demands and pressures

to succeed that have been mounting in recent years [231].

Further, these individuals tend to be on the later end of the chronotype

scale, which means they face a particularly high risk of circadian misalignment

[426]. In fact, this age group experiences the most severe symptoms and conse-

quences of social jet lag (which, as a reminder, is the work day versus free day

sleep schedule instability that stems from biological sleep preferences experi-

encing interference from social constraints, like the early start times of tradi-

tional school schedules) [426]. As mentioned, such disruption in younger pop-

ulations can increase the risk of drug and alcohol use [481, 528], produce cog-

nitive impairments and learning deficits [81, 115], and lead to problems with

attention and procrastination [130].

96

3.4.2 Soft Sensing for Health Assessment

Having decided what aspects of health to target and for whom, I next deter-

mined the assessment strategies I would take. In Section 3.3.4, I reviewed

how chronobiologists traditionally measure circadian rhythms. This knowl-

edge helped me decide whether I could apply these methods in a PHI sys-

tem or whether I would need to devise an alternative assessment strategy. I

chose the latter. Given the identified shortcomings of these traditional means

of assessment and chronobiologists’ expressed need for more scalable and in-

situ methods, my case study pursued a passive sensing approach to assessing

sleep and alertness. More specifically, I took a soft sensing approach, leverag-

ing technology-mediated digital traces to model individual behavior and infer

these chronobiology-related aspects of health.

Generally speaking, a soft sensing approach has a number of advantages

over the traditional assessment methods. First, this type of passive sensing is

more affordable, as specialized and expensive equipment is not required (only

the technology that individuals already possess and use). It also provides more

granular data given that collection is automated, which means it can be per-

formed continuously and over long periods of time. Further, soft sensing is less

susceptible to the self-reporting biases described earlier and is also less burden-

some and intrusive — although asking someone to grant access to large volumes

of personal and potentially sensitive information is intrusive in another sense.

To what extent this is considered intrusion is an individually-variable and open

question; in my case study work, at least, people were receptive to the approach.

For data, I focused on soft sensed usage logs for smartphone apps and social

media. The increasingly large and diverse user bases of these technologies were

97

a main draw in choosing their data. As just mentioned, a key advantage of soft

sensing approaches is that they enable real-time, continuous, and longitudinal

monitoring. However, this is only the case if the individual being monitored is

a regular user of the technology doing the passive monitoring; otherwise data

will be too sparse or skewed. The choice of these data was therefore motivated

by the fact that mobile and social technologies have reached deep penetration

within the target population of young adults. Specifically, the most recent statis-

tics report that smartphone ownership has already reached 86% for U.S. “mil-

lennials” in the 18–34 year old age range, who are also the heaviest and most

habituated users: 52% claim they could not last more than 24 hours without

their phones; 90% sleep with or next to their phones; 54% report checking their

phones “almost constantly”; and 90% check at least once an hour even during

social situations such as meals, meetings, and conversations [64, 462]. Similarly,

over 90% of 18–29 year olds in the U.S. hold at least one account on and regu-

larly use social media, and that percentage is still rising [142, 281].

Beyond the broad reach of these technologies and their deep embeddedness

into daily life, smartphone and social media data seemed promising for other

reasons too. As I overviewed in more detail in Section 3.2.3, a growing body of

research has had success in leveraging these or similar data to model traits and

behaviors related to sleep [24, 217, 224, 333] as well as attention and boredom

[309, 311, 383, 396, 397], which are aspects of cognition related to alertness.

I had a few additional reasons for selecting social media data in particular.

One, I felt confident working with it, given favorable past experiences I had

with it in the context of PHI assessment, including my work on smoking cessa-

tion described at the beginning of this chapter [355]. Second, social data seemed

98

appropriate for the case of assessing sleep-related circadian disruption, given

that such disruptions often stem from factors that are social in nature [528]. The

traditionally identified social constraints responsible for disruption are encoun-

tered in an offline context (e.g., school timings or evening social schedules), but

my thinking was that technology-mediated social interactions might be a con-

temporary source of disruption too, as I mentioned earlier.

Further, beyond my sense that social media data was familiar and fitting,

these data were also attractive because they included textual content naturally

expressed in a person’s own voice (e.g., in the form of Facebook posts, mes-

sages, etc). Advances in psycholinguistics have shown the effectiveness of using

the text people write to evaluate various behavioral, cognitive, and psycholog-

ical attributes, with analysis tools such as the Linguistic Inquiry Word Count

(LIWC) or Affective Norms for English Words (ANEW) well validated on such

assessment tasks [63, 391]. Further, researchers have found strong correlations

between language use and various aspects of health (e.g., with physical health

[78], cognitive processes [390], and emotional wellness [436]). Linguistic analy-

sis has also been used successfully in this way on the types of short texts typi-

cally found on social media specifically [120, 122, 265].

Altogether, this chapter has demonstrated how inquiry into domain knowl-

edge can support decision-making during PHI development. Specifically with

respect to this dissertation’s main case study, I have drawn from chronobiology

to devise a plan that can be used in the framework’s next stage of health assess-

ment — a soft sensing approach for studying sleep-related circadian disruption

as well as cognitive performance in young adults. The next chapter, “Domain-

Driven Health Assessment”, presents that assessment work.

99

CHAPTER 4

DOMAIN-DRIVEN HEALTH ASSESSMENT

In the previous chapter, I overviewed a process of domain inquiry: selecting a

compelling area for applying PHI technology, gathering salient domain knowl-

edge, and using that knowledge to inform subsequent stages of PHI work, in-

cluding a data collection and analysis strategy. In this chapter, that analytic plan

is executed. (See “Domain-Driven Health Assessment” portion of Figure 2.1).

To demonstrate this stage of the framework in action, I continue to use my

case study on developing chronobiology-driven PHI. Specifically, in this chapter

I present two experiments undertaken to investigate how technology-mediated

digital traces can be leveraged (i.e., passively collected and analyzed) in order

to assess idiosyncratic biological rhythms. In the sections that follow, I first

overview the protocols of these experiments, participant characteristics, and

data collection procedures. I then present findings from each experiment, along

with chronobiology-guided interpretations of their results. I conclude this chap-

ter with a discussion of these experiments, their limitations and future opportu-

nities, and broader implications for the domain-driven framework.

4.1 Method

In Section 3.4.2, I explained the merits of using passively sensed data from

smartphones and social media in the context of assessment. In Section 3.4.1, I

also justified a decision to scope collection and analysis of these data to a popu-

lation of young individuals, who could particularly benefit from chronobiology-

100

aware PHIs and for whom this sensing methodology is particularly apt. In the

following subsections, I overview how these and other relevant data were col-

lected from samples of this target group recruited to participate in two experi-

ments, which I refer to as Experiment 1 [352] and Experiment 2 [353].

Both experiments examined the interplay between biological rhythms and

technology use. Specifically, Experiment 1 explored how social-sensor data

can be leveraged to detect sleep-related behaviors and circadian disruptions,

and it took preliminary steps toward analyzing the impact of inadequate sleep

on cognition and mood. Digging deeper into daily functioning, Experiment 2

built on chronobiology about cognitive performance rhythms in order to ex-

plore and interpret a number of relationships among smartphone use, alertness,

sleep, and latent biological traits. Findings from both experiments contribute to

chronobiology-driven assessment by identifying ways to capture and analyze

usage patterns in order to passively monitor idiosyncratic biological rhythms.

4.1.1 Participants & Procedures

For both Experiment 1 and Experiment 2, public mailing lists, recruitment por-

tals, and snowball sampling were used to recruit participants who were in the

target age group, had been using smartphones for at least six months prior to

the beginning of the experiments, and were willing to participate for the exper-

iments’ full durations.

Experiment 1’s sample consisted of 9 participants (7 males, 2 females) aged

19–25 years old, and the study lasted 97 days from November 22, 2013 – Febru-

ary 26, 2014. Given that this experiment was interested in exploring how social

101

interactions and socially-defined demands impact circadian patterns, this study

spanned three phases of student life with varying scheduling constraints and

social environments: end of Fall semester (34 days), Winter break (24 days), and

start of Spring semester (39 days). All participants had standard class schedules,

except for one person who had an internship and attended no classes during the

Fall semester. Experiment 2’s sample consisted of 20 participants (7 males, 13

females) aged 18–29 years old, and the study lasted 40 days from February 13 –

March 24, 2015.

To onboard participants, they were invited to the lab, where I or a colleague

explained procedures and installed, tested, and demonstrated the experiments’

data collection tools on their phones. At the end of the experiments, participants

were compensated based on the number of completed sleep diary entries, the

amount of successfully logged data, and the number of conducted interviews.

(I describe all these data further in the following subsection). All collected data

were anonymized and encrypted, and Cornell’s Institutional Review Board ap-

proved all procedures for both experiments.

4.1.2 Data

Survey Measures

Chronotype. As previously described, an individual’s chronotype reflects his

or her unique circadian rhythms, which underlie numerous biological and be-

havioral processes including sleep and daily performance. To measure chrono-

type, participants took the Munich Chronotype Questionnaire (MCTQ) [428]

during recruitment. The MCTQ includes questions about sleep-wake behaviors

102

(e.g., timing, preferences) as well as daily activities (e.g., light exposure, lifestyle

details) for both work days (days on which alarms are used) and free days (days

without an externally-imposed work or school schedule — typically weekends)

[424]. The use of the MCTQ to assess chronotype has been clinically validated

in controlled settings against biomarkers, actigraphy data, and sleep logs [427].

To provide a quantified, comparable representation of chronotype, the

MCTQ estimates chronotype based on the halfway point between sleep onset

and waking on free days [528] (MS F). Previous studies have found this mid-

sleep point to be the best phase anchor for biochemical indicators of chronotype

[483]. MS F is corrected (MS FS C) to account for longer sleep durations taken on

free days; that is, except for extreme early chronotypes, most people accumulate

sleep debt during work days and then compensate (if possible) by oversleeping

on free days [428]. Thus, chronotype is a continuous variable quantified as:

MS Fsc = MS F − 0.5 (S DF − (5 ∗ S DW + 2 ∗ S DF)/7)

where S DF and S DW are sleep duration on free days and work days, respec-

tively, and (5 ∗ S DW + 2 ∗ S DF)/7 provides average sleep duration over a week.

Figure 4.1 shows chronotype according to MS FS C for each participant in Ex-

periment 1, and Figure 4.2 illustrates the distribution of chronotypes for partic-

ipants in Experiment 2. Both figures’ early-late key is based on an established

early-late spectrum for a general population [426].

For such a general population, average MS FS C falls closer to the 4:00–6:00

range. Here, participants trended later (average MS FS C = 5:34 in Experiment

1 and average MS FS C = 5:56 in Experiment 2). This was expected given their

ages; and considering the narrow chronotype range typically associated with

this age group, the samples actually provided a relatively wide variability of

103

Extreme Early

Moderate Early

Average

Moderate Late

Extreme Late

P1P2

P3

P4

P5 P6

P7P8

P9

9 am

8

76

5

4

3 am

P1 (6:59) P2

(6:41) P5 (6:38) P3

(6:12)P9

(5:43)

P6 (5:18)

P7 (4:54)

P8 (4:41)

P4 (3:02)

Figure 4.1: Chronotypes of participants in Experiment 1.

chronotypes [426, 428]. To verify that these samples were representative of my

population of interest, I also administered the MCTQ to over 200 additional

students for comparison. Generating an age and sex matched random sample

from that N=281 large survey gave a mean MS FS C = 05:46, which is indeed

similar to the average MS FS C of participants in both Experiment 1 and 2.

1 2 3 4 5 6 7 8 9

35

0

5

10

15

20

25

30

Chronotype

% o

f Sam

ple

Larks Owls

extremeEarlytype

moderateEarlytype

slightEarlytype

Normaltype

slightLatetype

moderateLatetype

extremeLatetype

Figure 4.2: Distribution of participant chronotypes in Experiment 2.

104

In determining a fair “early” versus “late” threshold for a particular popu-

lation of interest, chronobiologists consider the chronotype distribution of that

group [425]. Thus I followed prior work [422] and treated MS FS C ≤ 5:00 as

early and MS FS C > 5:00 as late, given participants’ young age range and corre-

sponding later-skewed chronotype distribution [203]. This split created groups

acceptably balanced in size and also provided the highest level of agreement

among MCTQ-measured chronotype, self-perceived lateness reported during

interviews, and earliness/lateness assessed via the Morningness-Eveningness

Questionnaire [213], which participants also completed during recruitment as

an additional check on their early/late classifications.

Personality. Given that previous research has found associations between

Big Five personality dimensions [111] and individual differences in circadian

rhythms [6], personality was captured as a control variable using the Big Five

Inventory (BFI) [226]. All personality factors showed good internal consistency

within participants (Cronbach Alpha between 0.74 – 0.89), and correlations be-

tween participants’ chronotype and personality were similar to findings from

prior studies [6, 511].

Daily Self-Reports

Sleep Logs. Throughout both experiments, each participant maintained a

once-daily online sleep diary. Guided by sleep diaries from prior work [127],

questions asked about the previous night’s bedtime, number of minutes to fall

asleep, number of wakeups during the night, details about any experienced

sleep disruptions, wake time, perceived feelings upon waking, presence and

duration of groggy feelings after waking, and overall alertness and sleepiness.

105

Participants received a reminder each morning to record this information;

reminders were sent via an email in Experiment 1 and via a mobile notification

in Experiment 2. Compliance rates were 76% and 73% in Experiments 1 and 2,

respectively. To ensure data quality, any retrospective entries were discarded

and only those that recorded the previous day’s sleep were retained. Though

tedious to collect, self-report journaling has been validated by prior studies as a

reliable approach for in-situ measurement of per-night sleep [58, 333, 389], and

is considered less intrusive than body-based sensors such as actigraphy.

Alertness EMA. Since Experiment 2 was interested in more deeply investigat-

ing circadian rhythms related to daily performance (specifically with respect to

relationships among alertness, sleep, and smartphone use), participants in this

study also completed a brief ecological momentary assessment (EMA) of alert-

ness on their phones. Specifically, the study used a three minute version of PVT-

Touch [243], a validated smartphone-based psychomotor vigilance task (PVT)

for objectively measuring alertness. The PVT, which is sensitive to changes in

alertness [35] and is immune to practice or learning effects [270], measures alert-

ness by displaying a visual stimulus and recording the elapsed milliseconds be-

fore a tactile response.

The alertness EMA was delivered four times daily at the start of time win-

dows defined by prior work [1] for morning, afternoon, evening, and late night

to increase the breadth of coverage across the day. Participants could complete

the assessment anytime within the time window, providing further variation in

the collection times. For the morning, afternoon, and evening windows, the av-

erage compliance rate was over 75%; the late night window overlaps with sleep

[230] and thus had an expected lower coverage of 14%.

106

Following an established formula from the chronobiology literature, I com-

puted alertness performance for a given session according to its percent devia-

tion from that individual’s baseline, where individual baseline was computed

as the mean reaction time across all test sessions, after removing false starts and

outliers 2.5 standard deviations above or below the mean [504].

Usage Logs

A core component of both experiments was technology usage data.

Computer-Mediated Communication (CMC) Data. In Experiment 1, participants

installed a smartphone application a colleague developed to run in the back-

ground and collect usage data. My analysis of this data focused on the probes

for technology-mediated social interactions: phone calls, text messages, and

social media app usage. For Experiment 1, I also requested participants’ per-

mission to download their Facebook friend data along with their logs of status

updates, posted comments and “Likes”, “Ask a Question” posts, location check-

ins, and outgoing private messages. I refer to all these Facebook data as “posts”.

Interested in participants’ Facebook interactions with other users, I filtered out

system-generated posts (e.g., tagged photo alerts). I focused on Facebook since

it was the most popular social network used by all of the study’s participants,

but it could be desirable in future work to incorporate data from additional so-

cial media platforms to further verify and compare results.

Smartphone Application Use Data. In Experiment 2, participants also installed

the robust AWARE framework, which provides fine-grained sensing of mobile

application usage [159]. I captured the timestamped log of any app coming

to the foreground (e.g., app launches or switches), which I refer to as “usage

107

events”, along with the duration of use. Most of my analyses focused on the

instances of app foregrounding since prior work shows this is an informative

portrayal of usage behaviors [54, 66, 199, 228, 454] and is less prone to measure-

ment errors than metrics like duration [134]. This also facilitated comparison

with related research using the same metric. As they are not indicative of user

behavior [54], I disregarded background apps with which the user did not in-

teract as well as system-generated activity such as automated notifications.

To categorize participants’ logged applications, I followed prior research [54]

and used the app’s developer-specified category in the official Android applica-

tion market, Google Play, where each app was associated with a single category.

I filtered out the Tools category since its apps related to launcher processes,

system activities, or settings, which were either not user-originated actions or

did not provide the sorts of insight I desired into individuals’ self-driven usage

behaviors [54], with the exception of clock and weather apps, which I relabeled

into a new Time & Weather category. I also filtered out Health & Fitness apps

since they comprised less than 1% of usage events and were used by a minority

of participants. Future work could do well to specifically recruit Health & Fit-

ness app users in order to explore the relation between app usage and physical

performance, which also exhibits well-known circadian fluctuations [140].

Again following prior work [54], I separated web browsers and email apps

from Communication into more fine-grained Browser and Email categories. Then,

to further facilitate analyses and because I hypothesized that circadian rhythms

of cognitive performance might be most strongly reflected by productivity and

entertainment-oriented application use, I consolidated a number of related apps

into higher level categories: Entertainment contains apps originally categorized

108

Category Example Apps # of Apps # of Usage EventsBrowser Chrome, Firefox 10 17683Communication Facebook Messenger, GroupMe, Phone, SMS 33 32906Email Gmail, Inbox 3 5142Entertainment Clash of Clans, Ebay, Netflix, YouTube 60 9863Productivity Evernote, OfficeSuite, To Do Reminder, Piazza 47 3146Social Media Facebook, Twitter, Yik Yak 14 27693Time & Weather Clock, Timely, Weather Channel 12 1702

Table 4.1: Categories of applications used by participants in Experiment 2,with examples and amounts of applications and usage events.

as Entertainment, Games, Media & Video, Music & Audio, Photography, or Shopping;

and Productivity contains apps originally categorized as Productivity, Business,

Education, or Finance.

Manually inspecting all apps within each category, a colleague and I inde-

pendently verified, discussed, and came to full agreement that similar kinds of

apps were folded together and that each resultant parent category fairly repre-

sented its contained applications. These categories are shown in Table 4.1, along

with information about the unique number of apps participants used from each

category and the total number of usage events.

Interviews

Both experiments included interviews with participants. In Experiment 1,

I or a colleague interviewed each participant three times: an initial interview

upon recruitment, a second interview at the end of the Fall semester prior to

the start of Winter break, and a concluding interview at the end of the study.

Participants were asked to discuss various aspects of their phone use, especially

109

in the morning and at night and across various contexts, as well as motivations

and any perceived patterns to their social media use specifically. In Experiment

2, I or a colleague interviewed each participant upon recruitment and again

upon completion of the study. Questions asked about sleep-wake behaviors

and perceived connections among one’s alertness, fatigue, sleep, and time of

day. We also asked about technology usage habits, including thoughts about

technology’s impact on alertness, fatigue, or sleep. Finally, we asked about ex-

periences with productivity software and reactions to ideas for chronobiology-

aware tools.

These interviews provided the opportunity to verify assumptions, validate

analyses against a self-reported “ground truth”, and seek explanations behind

participants’ observed technology-mediated behaviors. Interview data were

qualitatively analyzed using thematic analysis [61].

In the following sections, I report on the findings from both experiments,

contextualizing quantitative results with insights and representative quotes

from interviews where appropriate.

110

4.2 Experiment 1 Findings

4.2.1 Daily Rhythms in Sleep and Social Technology Use

To begin, I analyzed data from phone probes and social media logs (together

referred to as “CMC”) to gain a sense of participants’ typical usage trends. I also

compared my observations to results from prior studies (consistently finding

close alignment) in order to further support that the experiment’s small-scale

sample was representative of college students more generally.

Specifically, I observed the daily usage trends shown in Figure 4.3. Usage

was heaviest in the late evening, until about 11pm. Levels of social media app

use and Facebook posting activity in particular continued slightly later until

around 1am. These observations align well with prior studies on CMC use,

which find that Facebook usage increases through the evening until around

midnight [310], that social mobile applications have the highest probability of

being used from 9pm to 1am [54], and that text messaging frequently occurs late

at night and causes later bedtimes [495]. Following this CMC use, sleep diary

8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5

Overall

Phone Calls

Text Messages

Social Media Apps

Facebook Posts

14

7am pm 12am

12

10

8

4

6

2

Mea

n #

per d

ay

Figure 4.3: Daily trends in participants’ average CMC-based usage.

111

entries indicated participants go to sleep within an average of 49 minutes; prior

research similarly finds sleep occurs within 60 minutes of computer use for 60%

of 19–29 year olds [364].

Individuals in this 19–29 age range are known to go to sleep later than any

other age group, and adolescents in particular tend to delay their bed and wake

times as well as suffer from decreased sleep length and increased sleep irregu-

larity [113, 310]. Indeed, the average sleep onset of participants in Experiment 1

ranged from 1:36–2:14am (depending on weekday or weekend, semester or va-

cation) — quite late timings, likely since they are on the younger side of this age

range. Additionally, I found less than half (49.1%) of reported sleep durations

to be 7 or 8 hours, and 23.3% of reported durations were 6 hours or less. These

findings are close to those observed in prior studies [347] and indicate a con-

cerningly high incidence of insufficient sleep among participants. I also found

15.2% of sleep durations to be 10 hours or more, which is further troubling given

that exceedingly long as well as short sleep durations are detrimental to phys-

ical and mental health and are associated with a range of problems related to

academic performance, reckless behavior, and substance abuse [209].

Researchers have suggested that such sleep inadequacy may in part be due

to increased usage of the internet and social media [249]. Experiment 1 found

similar results — that social media may not only reflect but also modulate de-

layed sleep onsets. Specifically, on nights when participants used social media

apps and posted to Facebook after 12am, they reported an average of 34 minutes

less sleep. Using participants’ sleep diaries, I also compared each night’s num-

ber of reported sleep interruptions to the timing and amount of social media use

the prior day. I found that for nights during which participants experienced one

112

or more sleep interruptions, they used social media nearly twice as much the

day before (1.8 times more on average; Wilcoxon sign-rank test, p < .001). Di-

ary entries suggested that such behavior produced feelings of tiredness, as my

analysis showed that late-night social media use was associated with reports of

feeling “fatigued” as opposed to “refreshed” (χ2 = 10.21, d f = 1, p < .05). Face-

book updates from the following day sometimes expressed similar exhaustion

— as three examples: “Super sleepy”, “I am exhausted”, and “So tired and really

want another hour to sleep”.

During interviews, many participants explained that CMC use had become a

routine part of bedtime habits. For example, “I do my before-sleep routine, get into

bed with my phone, spend about fifteen minutes on Facebook, then set my alarm, put the

phone under my pillow, and am asleep”. The exact ways in which technology was

incorporated into late night behaviors varied across individuals depending on

lifestyle aspects or chronotype traits. For instance, all participants in relation-

ships reported using CMC to communicate with partners just before bed. Late

types noted using social media as something to do when unable to fall asleep

(due to their late biological clock), while as expected the early type participant

disagreed, “People usually keep me up not technology”, referring to evening social

schedules, which are shaped more by late types [426]. Most participants also

expressed that social media keeps them up longer than planned, for common

reasons such as “endless scrolling” social feeds that make them “feel like an

addict, obligated” to “need to know what’s going on”.

Based on these findings that CMC use related to sleep characteristics such as

length and quality, I next explored leveraging this data for sleep sensing.

113

4.2.2 Leveraging Social Data for Sleep Sensing

Inferring Sleep Onset, Duration, and Waking

I first attempted to infer sleep events from CMC patterns by implementing

the sleep-inference algorithm built on screen on/off patterns presented in prior

work [3]. I instantiated the algorithm using participants’ phone probe data, so-

cial media app use logs, and Facebook posts in order to model sleep events ac-

cording to the longest nightly gaps in usage. I pre-processed these social-sensor

inputs to filter usage events before 10pm or after 7am, which do not normally

coincide with sleep periods for non shift workers [230]. I also eliminated any

usage events with a duration of less than 30 seconds, which are likely due to

automated phone notifications rather than active user interactions [3].

Table 4.2 presents the accuracy of this sleep duration inference compared

with the screen on/off approach and with participants’ ground truth sleep di-

Social Screen Ground TruthData On/Off Diary

P1 8.44* 8.54* 8.13P2 7.64* 8.09 7.45P3 8.21* 8.33* 8.15P4 7.53* 8.02* 7.25P5 6.11* 5.44* 6.12P6 7.15* 7.17* 7.13P7 7.63 7.16* 7.14P8 7.38* 7.30* 8.14P9 7.48 5.42 6.25

Table 4.2: Average sleep duration for each participant according to social-sensor data, screen on/off data, and ground truth sleep diarydata. (* denotes inferences that fall within 95% confidence inter-val based on diary self-reports, p < .01).

114

aries. Results show my technique’s reliability, which achieved an average differ-

ence of only 23 minutes between socially-sensed and self-reported sleep dura-

tion. This prediction was more accurate than from screen on/off alone [3]; and it

also managed to outperform more complex algorithms based on environmental

factors such as light, movement, and sound as well as phone locking and charg-

ing events [89]. My approach is thus desirable for a few key reasons. First, this

technique is as reliable yet more unobtrusive and computationally lightweight

than those built upon frequent momentary assessments (EMAs), heavy instru-

mentation, or the use of wearable sensors. In addition, by leveraging web data,

it is able (unlike approaches based solely on mobile sensor data) to continue

capturing signals about a user’s behavior even if she is interacting through a de-

vice other than her personal phone such as a tablet, desktop computer, friend’s

device, or public computer.

My approach did overestimate sleep when the stop and start of CMC use

did not precisely adjoin sleep onset and wake, respectively. By incorporating

an error term to the calculation of sleep duration per participant (based on

chronotype and individual differences in pre-bed and post-wakeup CMC us-

age learned from the study’s first week of data), I was somewhat able to correct

for this non-usage gap, and more complex learning could further improve ac-

curacy. Conversely, there was sometimes an underestimation in sleep duration

when notifications were mistaken as active usage. By incorporating a threshold

for minimum usage duration, I attempted to filter out such device-generated

events, but more sophisticated instrumentation could further help eliminate

such misinterpretation of phone events that are not indicative of genuine user

activity.

115

Interview data allowed me to uncover other points of failure and opportu-

nities for improvement. For instance, one participant described pre-bed phone

use as a common tendency, identifying watching movies and using Twitter as

typical nightly activities; and she also noted normally checking email and texts

upon waking. Similarly, another participant explained that morning phone use

involved weather and calendar checking, and he discussed playing video games

before bed but explained that he did so on a desktop computer rather than the

phone. Thus incorporating into sensing both additional forms of social data

(e.g., Twitter, email) as well as broader non-CMC usage data (e.g., app logs,

web histories) and from across additional devices would be straightforward

next steps toward better sleep-event estimations.

This ability to infer sleep onset and waking introduces opportunities to as-

sess a number of other chronobiology variables that can be derived from this

information, such as chronotype, which is quantified using sleep duration and

midpoint (both of which can be computed from onset and waking). Given the

scope of this experiment, I focused on using this sleep sensing technique to as-

sess circadian disruptions related to social jet lag (defined in Section 3.3.3).

Assessing Social Jet Lag

As I discussed earlier, social constraints can result in later sleep onsets and

earlier required wake times that are in opposition to our own internal timings.

Figure 4.4 shows the average social-sensed sleep duration on work days and

free days for each participant — and illustrates the discrepancy between the

two. Duration is calculated as the amount of time between sleep onset and

waking [377]. (Note that participants’ work days were Monday through Friday,

116

9

8

7

6

5Sl

eep

Dur

atio

n (h

ours

)

Work daysWeekends

4

3

2

1

0 P1 P2 P3 P4 P5 P6 P7 P8 P9

Figure 4.4: Average sleep duration on work days and free days revealthe “scissors of sleep” phenomenon — a discrepant pattern ofsleep on work days versus free days that is reversed for earlyand late chronotypes.

and their free days corresponded to Saturday and Sunday; but generally speak-

ing, work and free days do not necessarily have to coincide with the standard

workweek and weekend days).

Chronobiologists refer to this reversed sleep pattern for early and late

chronotypes on work and free days as the “scissors of sleep” [422] — a phe-

nomenon my sensing approach was able to reveal. Specifically, these results

demonstrate that for later chronotypes in the sample (e.g., P1, P2, P5), sleep du-

ration was systematically shortened on work days, which led to accumulated

sleep debt that was then compensated for by sleeping more on the weekend.

This same effect has been observed in past research [428]. Excluding the early

chronotype participant (P4), participants slept an average of 67.8 minutes more

on weekends. In contrast, P4 exhibited precisely the opposite pattern. For this

117

individual, longer durations of sleep happened on weekdays while sleep was

shortened on the weekend. This was likely because P4’s work week schedule fit

better with his internal timing preferences while his weekend sleep was forced

to shift due to social engagements with later-type peers. Indeed, sleep onset for

P4 was 93 minutes later on weekends than during the week, plus sleep dura-

tion was reduced (by an average of 54 minutes) since the natural circadian drive

would still prompt an early wake up even after a later-than-preferred sleep on-

set [426] following a night of socializing.

Next, to quantify social jet lag and assess its severity across participants, I

computed the difference between mid-sleep (the halfway point between sleep

onset and waking) on free days (MS F) and on work days (MS W) per a formula

from the chronobiology literature [528]:

∆MS = |MS F − MS W |

Figure 4.5 shows the results of this calculation using the social-sensor data,

presented according to participant chronotype. Alarmingly, it is estimated that

Chronotype (MSF)

Soci

al Je

t Lag

(min

utes

)

8070605040302010

3:00-4:00am 4:00-5:00 5:00-6:00 6:00-7:00

Figure 4.5: Socially-sensed average social jet lag (discrepancy betweenmid-sleep on free days and work days) across chronotypes.

118

over 70% of the population suffers from social jet lag [424], and I unfortunately

observed it impacting Experiment 1’s participants too. My results compare well

to those from prior analyses on the MCTQ database, which similarly find social

jet lag ranging from approximately 1–2 hours [425]. In addition, I found that the

extreme ends of the chronotype spectrum experienced more social jet lag; and

it was most severe for the sample’s later types, as expected since their socially-

constrained days (work days) outnumber their free days (weekends) [528].

I also compared social jet lag across the three phases of Experiment 1 (Fall

semester, Winter break, and Spring semester) since academic responsibilities,

employment schedules, and social expectations vary across these periods. Fig-

ure 4.6 illustrates results. During the Fall and Spring semesters, sleep midpoint

was much earlier on weekdays versus weekends since imposed class sched-

ules forced earlier wake up times during the week. Further, it appeared more

sleep debt accumulated during work days in the Fall compared to the Spring

semester, as reflected by a considerable shift in weekend sleep midpoint during

Fall weekends in order to compensate. I believe this was due to the fact that the

Fall study phase overlapped with the demanding end-of-semester exam period

Work daysWeekends

Fall Winter Spring

Slee

p M

idpo

int

6:30

6:00

5:30

5:00

4:30

4:00

am

Figure 4.6: Shifts in sleep midpoint across study phases.

119

whereas the Spring study phase was during the (slightly) less intensive start of

the semester. A number of Facebook posts from the dataset suggested this to

be the case as well, for example: “lab exam. how much should i stay up to study

tonight??” (Fall) compared to “still just shopping for classes” (Spring).

On the other hand, when these external academic pressures somewhat sub-

sided during the Winter break and participants could more freely choose their

sleep timings, I found far less fluctuation between sleep midpoint on weekdays

versus weekends, which differed then by less than 10 minutes. Still, it is possible

for individual differences to exist in terms of social dynamics during vacation

periods. From interviews, I learned that some participants’ main social groups

were located in the place to where they were travelling, resulting in a substan-

tially reduced need to use CMC technology to stay in touch, as compared to

while away at school and disconnected from those groups. On the other hand,

other participants explained that their online networks were mainly comprised

of schoolmates, which meant that leaving during the break instead resulted in

increased CMC usage to maintain contact. Such results serve as an important re-

minder to avoid generic assumptions regarding technology usage and instead

to consider the variety of individual circumstances that can impact or reflect

widely different usage habits.

Nightly Sleep Disruption and Morning Sleep Inertia

A number of potential factors can contribute to sleep disruption. As men-

tioned, caffeine, exercise, napping, and alcohol are commonly studied by HCI

researchers, sometimes with an eye to designing technology to help users main-

tain sleep hygiene. Regardless of its culprit, the detrimental physical, cognitive,

120

and psychological effects of poor sleep are numerous [81, 377]; and such defi-

ciencies that follow a night of inadequate sleep can be initially observed during

the wake up process. Specifically, the term “sleep inertia” is used to describe the

time a person takes to become fully awake and functional, and prolonged sleep

inertia is a symptom of social jet lag [428]. Given that the duration of morning

technology usage has been shown to be a reasonable proxy of sleep inertia [3],

I investigated what specific technology-mediated activities typically comprised

morning usage for participants, along with the feasibility of using social-sensor

data to model this sleep-wake transition.

Analyzing rise time usage, I found that all participants used their smart-

phones within 10 minutes after waking up for activities such as browsing the

internet, checking email, and interacting with social media or communication

apps. Note that this usage was separate from alarm-related usage (7 of 9 partic-

ipants in Experiment 1 reported using their smartphones as their daily alarms).

The amount of morning phone use I observed is consistent with prior large-

scale studies on college students’ mobile device habits [279]. Prior work has also

found that communication applications are typically among the first apps used

upon waking from sleep [54], a tendency I found in Experiment 1’s phone probe

and social media data as well: on average, I detected some form of technology-

mediated social interaction within an hour of waking, with text messaging be-

ing the predominant form of social technology use (compared with phone calls

and social media) on more than two-thirds of mornings.

I attempted to operationalize sleep inertia according to the duration of morn-

ing CMC activity but did not observe the same strong association found in prior

work that bases usage on screen on/off events [3]. This suggests CMC-based ac-

121

tivities are a viable option for assessing wake events since they are frequently a

user’s first form of usage upon waking — but that attention soon turns to other

sorts of usage that may be more informative for measuring sleep inertia. For in-

stance, interviews revealed such usage often involved browsing news, weather,

and videos for this study’s participants. Going forward, it would therefore be

worthwhile to capture data about these types of interactions if the goal were to

build models for predicting morning inertia and transitional states out of sleep.

4.2.3 Sleep’s Links with Neurobehavioral Functioning

As described previously, social jet lag has numerous detrimental consequences,

with symptoms manifesting as cognitive difficulties and emotional problems.

Moving beyond morning rhythms, I therefore next used CMC data to explore

the impacts of sleep on such neurobehavioral functioning the following day,

specifically focusing on attention, cognitive functioning, and mood. These char-

acteristics are known to exhibit strong circadian patterns, suffer substantially

after sleep loss and interruption, and are considered especially important at-

tributes to evaluate for individuals in the participants’ age group [377].

I defined a number of socially-sensed variables in order to operationalize

activity levels, social interactions, cognition, and emotions, all of which prior

research and this own study suggested as relevant to performing such circadian

assessments. Here I present my analyses that revealed meaningful differences in

these variables on days following nights of varying sleep quality. Comparisons

were performed on medians using Wilcoxon sign-rank tests. Following estab-

lished guidelines, I treated sleep durations lasting 7–9 hours as “adequate” and

122

durations outside this range as “inadequate” [88] — though it is important to

note that just as our internal biological clocks direct our preferred sleep timings,

there are individual differences in sleep need as well [428].

Attention and Cyberloafing

A strong theme that emerged from the interview data was that participants

often turned to CMC when tired, bored, or unable to pay attention to tasks at

hand and that CMC was frequently used as a way for participants to procras-

tinate, entertain themselves, or simply pass the time. Cyberloafing is a term

used to refer to such procrastination and idling behaviors [289]. This tendency

to postpone tasks may be explained by a lack of attention and an inability to

focus that stem from insufficient self-regulatory resources, which drain over the

course of a day and require adequate sleep to become restored [37]. Both sleep

quantity and quality are important to this restoration [218], and an individual’s

failure to obtain both can result in increased levels of cyberloafing [351].

To represent cyberloafing behaviors, I therefore computed the following

measures based on a participant’s technology-mediated social interactions,

which I refer to as “CMC usage events”:

• Volume: The total number of CMC usage events a participant performs in

a given day between initially waking and eventually going to sleep.

• Burstiness: The maximum number of CMC usage events a participant

performs in any single hour between wake and sleep.

• Frequency: The number of hours between a participant’s successive CMC

usage events.

123

Adequate InadequateVolume ** 18 34Burstiness *** 6.12 9.54Frequency *** 0.71 0.43

Table 4.3: Median values of CMC-based activity levels following nights ofAdequate vs. Inadequate sleep. Significant differences in medi-ans marked on variable name (**p < .001, ***p < .0001).

As presented in Table 4.3, my analysis of these variables found that inad-

equate levels of sleep were associated with heavier use of technology the fol-

lowing day. Specifically, nights of insufficient sleep were associated with more

CMC-based usage events the next day, which were made more frequently and in

tighter temporal bursts. Correlating hours of sleep with the amount of next-day

cyberloafing activity showed the same negative relationship (r = −.52, p < .01).

During interviews, participants all mentioned checking social media when hav-

ing trouble focusing or concentrating, which they expressed often happens

when tired (e.g., “If I’m more tired, I’m less able to pay attention in class and more

likely to use phone to avoid falling asleep or get bored more easily.”)

Prior research has indicated that individuals with higher levels of consci-

entiousness may naturally possess more self-regulatory resources [111] and be

less susceptible to cyberloafing following lost or disrupted sleep. I therefore

performed linear regression between sleep duration and the amount of subse-

quent CMC activity while controlling for personality. I found sleep duration

(β = −.39, p < .001) and conscientiousness (β = −.16, p < .01) to be significant

predictors of subsequent CMC usage; and the negative direction of the partial

slopes again indicated that the less sleep an individual got, the more she used

CMC technologies the following day.

124

Cognitive Functioning

Sleep deprivation can also impair cognitive performance, including in aca-

demic settings, which is particularly salient for the students that comprised this

experiment’s sample. While people’s performance levels naturally fluctuate

throughout the day, sleep loss dampens overall performance; and the impair-

ment effects of fatigue coupled with these endogenous changes in daily brain

function have even been equated to alcohol intoxication [269]. Conversely, ade-

quate sleep duration improves learning and problem solving [507].

As a proxy for cognitive functioning, I assessed text-based content from par-

ticipants’ Facebook posts. I first performed standard pre-processing on these

posts (e.g., removing punctuation and URLs, handling spelling errors) and then

calculated the following measures, which represent the sophistication of a par-

ticipant’s posts and the cognitive effort required by the writing:

• LIX: A measure that indicates the linguistic sophistication of a piece of

text, computed as the percentage of words having 7 or more letters plus

the average number of words per post [47].

• TReDIX: A LIX-based measure adapted for use with social media content,

computed as a ratio of the total count of words having 7 or more letters

that appear in all posts made within a time period over the total number

of posts made in that time period [220].

As summarized in Table 4.4, I found that an adequate number of hours of

sleep related to higher levels of complex thought according to both cognitive

functioning measures. The greatest difference was in the TReDIX measure, and

linear regression confirmed a positive relationship; that is, the fewer hours of

125

Adequate InadequateLIX * 0.3592 0.3003TReDIX ** 0.2738 0.2144

Table 4.4: Median values of cognitive functioning variables followingnights of Adequate vs. Inadequate sleep. Significant differencesin medians marked on variable name (*p < .05, **p < .01).

sleep, the lower the subsequently demonstrated cognitive ability according to

social-sensor based assessment (β = 2.17, r2 = 0.12, p < .001).

Mood

Consequences of sleep reduction include negative mood, tension, nervous-

ness, and irritability [377]. Conversely, extending sleep improves mood [232].

To evaluate whether social-sensor data could be used to reflect patterns in

mood, I again turned to Facebook post data and this time applied psycholin-

guistic analysis techniques to compute the following measure, which prior work

has found to be a reliable representation of sentiment based on its strong corre-

lation with sentiment ratings from human judges [220]:

• Sentiment Intensity Rate: A measure of how intensely positive or negative

emotions are, computed as the ratio of the sum of valence intensity of

positive or negative language in posts to the total number of posts in a

period. Valence intensity can be determined from the ANEW dictionary

[63] and positivity and negativity from the LIWC dictionary [390].

To avoid skewed results due to participants with many more Facebook posts

than others, I normalized values of sentiment variables to between 0 and 1

(scaling in this way also makes results more interpretable — values closer to 1

126

Adequate InadequatePositive Sentiment Intensity *** 0.5373 0.3057Negative Sentiment Intensity *** 0.4176 0.8388

Table 4.5: Median values of sentiment expressed in Facebook posts fol-lowing nights of Adequate vs. Inadequate sleep. Significant dif-ferences in medians marked on variable name (***p < .0001).

indicate levels of the sentiment variable are nearer to the maximum value ever

observed for that individual and values closer to 0 indicate levels nearer the

minimum). Table 4.5 shows the differences in positive and negative sentiment

expressed after adequate and inadequate sleep.

I found that positive sentiment following nights with adequate sleep was

1.75 times higher than following nights with inadequate sleep, after which neg-

ative sentiment was instead over twice as high. Figure 4.7 illustrates the differ-

ence in negative sentiment on days following nights of varying sleep duration.

A similar relationship between insufficient sleep and negative affect has been

observed in prior studies that required participants to take daily EMA-based

mood assessments [347]. Interviews agreed, with Experiment 1’s participants

consistently noting their usage was higher when energy and mood were lower

PositiveNegative

Slee

p D

urat

ion

(hou

rs)

-1 0 0.5 1Sentiment

-0.5

Over 9

Under 5678

Figure 4.7: Sleep duration and sentiment the following day.

127

(e.g., feeling “more down” or “down and frustrated”) and also describing using

social media to “vent” or seek social support when tired and irritated.

Subsequent daily mood also showed an association with the timing of sleep

onset, with 3am showing the strongest signal during exploratory analysis.

Specifically, I found that participants whose final CMC activity happened af-

ter 3am had the lowest levels of measured sentiment the following day, while

posts from individuals who went to sleep at a relatively earlier time were 2.2

times more positive the following day. (Note that “earlier” here means rela-

tive to other participants’ timing of sleep onset and does not necessarily mean

“early” as in “early to bed and early to rise”, given these participants have later

chronotypes compared to a general population). More obtrusive studies admin-

istering end-of-day mood surveys and employing a wide array of sensors (e.g.,

computer logging, heart-rate monitors) have similarly found that people who

go to bed earlier are also happier [310], and my observation also aligns with

prior work associating late-night social media usage with depression and stress

[121, 355] — though the cause versus effect remains unknown. My results thus

complement prior findings about a connection between sleep and mood as well

as demonstrate how social media data can reveal this relationship.

128

4.3 Experiment 2 Findings

4.3.1 Smartphone Application Usage Patterns

Similar to Experiment 1, I began my analysis by exploring the types of smart-

phone apps participants in Experiment 2 used, along with temporal trends in

those usage patterns both within and over days. When possible, I again com-

pared my findings to those from prior work, both to affirm the sample was rep-

resentative of larger populations and to highlight and interpret new findings.

Daily Rhythms in Smartphone Application Usage

Aggregating participants’ smartphone app use events, I found the trends

illustrated in Figure 4.8, with app use overall at its lowest in early morning,

steadily rising and remaining relatively high from approximately noon until

late evening, and then dropping off. These trends are similar to those observed

in prior studies on daily mobile, computer, and internet usage [54, 310].

The most heavily used types of applications across all hours of the day were

communication and social media apps. Communication app use was highest

between late morning and midnight, with peaks mid-afternoon and evening,

similar to trends Experiment 1 found for phoning and texting. Usage of social

media apps (which, compared to those in the communication category, are used

more for consuming social content rather than communicating) had maximum

usage levels between 7pm and midnight, similar to findings observed in prior

studies [54, 310] as well as Experiment 1, that social media is most heavily used

in the late evening.

129

230 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

2500

0

500

1000

1500

2000

App

licat

ion

Usa

ge E

vent

s

Entertainment Time & WeatherCommunicationProductivity BrowsingEmailSocial Media

Figure 4.8: Hourly smartphone app usage by category.

Browser use was relatively stable from morning until late night, except for

dips around 3pm and 10pm. I observed the same for email use, which declined

gradually from late morning onward. The use of time & weather apps spiked

around 8–9am, which makes sense since participants’ sleep diaries indicated

nearly 60% of wake times were within an hour of 8:30am and over 75% of par-

ticipants used the phone as an alarm (often a gateway to usage, they reported),

similar to past findings that over 80% of people use a workday alarm [424].

Finally, entertainment apps were used more during the same morning pe-

riod as well as in mid-afternoon and late night, while productivity apps showed

usage peaks at points later in the morning, afternoon, and evening with a dip

mid-day and dropping off past late evening. These patterns are similar to those

found in prior work [54, 309], though shifted an hour or so later, likely because

this experiment’s sample was younger and therefore trended later in terms of

timing [428].

130

Weekly Usage Trends

I also found a distinction in the use of entertainment and productivity apps

across days of the week. Figure 4.9 presents the percentage of usage for each

category on each day, showing a reversed “scissor” pattern also found in other

work [422].

For participants in Experiment 2, work days were Monday–Friday and free

days corresponded to the weekend, Saturday–Sunday. At the beginning of the

work week on Monday and Tuesday, I saw over 40% of productivity-based us-

age events occurring, while Friday and weekend days showed the least use of

productivity apps — except for Wednesday, when only 8% of usage events were

productivity-related. Inversely, Wednesday was the day when entertainment

apps were used the most, followed by Friday and weekends. This mid-week

dip resembles a common mid-week sentiment dip found in other work [4]; and

in interviews, participants expressed experiencing a high degree of fatigue on

20

0

5

10

15

% o

f Usa

ge E

vent

s

EntertainmentProductivity

Mon Tues Wed Thu Fri Sat Sun

Figure 4.9: Use across the week of productivity and entertainment appsshown with standard error.

131

Wednesdays related to their class schedules — though further study is required

to see if this mid-week effect is consistent inside and outside of other college

populations.

4.3.2 Circadian Alertness Rhythms Reflected in App Use

Overall, these patterns replicated and expanded past findings and provided de-

scriptive insight into types and temporal patterns of mobile application use. Yet,

while this helps increase understanding of what individuals are doing with their

phones over the course of hours and days, more work was necessary to under-

stand why. I therefore next looked to chronobiology to add explanatory bite.

Usage Relative to Internal Time and Alertness

I first explored how usage patterns varied for different chronotypes. As men-

tioned, chronotype modulates nearly all biological functions [504], including

alertness performance [109]. Simply put, earlier chronotypes are more alert ear-

lier in the day, and later chronotypes function at their peak alertness later [212].

My comparison between the amount of usage events between early and late

types across parts of the day suggested this distinction might be reflected in

differing usage patterns, particularly of productivity and entertainment apps.

Figure 4.10 shows these statistically significant differences in app use (p < .05

using Wilcoxon sign-rank tests) between early and late types, broken down by

application category and time of day. Bars above (or below) the y axis indicate

early types used that type of app at that time of day the indicated percentage

more (or less) than late types.

132

100

-100

-80-60-40-20

020

406080

Earl

y-La

te U

sage

Cha

nge

(%)

Morning(6AM-12PM)

Afternoon(12PM-6PM)

Evening(6PM-12AM)

Night(12AM-6AM)

EntertainmentProductivity

Figure 4.10: Percentage increase (positive y value) or decrease (negative yvalue) in amount of usage by early types compared to usageby late types of productivity and entertainment apps acrossthe day.

That is, in the earlier half of the day, early types used approximately 25%

more productivity apps than late types and 19–29% fewer entertainment apps;

while I observed the opposite effect for evening and night usage, when early

types used 15–50% fewer productivity apps and 22–68% more entertainment

apps than late types.

As mentioned, alertness exhibits well-known fluctuations over the course of

a day [14]. The pattern of peaks and dips in alertness is roughly the same for

everyone, but there are individual differences in the phase of these rhythms that

are reflected by one’s chronotype. To align the phases of circadian rhythms for

different chronotypes, temporal analyses of usage patterns can be shifted to a

measure of time that is adjusted to take chronotype into account.

“External time” (ExT, also known as “clock time” or “local time”) is the num-

ber of hours that have elapsed since midnight (the midpoint of nighttime) [117].

“Internal time” (InT, also known as “body clock time” or “biological time”) re-

133

flects the phase difference (i.e., the “phase of entrainment”) between time cues

from the environment (e.g., the cycle of the sun) and the timing of an individ-

ual’s biological clock and is therefore computed as the number of elapsed hours

since an individual’s sleep midpoint, MS FS C (the midpoint of a person’s biolog-

ical night) [426, 504]. That is, internal time is a corrected measure of time that

reflects individual chronotype, calculated as:

InT = ExT − MS FS C

Reconsidering mobile app use in terms of internal time rather than exter-

nal time, I now saw associations with participants’ innate biological rhythms of

alertness (based on PVT performance — described in the “Daily Self-Reports”

portion of Section 4.1.2, under “Alertness EMA”). Productivity and entertain-

ment apps specifically showed the strongest associations with performance of

all app types in Experiment 2’s dataset. Specifically, I found a strong positive

correlation between performance and productivity app usage (r = 0.52, p < .001)

— that is, higher alertness performance was associated with more usage of pro-

ductivity apps. I also found an inverse relationship between performance and

entertainment apps (r = −0.31, p < .05), indicating that lower alertness was

related to increased entertainment app use, though the correlation was more

moderate. Important to note is that I did not find any such strong nor statisti-

cally significant associations between alertness and app use when I did the same

analysis using external time.

Figure 4.11 illustrates alertness together with usage of productivity and en-

tertainment apps over the course of the biological day. Inspecting these trends

beginning with the midpoint of biological night (hour 0 of internal time) showed

how alertness levels gradually rose from the end of sleep through the wakeup

134

230 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9Perform

ance6

-8

-7

-6

-5

-4

-3

-2

-10

1

2

3

4

5

Usa

ge

EntertainmentProductivityAlertness

Internal Time (InT)

Biological… Morning Mid-Day Night

Figure 4.11: Temporal trends in application use (Usage) and alertness per-formance (Performance) across internal body clock time (InT).Usage axis is proportion (normalized to [0,1] scale) of all anapp category’s usage events that occurred in a given hour.Performance axis is percent deviation from individual base-line of alertness measured in a given hour. Internal time axisis number of hours since biological midnight, and accompa-nying spectrum indicates periods of the biological day.

phase. During this same period, usage of entertainment apps was over 2.4 times

higher compared to productivity apps. These findings resonate with the concept

of sleep inertia, which can last for hours, reflects the transition period from sleep

to full wakefulness, and is characterized by diminished alertness and vigilance

in attention [428]. Nearly three-quarters of interviews supported this idea of an

association between groggy wakeups and morning entertainment app use (e.g.,

“I’ll stay on the phone longer, browsing YouTube, etc, if I’m more tired.”)

Following this wakeup period, I found that alertness performance eventu-

ally peaked approximately 7 hours after sleep midpoint, which agrees with

trends found in prior research [79, 183, 504]. At the same time, the use of pro-

ductivity apps also ramped up and reached its own daily maximum, while the

135

use of entertainment apps fell to one of its minimum levels. Both the well-

studied mid-day alertness dip (during which performance is known to drop

[79, 341, 342]) and evening rebound [138, 287] were also observable in partici-

pants’ alertness patterns and aligned with a productivity app use dip and peak,

respectively.

Finally, as biological night approaches, alertness is known to fade [504], and

participants’ data showed this same trend in diminished alertness. (The outlier

spikes at InT=1:00 and InT=23:00 resulted from sparse data since this period

overlaps with sleep). In parallel, productivity app use also fell off while en-

tertainment app use stayed more elevated. Accordingly, in interviews, partici-

pants commonly mentioned nightly habits related to watching videos or play-

ing games (e.g., “Every time before I go to bed, I play a card game until I feel sleepy.”)

Gauging Alertness Level from App Use Features

I next explored how alertness may be reflected through additional usage fea-

tures beyond the time of day an app is used. In Experiment 1, using technology

for a longer amount of time was associated with procrastination, inattention,

and lack of devoted concentration. In addition, switching among different tasks

and computer windows has shown relations with capacity for sustained atten-

tion, distractibility, and boredom [309, 310]. Such findings suggest that the fol-

lowing metrics of app use duration, diversity, and switching may therefore be

particularly relevant to alertness:

• Duration: Mean # of seconds per usage session during T

• Diversity: Total # of distinct apps used during T

• Switching: Total # of app switches during T

136

I calculated these features based on usage in a given hour window (T) sur-

rounding an alertness measurement. To clarify, a usage “session” represents

a period of interaction marked by unlocking the phone and is comprised of

any number of app foreground events. I used Mann-Whitney-Wilcoxon tests to

compare these features during low and high alertness states. Guided by prior

research, I set thresholds for low and high alertness according to whether a PVT

measurement was below or above that participant’s individual baseline — i.e.,

was a negative value in the range [-1, 0) or a positive value in the range (0, 1],

respectively [504].

Participants’ overall durations of usage showed good agreement with dura-

tions found in prior work [54, 158, 541]. During periods of low alertness, I found

duration of use was over 20% higher as seen in Table 4.6. Interviews agreed that

usage became more “bottomless”, “stuck”, and “harder to get off”. I also found

that participants also switched apps 33% more when alertness was low, though

they did not necessarily switch among a larger set of distinct apps, as app di-

versity showed no significant difference between alertness states.

Low Alertness High AlertnessDuration* 103.4 seconds 85.8 secondsDiversity 2.87 apps 2.82 appsSwitching* 32 switches 24 switches

Table 4.6: Median values of usage features during low vs. high alertness.Significant differences in medians marked on variable name(∗p < .05).

137

4.3.3 Connecting App Use and Alertness with Sleep

Lastly, I studied connections among app use, alertness, and sleep. Variation in

performance is made most evident by sleep loss [184], with the largest effects

on alertness, working memory, and cognitive throughput [288]. In addition to

impairing performance directly, inadequate sleep is also associated with subse-

quent feelings of fatigue [269]. Conversely, extending sleep enhances learning

and problem solving [509], with adequate sleep duration improving alertness,

energy, and reaction time [232, 428].

Comparing sleep duration according to participants’ sleep diaries with their

app use the following day, I found that less sleep was correlated with less

productivity-oriented usage (r = 0.43, p < .05) and more use of entertainment

apps (r = −0.19, p < .05), for both weekdays and weekends.

A more coarse grained measure of sleep duration adequacy also showed

statistically significant differences. Specifically, considering (as in Experiment

1) sleep lasting 7–9 hours as “adequate” following established guidelines and

prior research [88], I found participants used productivity apps an average of

61% more after nights of adequate as opposed to inadequate sleep (Cohen’s

d = 0.48, p < .05), while entertainment apps were used 33% more on average

after an inadequate amount of sleep (Cohen’s d = 0.24, p < .05).

Interviews provided qualitative detail about sleep loss and subsequent fa-

tigue manifesting through increased usage of entertainment-based apps, which

participants described as enabling “mindless”, “passive” interactions; while on

the other hand, they associated feeling rested, energized, and alert with more

“intentional”, “directed”, and “productive” usage.

138

Finally, recent studies continue to suggest links between nightly technology

usage and sleep problems — particularly when it comes to usage delaying sleep

onset and cutting into sleep time [68, 415, 485]. I found a mild relationship

between the number of experienced sleep interruptions according to daily sleep

diaries and the number of app usage events sensed between sleep onset and

waking (rs = 0.46, p < 0.05). This indicates app use data might help to assess

sleep disruptions — though, the phone itself might be a culprit of disruption

in the first place. In interviews, the majority of participants described turning

their phone to silent overnight to avoid such sleep interruptions. However, even

phantom notifications sometimes awoke them (e.g., “Sometimes I wake up as if I’m

expecting something, like an email or a text, and will check my phone. I imagine that if

I didn’t have technology, I’d have a sound sleep.”)

These results demonstrate how app usage can provide informative signals

when assessing sleep duration, interruption, and associated feelings of alert-

ness and fatigue. However, they also reveal a disruptive potential of mobile

devices, which deserves careful consideration from researchers who leverage

digital traces from these systems. I examine this tension further in the disserta-

tion’s concluding chapter (see Section 6.2.1).

139

4.4 Discussion

Having presented a variety of results from both experiments, it would be worth-

while to now reflect on this research. In this section, I first provide a brief sum-

mary of these studies’ findings and their commonalities, then identify their lim-

itations together with opportunities for future work, and conclude with broader

implications for the domain-driven framework.

In Experiment 1, I focused on examining sleep, including sleep-related cir-

cadian disruptions and how inadequate sleep relates to changes in cognition

and mood the following day. To begin, I analyzed temporal trends in the use

of computer-mediated communication technology, along with associations be-

tween usage and sleep timing. Leveraging these usage patterns, I then inferred

sleep onset and wake events with an accuracy comparable to approaches that

are more obtrusive for users and less feasible to deploy on a mass scale. Ap-

plying this sleep sensing technique, I was next able to detect chronobiology

phenomena such as the “scissors of sleep” and “social jet lag”. I further used

social-sensor data to explore the impacts of sleep on variables defined to repre-

sent neurobehavioral functions known to exhibit strong circadian patterns and

suffer substantially after sleep loss and interruption: attention, cognition, and

mood. My analyses revealed significant differences in these variables following

nights of adequate vs. inadequate sleep — specifically, that lack of quality sleep

manifested in cyberloafing behaviors (based on an increased amount, frequency,

and burstiness of technology usage the following day), diminished cognitive

functioning (based on the expression of complex thought in text-based social

media content), and more negative mood (based on sentiment analysis of that

same content). Overall, my findings suggest that social-sensor data can serve as

140

a proxy to measure sleep timing, the extent of circadian disruptions like social

jet lag, and the impacts of inadequate sleep on neurobehavioral functioning.

In Experiment 2, I explored daily functioning further. To scope, I focused on

cognitive performance and even more specifically on alertness given how many

other aspects of performance it underpins. Analyzing smartphone app use logs,

I found that early and late chronotypes displayed inverse patterns of app use;

that usage features such as duration and switching could distinguish periods

of low and high alertness; and that app use reflected aspects of sleep includ-

ing duration, interruptions, and subsequent fatigue. Notably, I demonstrated

that adjusting analyses to biological “internal time”, as defined by the chrono-

biology literature, surfaced otherwise undetected correlations between alertness

and app use, with individuals using productivity apps during their optimal per-

formance periods and turning to entertainment apps when less alert.

Before considering the limitations of these experiments and the attendant

directions for future work, it is worth relating their results to each other a bit

more. In fact, the two experiments’ findings did align in interesting ways.

In both experiments, I found that less sleep correlated with particular usage

patterns: elevated CMC usage in Experiment 1 and elevated entertainment app

usage in Experiment 2, suggesting that CMC and entertainment software might

have similar use cases and meet similar needs under some circumstances. Fur-

ther, inadequate sleep was not only associated with a greater overall volume of

CMC-based usage events the following day in Experiment 1, but those CMC

usage events occurred at more frequent intervals too. Additional data captured

during Experiment 2 further allowed me to find associations between inade-

quate sleep and lower alertness, which in turn related to longer usage sessions

141

that involved more frequent, defocused switching among apps. An interpreta-

tion supported by these data from usage logs as well as qualitative data from in-

terviews, is that inadequate sleep is linked to subsequently diminished alertness

and that CMC and entertainment apps are utilized as a means of cyberloafing

to manage this experienced fatigue.

Because data about sleep, alertness, and CMC and entertainment apps were

all collected in Experiment 2, it is possible to look at potential connections di-

rectly. Many of my reported findings for Experiment 2 focused on productivity

and entertainment apps because they exhibited the strongest correlations with

alertness, but social media and communication apps (i.e., CMC apps) together

accounted for over half of that study’s usage events and sent some informative

signals too.

To begin, I found usage of these CMC apps was elevated during biological

morning, and nearly all participants in Experiment 2 described using social me-

dia apps as a way to “ease” themselves into the day (e.g., “To wake myself up,

I’ll have to look at things on the phone like Facebook or Tumblr.”) Also notable is

that these participants’ morning classes often fell within their phase of sleep in-

ertia; and the majority of participants described using their phones in lectures

when tired, bored, or unable to concentrate, for instance to help keep them-

selves awake — which they explained was particularly necessary for morning

classes (e.g., “In morning classes, I have less attention and am very tired so I’ll browse

the phone. Using tactics like social media, I focus on the screen to try to keep my eyes

open.”) As presented earlier, participants in Experiment 1 described a similar

“tactical” use of CMC to manage fatigue (e.g., “If I’m more tired, I’m less able to

pay attention in class and more likely to use phone to avoid falling asleep or get bored

142

more easily.”) Given the negative impacts on learning associated with this type

of distracting technology usage in the classroom [207], such findings suggest

that the learning impairments associated with lectures being scheduled at bi-

ologically unsuitable times may be further compounded by this compensating

phone usage.

Then throughout the day, participants in Experiment 2 appeared to continue

turning to CMC apps when experiencing low alertness (e.g., “I go for apps that

don’t require much mental energy when fatigued. Facebook, Yik Yak.”), including dur-

ing the mid-day dip when these apps were used more than any other category.

Reaching the day’s end, I found CMC apps also interplayed significantly with

behaviors before and during sleep, with over 50% of sensed sleep interruptions

corresponding to social media app use. When describing the amount of time

between phone use and sleep in interviews, half of Experiment 2’s participants

reported that usage was immediately before sleep and all but two reported it

was within thirty minutes; and they noted this usage was often related to CMC

(e.g., “I use my phone directly before bed — Messenger, email, Facebook. Any notifica-

tion.”) Such findings resonate with those from Experiment 1 regarding ways in

which CMC interacts with (including sometimes interrupts) sleep.

Altogether, these experiments provided evidence that biological rhythms ex-

ert a strong influence on patterns of sleep and alertness, and they demonstrated

how these relationships and other phenomena well-known in a domain of inter-

est may manifest through technology use. As a result, these soft sensor signals

provide opportunities for more unobtrusive, scalable, and personalized assess-

ment and intervention.

143

4.4.1 Limitations and Future Work

Broadening Samples

As explained, individual differences in circadian variables can vary dramat-

ically. For reasons provided in Section 3.4.1, I focused on college students as a

vulnerable population to study; but patterns are likely different for people of

other age groups or who have different roles and work responsibilities. It is also

possible that users of non-Android phones behave differently.

Thus extending this work to larger samples would be a natural future di-

rection in order to see how well these experiments’ findings generalize across

a wider sample of participants (e.g., of more diverse age groups, genders, and

chronotypes as well as individuals living with affective illnesses such as bipo-

lar disorder, who could benefit immensely from technologies designed to sup-

port circadian rhythm stabilization). Similarly, extending the study to a longer

time frame and to additional geographical regions would allow measurement

of circadian variations over the course of seasonal and yearly cycles and across

multiple time zones and latitudes.

Further, I did not explicitly control for characteristics like class schedules,

course load, or a number of other factors that might exert an influence on par-

ticipants’ observed behaviors. Future work could extend models to include ad-

ditional characteristics of participants and their contexts. Similarly, it would

be desirable to examine the effects of light — including light emitted from de-

vices, especially given my findings regarding the use of technology at night.

Light plays a central role in setting the biological clock and the timing of sleep

and is also known to impact alertness. While the MCTQ contains a question

144

about daylight exposure on average, measurement of daily sunlight as well as

artificial light could be incorporated into analyses by including a question in

participants’ self-reported sleep diaries or by using data from a phone’s light

sensor. Such consideration of light could be particularly valuable to consider

depending on a participant’s geographical location or the time of year, when

differences in light-dark patterns may impact measures. Finally, though more

burdensome for participants, comparison with actigraphy measures as well as

exploring the use of body or environment based sensors could also contribute

to a more holistic representation of rhythms.

Broadening Data and Analyses

Future work can also expand the types of data collected and the analyses

performed. Additional qualitative data from diaries, EMA prompts with open-

ended questions, or more interviews could help to further unpack and explain

my quantitative observations regarding relationships among technology use,

sleep, alertness, and other neurobehavioral functions. Such qualitative data

might also enable the identification of additional edge cases in order to incorpo-

rate more informative features and make sensing more robust.

Next, future iterations of Experiment 1’s sleep sensing algorithm can in-

corporate additional socially-computable signals, for instance from emails, so-

cial media platforms beyond Facebook, or other data that interviews suggested

would be informative. This may not only improve sleep inference accuracy from

relevant behavioral data; but it would also allow examination of if, how, and

why individuals exhibit different behaviors in different technology-mediated

social contexts and whether such variations relate to circadian factors.

145

Similarly, in Experiment 2, I focused on smartphone usage because of a num-

ber of advantages outlined in Section 3.4.2; however, relying solely on behav-

ioral signals from phones may ignore useful data from other sources. In partic-

ular, I likely missed some use of productivity tools designed for devices with

larger screens and better input methods (e.g., laptop and desktop computers),

which would be desirable to consider in future work. Further, the mobile-based

alertness EMA was a source of potential confounds as it can impact attention

and usage [486], suggesting the value of exploring more unobtrusive measure-

ment. Such alternative sensing strategies could also capture aspects of alertness

that may manifest through non-use of technology.

Likewise, it would be worthwhile to study phone usage behaviors beyond

instances of app use (i.e., an app taking the foreground). Many of my analyses

focused on this initiation of app use since it is a well accepted, easily compara-

ble, and reliably captured metric, as described earlier. Theoretically, this does

mean a 15 second interaction and a 15 minute interaction could potentially ap-

pear the same way, as single usage events; however, the former case would

likely look different in practice because other app use events would tend to

happen during the additional 14.75 minutes, especially considering the amount

of app switching I observed. Still, I recognize the value in future work to look

deeper into other metrics of usage such as duration, revisitations, or chains of

app use that may offer additional insights.

In addition, although my application categorization was broadly useful, it

was unable to account well for apps that can be used in ways that map to both

low and high alertness. For instance, I observed elevated use of entertainment

apps in both low and high alertness states. A potential explanation suggested

146

by interviews is that “lightweight” games that do not require much mental en-

ergy (e.g., “mindless puzzle games”) are used primarily when fatigued, bored,

distracted, and for procrastination; whereas games played when feeling en-

ergetic and alert tend to require more focus and attention (e.g., “active strat-

egy games”). Similarly, correlations between participants’ email use and alert-

ness, together with interview data, suggested that checking email may be more

productivity-oriented earlier in the day and more about “killing time” or social-

izing later on and especially before bed. Modeling the actual behaviors enacted

in apps, though challenging, might therefore give a clearer picture of the rela-

tionships between technology use, alertness, and biology.

Still, as the first studies looking at social media and smartphone usage as soft

measures for studying circadian rhythms of sleep and alertness, I have obtained

a variety of useful findings relevant to researchers interested in chronobiology,

mobile sensing, or personalized technology design.

4.4.2 Implications for Domain-Driven Health Assessment

My overall motivation in presenting these experiments was to demonstrate how

domain knowledge can guide a process of data collection and analysis — specif-

ically, by informing important health determinants to analyze, avoiding compu-

tational costs of modeling unneeded features, guiding analyses, aiding interpre-

tation of outcomes, and generating implications for effective interventions.

147

To begin, this domain-driven approach enabled me to capture data and ex-

tract metrics that are believed to have a greater impact on health than indicators

I might have measured if I had followed an approach disconnected from the

domain of chronobiology. For example, extant sleep technologies, like those

I reviewed earlier in Section 3.2.3, have fixated on improving sleep hygiene in

terms of metrics like sleep duration. The previous chapter’s inquiry into chrono-

biology revealed other indicators (e.g., the experienced degree of social jet lag)

that are equally (or arguably, more) important to assess when it comes to eval-

uating sleep and overall wellness. Therefore in Experiment 1, I explored how

technology-mediated social interactions and communication patterns could be

leveraged to provide a means of assessing socially-rooted disruptions to circa-

dian rhythms — disruptions that the chronobiology literature helped me rec-

ognize as key determinants of overall health but that I would have failed to

include in an analysis that was guided instead by layman intuition.

Domain knowledge also helped me avoid unnecessary computational costs.

After finding usage of social technology was coupled with sleep behaviors, I

next harnessed this social-sensor data as a means of assessing sleep events and

quality. My algorithm was able to infer sleep events to a level of accuracy

comparable to (sometimes better than) prior work’s domain-disconnected tech-

niques — plus, this leaner, more targeted assessment approach was less compu-

tationally expensive and privacy intrusive than the more data-driven strategies.

In addition, domain knowledge provided implications for alternative ways

to analyze data. For example, my analysis of temporal trends in smartphone us-

age data in Experiment 2 revealed patterns that correlated with personal alert-

ness rhythms, suggesting such data could be leveraged to passively sense per-

148

sonal performance. However, such relationships became apparent only when

analyzed through the lens of internal time, a measure of time that inquiry into

chronobiology made available to me. As with Experiment 1’s findings about

disruptions, I would have stopped short of discovering Experiment 2’s find-

ings about a link between alertness and smartphone use if I had followed an

approach disconnected from the domain of chronobiology. Further, the notion

that analysis of usage trends can be adjusted to take the internal time of the body

clock into consideration has broad implications for PHI work that assesses hu-

man behavior across time. For example, a number of studies have aggregated

social media data across external clock time to study daily rhythms in mood

[187]. The same analyses corrected to internal time (i.e., using a timescale based

on the biological timings of those social media users) would be intriguing; and

given the extent mood correlates with circadian rhythms [232], I suspect the

results would be particularly striking in this case.

Next, both experiments demonstrate how domain knowledge can aid the in-

terpretation of findings. While my results aligned with usage patterns found in

related HCI research, domain knowledge about the biology behind sleep and

wake behaviors provided explanatory power. For example, in Experiment 2, I

brought a biological perspective on daily performance patterns to the interpre-

tation of how and why individuals use their smartphones in particular ways.

A domain-driven approach enabled me to go beyond prior works’ descriptions

of diurnal variations in app use in order to offer biological factors behind those

variations. Higher level constructs such as cognitive engagement and boredom

have been used in prior HCI work studying digital activity [309]; however, I

learned from domain knowledge that such constructs are underpinned by lower

level processes like sustained vigilance in alertness [145]. I expect that through-

149

out the day, people fluctuate to an extent across their boundaries of alertness

depending on their tasks, interactions, and other contextual factors — but that

circadian rhythm factors present a consistent limit to one’s cognitive function-

ing and hence are most helpful in understanding individual performance (and

in turn, could be practically useful in the planning of cognitively demanding

activities).

Further, these experiments show how a domain-driven approach can facil-

itate direct contributions back to that domain. For example, my scalable soft

sensing techniques for sleep and alertness assessment could benefit the research

of chronobiologists who express a pressing need to capture in-situ data from

large populations spread across diverse locations and time zones [423]. This

ability to detect and infer behavioral traits of circadian biomarkers opens up

new opportunities to answer fundamental scientific questions about sleep and

circadian rhythms in real-world settings. Thus another strength of a domain-

driven assessment approach is that the outputs of analysis are relevant not only

to the PHI enterprise at hand, but they can have broader ramifications too by

contributing novel methods or knowledge back to the informing discipline.

Finally, beyond contributing computational assessment techniques and em-

pirical findings, both studies’ domain-driven approach also allowed me to de-

rive a variety of design implications for novel systems that may considerably

enhance monitoring and intervention. In the next chapter, “Domain-Aware In-

tervention Design”, I present in more detail the research I have undertaken to

pursue those ideas.

150

CHAPTER 5

DOMAIN-AWARE INTERVENTION DESIGN

In the previous chapter, I showed how domain knowledge can support the col-

lection and analysis of personal data in order to extract meaningful insights

about health. For that information to be useful to individuals, it must be com-

municated in an understandable, actionable way. Therefore, this chapter moves

along to the remaining component in the personal health informatics pipeline

and, correspondingly, the final piece in this dissertation’s domain-driven frame-

work: engaging in domain-aware design processes to deliver user-facing feed-

back. (See “Domain-Aware Intervention Design” portion of Figure 2.1).

This design stage begins in a phase of planning, where the goal is to gain

an initial sense of what the PHI system will do and how. To better understand

a system’s “requirements”, designers can gather information through several

typical methods. For example, behavioral or social science theories can pro-

vide high level design principles applicable to a broad class of systems (e.g.,

using Fitts’ law of human psychomotor behavior to inform the design of point-

ing and input devices or using social science concepts about social capital or

self-presentation when designing online communities) [441]. One might simi-

larly look to sources of “textbook” design knowledge about best practices and

broad rules of thumb (e.g., style guides or heuristics).

Also important at this stage is engaging with anticipated users, for instance

through questioning (e.g., via surveys, interviews, or focus groups) or observa-

tional fieldwork in order to ensure a system meets their envisioned needs and

respects their extant practices. Earlier, I described a recent shift in medicine to-

ward a more person-centric model of care aimed at meeting individual health

151

needs and preferences. Similarly, a core principle in human-centered design is

that technology should be user-centric, designed around users’ needs in order

to create a positive user experience that demonstrates a deep understanding of

their perspective.

A domain-aware approach to planning additionally integrates knowledge

acquired during other stages of the framework (e.g., information gathered dur-

ing domain inquiry or empirical insights generated during health assessment).

In Section 2.3.2, I discussed the merits of informing design work with such do-

main knowledge in the context of PHI development, for instance, to ensure a

system is both targeting the factors that are relevant to a health outcome of in-

terest as well as accommodating the idiosyncratic characteristics that can impact

how a person will respond to feedback.

Altogether, this information can then support a phase of ideation to gener-

ate implications for design, such as speculative guidelines for what feedback

to present and how. A phase of building then involves the realization of these

ideas, by creating low or high fidelity mockups and prototypes or even full-

fledged systems where user models are instantiated. Finally, a participatory

review process supports the evaluation of these built artifacts — information

that can loop back and continue to fuel this ongoing design process aimed at

regularly checking in with user needs and iteratively refining the PHI system.

Overall, these domain-grounded, user-centered steps support the design of

systems that supply effective interventions as well as positive user experiences.

In this chapter, I demonstrate such domain-driven design work in practice by

presenting my research on creating chronobiology-aware PHI technology.

152

5.1 Chronobiology-Aware PHI

As described in Section 3.3, the chronobiology literature establishes that our cir-

cadian rhythms influence nearly all aspects of physiological and neurobehav-

ioral functioning, including our sleep-wake patterns, cognitive performance,

and mood. In addition, individual differences exist in these functions (e.g.,

the timing of sleep-promoting hormone secretions), as reflected by a person’s

chronotype. Designing PHI technologies that are aware of such biologically-

rooted, idiosyncratic characteristics opens up numerous opportunities for mon-

itoring, stabilizing, and helping individuals work in better alignment with their

innate biological rhythms — ultimately (ideally) to improve everyday life on a

broad scale. The following subsections discuss particularly promising applica-

tion areas related to sleep, daily performance, and emotional wellness.

The chronobiology-aware designs I present are of various fidelities — rang-

ing from speculative design implications (e.g., unimplemented design ideas), to

mid-level mockups (e.g., wireframes, storyboards), to working prototypes. In

describing these designs, I indicate how they embody the dimensions outlined

in Section 2.2.3 regarding format, delivery medium, attentional demand, room

for interpretation, and level of personalization. Finally, while some designs (e.g.,

the calendar system) might not seem like conventional types of personal health

technology, it is worth noting that the designs I present do meet PHI criteria —

they use personal data to assess physical and neurobehavioral aspects of health

and deliver feedback that can support individuals in self-managing wellness.

Regardless of the fidelity, embodied dimensions, or conventionalness of the

designs, they illustrate domain-driven design practices, in that they reflect in-

153

tentional choices informed by available literature (e.g., chronobiology theory,

clinically-validated interventions, condition pathology, etc.) as well as gathered

evidence (e.g., the empirical findings presented in the previous chapter) and are

grounded in user feedback that helps to verify a priori design ideas.

The following sections demonstrate various phases of the domain-aware de-

sign process. My descriptions of systems for supporting sleep (Section 5.2) focus

on the planning and ideation phases, while my mid-fidelity mockups and pro-

totypes for biologically-friendly productivity technology (Section 5.3) demon-

strate more of the building phase. Finally, the participatory design and deploy-

ment of MoodRhythm (Section 5.4) showcase the building and review phases of

a more high-fidelity, full-fledged system and also illustrate how insights gath-

ered during this review can loop back into the iterative design cycle, in order to

refine the system further and inform future design opportunities.

5.2 Chronobiology-Informed Sleep Support

As described in Section 3.2, given the current prevalence of sleep problems and

comorbid conditions, PHI developers are keen to measure, assess, and improve

various aspects of individuals’ sleep habits. However, (1) generic sleep hygiene

systems that do not consider circadian factors are missing the full picture, and

(2) interventions that only target sleep disturbances may merely be treating the

symptoms of a misaligned biological clock rather than helping to address the

roots of circadian disruption. In developing chronobiology-aware sleep tech-

nology, these two issues stand out as initial directions to pursue.

154

5.2.1 Personalizing Sleep Hygiene Recommendations

For patients who complain of sleep problems, clinicians often recommend that

they should work to improve their “sleep hygiene.” Sleep hygiene is defined

by the National Sleep Foundation as the “practices that are necessary to have

normal, quality nighttime sleep and full daytime alertness”. Examples of these

recommended practices include getting 7–8 hours of sleep per night, avoiding

exercise within 3 hours of bedtime, and generally limiting caffeine especially

within 4–6 hours of sleep.

Most of today’s sleep technologies are built with some subset of these sleep

hygiene recommendations in mind. For example, BeWell computes a sleep qual-

ity score based on adherence to this ∼7 hour ideal sleep duration [275]. “Got

Sleep?” computes a similar score but factors in adherence to additional rec-

ommendations related to caffeine and alcohol intake, meals, exercise, napping,

sleep environment, and electronics use [438]. As described earlier, ShutEye uses

a smartphone’s wallpaper and lock screen to show a visualization of how likely

it is that various activities including caffeine, exercise, and napping will nega-

tively impact sleep at that point in the day [36].

These sleep hygiene recommendations are a good place to begin thinking

about the design space of sleep technology at a high level, though they tend

to take a relatively generic perspective. In this case, generic advice is likely

better than no advice at all; but it would be desirable to make these recom-

mendations more personalized, given the highly individualistic nature of sleep

requirements. Specifically, chronobiology research suggests that how we sleep

is determined by multiple factors and contingent, in large part, on biological

attributes such as a person’s genetic makeup, age, and gender.

155

For instance, genetics play a role in determining the effects caffeine has on a

person, including the time at which intake will affect sleep [542]. Using an ap-

proach similar to the experiments I presented in the previous chapter, a PHI sys-

tem could attempt to translate personal data streams into behavioral biomarkers

of caffeine-related traits. This information could then help a system move away

from more blanket recommendations (e.g., ShutEye’s “End caffeine consump-

tion 8–14 hours before bedtime” [36]), in order to tune this window to a user’s

predicted genetic response to caffeine intake and its personal impact on sleep.

As another example, every person has a distinct chronotype, as described

earlier, with many individuals (especially young adults) falling closer to the

late end of the spectrum. Therefore, common recommendations for early sleep

timings as well as maxims like “The early bird catches the worm” and “Early

to bed and early to rise makes a man healthy, wealthy, and wise” perpetuate

normative values that do not necessarily fit with everyone’s biology [426] — yet

many systems supply sleep hygiene advice based on these generic perspectives

(e.g., Jawbone UP’s “Early to Bed” goals encourage users to adhere to earlier

and earlier sleep times, as mentioned earlier). Instead, chronobiology-aware

systems could supply tailored timings for sleep onset and waking that better fit

with an individual’s unique chronotype signature.

Similarly, recommendations could be further tweaked based on gender. Cur-

rently, sleep hygiene recommendations are not gender specific, even leading

ShutEye’s designers to explicitly discount the need to consider gender differ-

ences: “Sleep hygiene recommendations are not gender specific, and thus we

did not attempt to recruit an even number of males and females for the study”

[36]. However, men and women are different when it comes to sleep; women’s

156

circadian clocks tend to run earlier (by about an hour) and shorter (by about

6 minutes) than men’s [141]. A PHI system providing advice about sleep onset

and duration could therefore incorporate gender features into its algorithm, rec-

ommending timings that fall a bit earlier and last a bit shorter for female users

than for their male counterparts.

Overall, such examples illustrate how sleep hygiene advice could be pre-

scribed in more biologically-personalized ways. A natural starting point for

chronobiology-aware sleep support is to design tools in the style of extant PHI

sleep hygiene applications — i.e., suggesting the timing and duration of sleep

or computing a sleep quality score. But instead of providing one-size-fits-all

guidelines, sleep-related recommendations would be tuned based on a combi-

nation of any available information about biological traits (e.g., to provide a

genetically-appropriate caffeine cut-off or to recommend sleep timing and du-

ration advice that factors in chronotype, age, and gender). Responses from post-

study interviews in Experiments 1 and 2 indicated that all but two participants

would find this information meaningful.

Moving beyond timing and duration of sleep, metrics of sleep quality could

also be made more chronobiology-relevant. This dissertation’s inquiry phase

revealed other indicators that chronobiologists agree are equally important to

assess when it comes to evaluating sleep and its impact on overall wellness,

including sleep inertia, social jet lag, and sleep debt. Furthermore, such indica-

tors are often symptoms of a misaligned biological clock, which chronobiology-

aware interventions could help to stabilize.

157

5.2.2 Stabilizing Circadian Disruptions

As described in Section 3.3.3, disruption of our circadian timing system (i.e., cir-

cadian disruption or circadian misalignment) is associated with a wide range of

negative health impacts. Aging as well as neurological and psychiatric disor-

ders can lead to a breakdown in normal sleep-wake cycles; and even for healthy

individuals, modern living’s biologically-unsuited work schedules, social con-

straints, and late-night technology use can all contribute to circadian disruption

[528]. Therefore, another fertile area PHI sleep support can target is “fixing

the broken clock” — i.e., stabilizing circadian disruptions through lifestyle in-

terventions. One immediate design strategy is to create applications that can

facilitate the broad-scale, real-world deployment of circadian interventions that

have shown promise in small lab-setups or in animal studies.

Adjusting Sleep Schedules

Consider interventions to reduce social jet lag — the discrepancy between

sleep patterns on workdays and free days, with the former typically charac-

terized by chronic undersleep and the latter by compensating oversleep. In a

recent study aimed at reducing this mismatch and promoting more stable sleep

routines, factory shift schedules were adjusted to eliminate highly disruptive

shift assignments. That is, early chronotypes were not scheduled for night shifts

and late chronotypes were not scheduled for morning shifts. After five months,

sleep duration and quality increased, as did wellbeing ratings; and participants’

social jet lag was reduced by one hour [503]. Technology could help deliver such

chronotype-adjusted (CTA) sleep schedules to a wide user base. Given the in-

forming research was conducted with factory shift workers, this intervention

158

may be particularly well-suited to individuals from that population; but others

could benefit too. For example, doctors might use their recommended CTAs

when appealing for a shift change or students could reference their CTAs when

making decisions about whether to choose morning or night classes.

For a person unable to do this much of a schedule overhaul, a chronobiology-

aware system might instead target the reduction of sleep debt. Sleep debt (also

known as sleep deficit) refers to the accumulation of undersleep that people typ-

ically experience on workdays, and its misaligning effects are similar to crossing

time zones [91]. Considered a central factor behind a number of adverse health

outcomes, sleep debt is associated with increased daytime sleepiness, fatigue,

mental exhaustion, confusion, mood disturbance, tension, and stress; and the

effects of chronic sleep debt (lasting at least ten days) are similar to experienc-

ing total sleep deprivation [133].

Recent chronobiology research suggests that helping individuals take on

schedules that prevent the buildup of chronic sleep debt may be a promising

strategy for reducing the detrimental effects of ongoing, unavoidable disruption

(e.g., due to fixed work or class schedules that are ill-suited to one’s chronotype).

Specifically, by adopting a rotating work schedule that gave at least 24 hours off

after each night shift, study participants (shift workers) were able to immedi-

ately recover from sleep debt incurred from that night shift [160]. For a late

chronotype college student who must take an early class every, say, Monday,

Wednesday, and Friday, a chronobiology-aware sleep coach could help plan a

schedule that ensures activities on Tuesdays, Thursdays, and Saturdays begin

late enough to prevent the accumulation of sleep debt. Such a system might also

provide a user with chronotype-tailored napping schedules, which the same

159

study found were effective at compensating for the ill-effects of shortened sleep

after a night shift (e.g., napping up to three hours before a night shift for early

chronotypes [160]). This finding about the benefits of naps also reveals how cur-

rent sleep hygiene recommendations, which sometimes recommend the avoid-

ance of any napping more than 8 hours past waking [36], could be made less

brittle by supporting naps during later portions of one’s day, as long as they are

timed in a biologically suitable way to stabilize rather than disrupt sleep.

Finally, another effective way to support healthy circadian rhythms is to

maintain regularity in certain sleep-related “anchors”. In particular, keeping

mid-sleep (the halfway point between sleep onset and waking) consistent each

night can help stabilize circadian misalignment [528]. Even if a person is unable

to go to sleep at the same time each night or wake at the same time each morning

(e.g., due to work, school, family, or travel constraints), a chronobiology-aware

system could recommend sleep and wake times that minimize circadian disrup-

tion by keeping mid-sleep anchored as much as possible. For example, again

consider the case of the late type college student with morning classes three

days a week (M, W, F). Say that this individual normally sleeps at 12am, that

early morning classes require her to rise at 6am, and that she sleeps until 8am

on non-class days. A bit of arithmetic shows that her mid-sleep point is 3am on

Sunday, Tuesday, and Thursday nights and 4am on others. To eliminate these

mid-sleep fluctuations, a chronobiology-aware sleep advice tool could suggest

this user shift, for instance, her bedtime to 11pm on the nights when she does

not have class the next day and her wake time to 7am on her non-class morn-

ings. A quick recalculation confirms mid-sleep stability at 3am across the week.

While this example is simplistic for illustrative purposes, an adaptive system

could handle more realistic sleep patterns with more variable daily sleep and

160

wake times by determining on-the-fly an optimal sleep schedule to maximize

mid-sleep stability based on the current time, the user’s historical sleep-wake

data, and her known future schedule constraints.

Delivering Zeitgebers

Next, tools can supply interventions that help people temporally structure

activities beyond sleep in ways that reduce circadian misalignments. As de-

scribed in Section 3.3.1, our circadian rhythms are maintained by a process

known as entrainment, whereby a group of nerve cells in the human brain use

external information to keep our body clocks synchronized with changes in our

environment (e.g., the Earth’s 24 hour light-dark cycle). The term “zeitgeber”

(zeit: time, geber: giver) is used to refer to these external cues. Light is the most

dominant cue; but a number of factors including temperature, food intake, and

exercise work as zeitgebers too. The following zeitgeber-inspired design im-

plications consider how chronobiology-aware technology could deliver these

circadian cues in order to minimize disruptions and promote stability.

Using light to stabilize the circadian system has been a regular practice in

order to re-establish healthy sleep-wake cycles for sleep phase syndromes [32],

people with travel-induced jet lag, and shift workers [72]. Light is also com-

monly used to improve mood for conditions such as seasonal affective disor-

der [305] and depression [527], including for patients resistant to antidepressant

medications [42]. Recently, light therapy has been extended to other conditions

with symptoms of circadian disruption as well, including to improve sleep, mo-

tor skills, and cognitive abilities in patients with neurodegenerative disorders

(e.g., Alzheimer’s and Parkinson’s) [525, 536] and to improve sleep and day-

161

time alertness in aging individuals [435]. Healthy, young individuals can expe-

rience positive benefits from light therapy too, including improvements in sleep

quality, alertness, and fatigue [506]. Recent research has also demonstrated the

efficacy of morning blue light therapy in reducing social jet lag as well as mini-

mizing the associated losses in sleep quality and daily performance [178].

This encourages the design of chronobiology-aware light therapy tools,

which could assist individuals in getting exposure to either outdoor light or ar-

tificial sources at the times that would help restabilize their circadian systems.

For example, automated smartphone notifications or calendar tasks might re-

mind a person to adhere to a personalized schedule of exposure. Advances in

screen hardware on personal devices (e.g., smartphones, tablets) might even

allow applications to deliver clinically-robust light therapy sessions without

the use of specialized equipment. Novel light-emitting wearables (e.g., glasses,

hats, visors, scarves, wristbands) could be another portable option [405]. Sim-

ilarly, chronobiology-aware homes and offices might adapt the intensity and

wavelength of lighting settings to cue exposure at opportune moments for re-

alignment; however, while this might be straightforward for personal “smart

lamps”, designing smart lighting for larger environments is a challenge, espe-

cially if individuals of different circadian profiles are together in the room.

However, improperly timed exposure to light can exacerbate circadian mis-

alignment [448]. In particular, light at night is a known circadian disruptor.

Given that many of today’s popular electronic devices for reading, communica-

tion, and entertainment have been identified as a main culprit of such disrup-

tion [84], building chronobiology-awareness of their users directly into these

devices could help both to eliminate circadian disruptions in the environment

162

and to move people back toward stabilization. Software applications seem to be

the best approach to controlling a device’s emitted light [151], and some tools

do exist to automatically dim or adjust a screen’s white-blue light at appropri-

ate times of day (e.g., iPhone’s Night Shift feature). The “f.lux” application is

similarly designed to match the light spectrum generated by a screen with the

natural light spectrum of the sun at any given time. This can help reduce sleep

disruptions — assuming, however, that we all sleep and wake by the sun. A

personalized, circadian-attuned version of the software could have default set-

tings based on our available knowledge about the circadian effects of light but

then automatically calibrate (or at least allow users to manually specify) “sun-

rise” and “sunset” times that match biological sunrise and sunset and individ-

ual sleep-wake cycles.

While not as potent a circadian cue as light, temperature is also a major reg-

ulator of sleep timing and duration in humans. In nature, the daily rhythm of

environmental temperature is tightly coupled with the rhythm of sunrise and

sunset. A recent study on the effects of temperature on sleep in pre-industrial

societies found that sleep onset coincided with a nightly reduction in ambient

temperature and that waking occurred just before ambient temperature started

rising for the day [545]. However, this cycle of temperature change is largely

absent in the modern sleep environments of most industrialized societies, with

insulated buildings and artificial heating and cooling systems. The study’s

authors suggest that recreating temperature conditions that the human body

would experience in the natural environment of temperate climates might be

greatly beneficial in stabilizing our own biological rhythms. Existing sleep hy-

giene tools often do encourage users to keep their bedroom at a “cooler tem-

perature”, though their suggested range is a bit loose (60–75◦F); instead, a

163

chronobiology-aware heating system could more closely simulate the nightly

temperature fluctuations experienced in natural conditions.

Another feature of industrialized societies is that food is always available

for most people, who can eat any time they choose. However, animal stud-

ies suggest that eating too much food, too frequently, or at inappropriate times

(e.g., during the day rather than night for a nocturnal animal) can lead to cir-

cadian disruption, particularly with respect to metabolic imbalances [17]. In

humans, regularity in daily routines, including stable meal timings, correlates

with higher subjective sleep quality [343]. Eating meals at times that reinforce

our biological clocks’ innate oscillations may therefore be an effective lifestyle

choice for maintaining healthy circadian rhythms [448]. Further, considering the

public interest today in using technology to manage diet and weight, systems

that provide chronobiology-aware mealtime interventions might pair well with

other PHI applications for healthy eating.

Finally, there are opportunities for chronobiology-aware exercise coaches

to improve sleep and reduce circadian misalignment. Exercise can accelerate

stabilization of circadian systems that have become desynchronized [336] as

well as minimize the circadian disruptions of shift work [144]. In older men,

mid-day fitness training has been shown to improve sleep-wake rhythms [497],

while older adults with insomnia have experienced improved sleep quality af-

ter doing moderate aerobic exercise in the afternoon and early evening [412].

Chronobiology-aware technology could take into account personal characteris-

tics like age and gender in order to supply information about the type, intensity,

timing, and duration of exercise in which a user should engage in order to max-

imize the circadian stabilization benefits.

164

5.3 Biologically-Friendly Productivity Technology

Rather than an immutable trait, cognitive performance is coming to be seen as

a critical component in the overall notion of wellness, with performance opti-

mization emerging as a new frontier in health [475]. At the same time, pressures

to boost work output in today’s increasingly technological and always-on cul-

tures often endorse a mindset that it is possible to maximize — even “hack”

[375] — human performance in a way that would allow someone to sustain

high levels of lasting productivity.

However, such a perspective toward optimizing performance does not ac-

count for both inter- and intra-individual variability in biological characteristics

[504] — i.e., the “internal timing” of the body clock. Beyond sleep, biological

clocks also influence our cognitive performance levels, which naturally rise and

fall throughout the day [79], as described in previous chapters. Alertness, atten-

tion, reaction time, response inhibition, short-term and working memory, and

higher executive skills all follow rhythmic patterns [49].

Relatedly, technologies aimed at supporting productivity are typically de-

signed on assumptions that our capabilities over the course of a day are steady

or could be made steady. Calendars, for instance, typically treat hours and tasks

as commodities instead of helping people schedule in accordance with their

own historical patterns of performance. Notifications arrive at any time of day

or night on the sender’s schedule, not the receiver’s; and though there has been

much research around interruption management [25], it tends to focus on min-

imizing disruption rather than whether a person has the biologically-regulated

cognitive capacity to respond to a particular kind of notice.

165

A greater awareness of our innate biological rhythms could positively

change the way we design such technology. For instance, tools might dynami-

cally adapt to the idiosyncratic needs of their users based on their current or pre-

dicted levels of performance, which could in turn support improved productiv-

ity on a broadly deployable scale. In the previous chapter, I demonstrated how

circadian rhythms of alertness, a cornerstone of cognitive performance [445],

could be measured using smartphone application logs. In this section, I dis-

cuss design opportunities for chronobiology-aware technologies that build on

such assessments in order to support personal productivity in a biologically-

friendly way. Particularly promising classes of technology include those that

raise (inter)personal awareness of performance rhythms, scheduling tools, and

performance-predictive systems. Below, I consider each in turn.

5.3.1 Self- and Social-Awareness

To begin, this line of chronobiology-aware designs aims to help people gain a

better understanding of personal performance rhythms. Since people may not

be aware of their alertness in the moment [139] nor have a good sense of why

and when they experience alertness fluctuations, such systems could help indi-

viduals become cognizant about personal characteristics into which they would

otherwise have little insight. Such self-knowledge could in turn empower a

person to make more biologically-informed decisions when it comes to produc-

tively managing activities — or increase personal empathy and one’s capacity

to understand, accept, and even embrace productivity dips.

166

Given that part of the goal here would be to help people learn about and

come to appreciate their rhythms, a desirable strategy is to convey personalized

alertness levels in an open-ended way that leaves room for interpretation and

self-driven decisions. In the following designs, I therefore represent alertness

levels through a peripheral background that uses a customizable color scale.

By default, a brighter, more yellow saturation corresponds to higher alert-

ness while a faded blue-gray color corresponds to lower alertness. My initial

design used a red color scale (i.e., more or less saturated shades of red). How-

ever, informal feedback-gathering sessions indicated that red was confusing or

put people off; for example, individuals considered high alertness a desirable

state, but that clashed with negative connotations (e.g., danger) that they asso-

ciated with red. I therefore adopted the current color scheme based on prior

design work that found people associate bright yellow with liveliness and en-

ergy, muted blue with the opposite (e.g., low energy, calm, relaxation), and both

with agreeable perceptions [245]. Collecting some follow-up feedback, I veri-

Figure 5.1: To provide peripheral self-awareness, this live wallpaper’scolor transitions in real-time in accordance with the user’salertness levels at that moment. By default, brighter, more yel-low saturation corresponds to higher alertness while a fadedblue-gray color corresponds to lower alertness.

167

fied that the new scheme fit individuals’ mental models of how high versus low

alertness would look as colors, and I confirmed that neither end of the spectrum

was upsetting in the way the red scheme sometimes evoked negative emotions.

I am exploring how individuals react to this idea across several different

media. The smartphone version, illustrated in Figure 5.1, instantiates the back-

ground in a live wallpaper whose color transitions over the course of the day

to display the user’s real-time alertness information. I am also instantiating this

feature in a chronobiology-aware calendar, as shown in Figure 5.2. Compared

to the smartphone wallpaper’s temporary, moment-to-moment view of alert-

ness, the calendar background presents a more holistic overview of entire days,

Figure 5.2: This chronobiology-aware calendar background scaffolds self-awareness of personal alertness levels, which are representedusing a customizable color scale. Visual indicators on eventsprovide an at-a-glance sense of whether scheduling aligns withpersonal alertness at that time.

168

weeks, or longer time periods and also enables a user to see archived views

of past alertness patterns. In addition, visual indicators on calendar events re-

flect the required alertness of that event (using the same yellow–blue/gray color

scheme) in order to provide an at-a-glance sense of whether the timing of an

event aligns with personal alertness levels at that time.

This same information can also be communicated in ways that more inten-

tionally expose it to other people, for instance through personal “beacons”: ob-

jects that are located in one’s environment (e.g., a desk ornament or room light)

or used as a wearable charm (e.g., a necklace, pin, bracelet, ring, etc.), as illus-

trated in Figure 5.3. By providing peripheral cues about internal states, such

displays could not only increase personal self-awareness but could also socially

communicate normally invisible characteristics, for example to create a shared

awareness of alertness profiles among co-workers.

Figure 5.3: The color of personal “beacons” transitions over the course ofthe day to display personal alertness levels in a way visible toother people. Beacons could be objects placed in the environ-ment (e.g., a desk ornament, lamp, etc.) or wearables worn onthe body (e.g., a necklace, pin, bracelet, ring, etc).

169

Such shared awareness could also be valuable in a family or other co-living

setting (e.g., roommates). Viewable from a digital display in the home and

synced to personal devices, a “family portrait” interface as illustrated in Fig-

ure 5.4 could show the individual alertness and sleep-wake patterns of each

household member as a way to coordinate scheduling and foster empathy for

each other (e.g., increasing a family’s tolerance of their teenager’s late sleep-

wake schedule by improving their understanding that it is driven by biology,

not laziness).

Figure 5.4: This “family portrait” (or “roommate portrait”, etc.) interfacesyncs to personal devices and uses a shared digital displayin the home to deliver chronobiology information (e.g., per-formance and sleep-wake patterns) about household membersas a way to coordinate scheduling and improve interpersonalawareness and empathy.

170

5.3.2 Scheduling and Activity Management

In this series of designs, I consider how technology for scheduling activities

could take into account the cognitive demands of those activities and the circa-

dian profiles of involved individuals. If the designs described in the previous

subsection are in fact effective at improving people’s awareness of their personal

rhythms and the substantial impact these rhythms can have on functioning, it

will hopefully increase users’ receptivity to such scheduling suggestions and

motivate the adoption of this more prescriptive form of feedback.

For example, the chronobiology-aware calendar introduced above could as-

sist with scheduling cognitively-intensive versus rote tasks, based on a user’s

chronotype, sleep-wake patterns, and historical alertness rhythms. In the proto-

type I am building, pull-based assistance is provided in two ways. First, a user

is able to specify an event’s chronobiology-relevant information (e.g., required

alertness in Figure 5.5, or required physical exertion, etc.) in order to receive

recommended times to schedule that event based on her circadian profile. An

Figure 5.5: Specifying performance-related parameters (e.g., alertness, asshown here) enables scheduling recommendations that alignwith personal rhythms.

171

attempt to schedule an event at an ill-suited time prompts an overt alert with a

warning and a suggestion for a more biologically-optimal time.

In addition, the calendar’s “Fix My Day” feature offers rescheduling sug-

gestions to better align events with personal performance levels. In the case

seen in Figure 5.6 (where the background gradient is toggled off), the rearrange-

ment eliminates overlaps between the user’s morning sleep inertia phase (as a

reminder, a period of diminished alertness and functional ability) with a meet-

ing and a workout that the user has specified as cognitively and physically in-

tensive, respectively. That is, the calendar suggests alternative times for these

events when it predicts alertness and athletic performance will be higher.

Figure 5.6: This “Fix My Day” scheduling assistant analyzes the day’sevents in order to provide suggestions for rearrangements thatoptimize alignment with personal performance patterns.

172

It is worth noting that the calendar medium could also be well-suited to de-

livering several of the sleep and circadian stabilization interventions previously

overviewed in Section 5.2.2 (e.g., by scheduling recommended sleep times or

zeitgebers such as when to get light exposure, eat, or exercise). This calendar

could also be extended by building in social features, which could go beyond

traditional calendars’ focus on mutual availability when recommending time

slots for group-based activities (e.g., corporate meetings or student study ses-

sions) to additionally consider if most participants are likely to be closer to peak

performance. The user models underlying such a system could further facilitate

team management by helping to pair collaborators or form groups whose mem-

bers are better synchronized in terms of chronotype and performance patterns.

To move beyond the calendar medium and support more flexible timeframes

and high-level activities, Figure 5.7 illustrates mockups of smartphone and

smartwatch clockface widgets that deliver these types of chronobiology-aware

activity recommendations. Both interfaces provide a glanceable view of person-

Figure 5.7: Smartphone and smartwatch clock widgets show glanceableviews of personalized activity recommendations.

173

alized suggestions for the day. As these designs are intended to give users a bit

more agency in their decision-making, the system leaves activity suggestions

more open-ended (e.g., to “work” or to “exercise” during time windows with

fuzzy boundaries). If more hands-on guidance is desired, a user might enable

the tool to deliver notifications with more specific directives (e.g., “go for a 30

minute brisk walk at 12pm on Stewart Avenue”).

Figure 5.8 illustrates an application I am developing to explore the deliv-

ery of these same recommendations through a more playful experience that hits

different design dimensions. In the context of activity management, systems

Alice has some free time…

…so she pulls up her fortune teller-scheduler

She selects activities she’s interested in,,,

…which drop into the fortune cup

She shakes her phone…

…and her fortune drops out

Alice goes for it, or she could have shook again for a different suggestion!

Ludic Activity RecommendationIn-the-moment activity recommendation through a ludic experience

Figure 5.8: Storyboard of a playful chronobiology-aware activity recom-mender that provides in-the-moment feedback through an in-teractive, haptic experience.

174

tend to have primarily pragmatic, optimization-oriented goals. The purpose of

this design is to explore a more ludic approach, for cases when the user is as

interested in fun as in efficiency and when the idea of productivity might re-

late less to work output and more to experiencing a delightful use of one’s time.

Brainstorming sessions with users produced design ideas related to crystal balls

and fortune tellers, which led to this interactive smartphone app that ranks cus-

tomizable candidate activities based on a user’s alertness profile. When the user

desires activity suggestions and is in a playful frame of mind, haptics like shak-

ing afford engagement in this fortune telling experience.

5.3.3 Performance-Predictive Systems

Another fertile area is the development of adaptive tools that can automatically

alter system behavior based on the user’s current alertness levels or predictive

systems that can do the same based on inferred, future alertness levels (or other

personal attributes or indicators that are salient to and valued by a user). For in-

stance, productivity tools that block access to potentially distracting websites or

software might adjust their restricted usage times to match those when it senses

an individual should protect a period of high alertness. Chronobiology-aware

mobile notification could similarly delay the delivery of potentially distractive

interruptions until an alertness lull was detected.

Systems capable of momentary alertness detection as well as future alertness

prediction could also help individuals make more informed choices for them-

selves. For instance, when contemplating whether an all-nighter will either be

productive in the long run or instead lead to diminishing returns, the sleep-

175

decision-support application illustrated in Figure 5.9 would provide a user with

feedback about the “net gains” (or losses, i.e., negative consequences) of stay-

ing awake, in terms of the impact the system predicts potential sleep choices

will have on performance the next day.

Finally, impaired alertness performance can be a serious issue when it comes

to safety, increasing the risk of occupational injury, industrial disasters, and ve-

hicular crashes. Performance-predictive technology might help prevent such

accidents. For example, a driving-intervention application could apply predic-

tive models about a user’s alertness in order to determine whether to deliver a

recommendation to avoid the road until resting, if assessed accident risk was

too high. A more controlling instantiation in a smart car might go even fur-

Figure 5.9: This interactive sleep-decision-support tool provides feedbackabout the net gains or losses of staying awake (e.g., to studyfor a test) in terms of the predicted impact of specified sleepchoices on performance the next day.

176

ther, locking the car’s doors to prevent driver entry or automatically pulling the

car over. Moving beyond the soft sensing approach based on social media and

smartphone app data that I took in the previous chapter’s experiments, a PHI

system’s assessment component could swap in data from alternative passive

sensor streams more suitable to the applied context (e.g., acceleration sensors,

steering patterns, or radio usage for the driving context) that could be used to

continuously monitor performance and deliver just-in-time interventions.

Altogether, the designs presented in this section aim to help individuals

work in better alignment with their natural performance rhythms — either

through adaptive suggestions or by simply helping people become more aware

of their personal rhythms in the first place. An accompanying goal is then to

help optimize performance in a personalized, biologically-friendly manner. In

the face of heavy work pressures, individuals today are increasingly turning to

stimulants to artificially heighten performance and extend working hours [18];

however, a domain-driven perspective suggests that achieving consistently el-

evated alertness is unrealistic and contradicts our biology. I therefore feel it is

important to emphasize that my design ideas should not be posed as helping

people work harder, longer hours. Rather, by incorporating an awareness of

biological rhythms into research on performance and technology, my intention

is to support a vision of systems that are designed to help people adopt more

biologically-suited working schedules and realistic, healthy productivity goals.

177

5.4 Mental Health Management

For this dissertation, my case study research has focused on bringing chronobi-

ology to the development of PHI technology that supports sleep, performance,

and emotional wellness. In demonstrating my domain-driven framework’s

health assessment stage, sleep and performance played the major role. My work

on emotional wellness (including mental health) comes more into the picture

now, in illustrating the framework’s design stage. While my design work in the

contexts of sleep and performance helped demonstrate planning and ideation

as well as building at the level of guidelines, mockups, or barebones prototypes,

my design work in the context of mental health speaks to other aspects of the

framework’s domain-aware design process: building at a high-fidelity, deploy-

ment and review, and iterative trips through the design cycle.

Specifically, Section 5.4.1 presents the building of the high-fidelity system

MoodRhythm, including descriptions of how design goals developed during

planning were translated into system features and how reviews with users and

clinicians helped to further hone these elements. The work presented in Section

5.4.2 then attends to the boundary between review and planning, illustrating

how the design cycle can be re-entered through continued user consultations

(both large and small scale) to inform further refinement of a system’s require-

ments. The end of that subsection then moves into ideation by providing design

implications applicable to MoodRhythm as well as self-monitoring technologies

for managing mental health more generally. Finally, Section 5.4.3 continues to

focus on ideation but demonstrates how the design process is not a unidirec-

tional nor serialized pipeline — in this case, by revisiting the review phase after

ideating in order to corroborate design ideas before undertaking any building.

178

5.4.1 Participatory Design of MoodRhythm

In my research on mental health, I focus on bipolar disorder (BD), which is

recognized as one of the ten most debilitating illnesses worldwide and affects

approximately 60 million people [358]. BD is characterized by episodes of mania

and depression that are separated by periods of normal mood. Manic symptoms

include elevated or irritable mood, hyperactivity, impulsivity, and sleep loss,

while depressive symptoms include inactivity, fatigue, and social withdrawal.

BD is chronic, and there is no cure. People diagnosed with BD expect to manage

their condition for the rest of their lives.

As described in Section 3.3.3, substantial evidence shows that circadian dis-

ruptions are associated with a number of mental illnesses including BD. Sta-

bilization of an individual’s circadian rhythms can be an effective strategy to

reduce condition symptoms — meaning chronobiology-aware designs can play

a major role in supporting self-care and condition management in this context.

Specifically, the Social Zeitgeber hypothesis suggests that the disruption of

certain behavioral, social, and sleep-wake events can disturb circadian rhythms

and, as such, is a causal factor in triggering mood symptoms in vulnerable in-

dividuals [146]. Tracking and stabilizing these routines is therefore considered

a particularly effective non-pharmacologic, chronobiology-based treatment for

diminishing symptoms of BD. The standard practice for tracking lifestyle reg-

ularity and BD symptoms involves paper-based diaries, such as the Social

Rhythm Metric (SRM) [344], which is a central element of Interpersonal and

Social Rhythm Therapy (IPSRT), a clinically validated psychosocial treatment

for BD [170, 331]. However, nonadherence to the paper-based SRM is com-

mon, especially when concentration is compromised during manic or depres-

179

sive episodes; plus, the paper format hinders the synthesis of data into easily-

digestible summaries and feedback.

One promising alternative to paper-based diaries is using smartphone tech-

nology for behavioral tracking and intervention delivery, given its high owner-

ship levels including among individuals with BD [41]. Indeed, mental health

treatment protocols that involve such technology are becoming increasingly ac-

ceptable and advocated by organizations such as the Institute of Medicine and

the National Council for Behavioral Health. A smartphone application also

seems a well-suited medium because it not only permits completion of the self-

report parts of the SRM on a device that is (near) constantly in the patient’s

possession, but a range of BD-relevant parameters (particularly, sleep-wake be-

haviors, activity levels, and social interaction) map well to smartphone sensors

that could automatically detect such information, in ways I describe later.

MoodRhythm [312] is a patient-facing, cross-platform smartphone app built

on the Open mHealth Architecture and developed by a large team of collabora-

tors as part of a participatory design process involving clinicians, professional

psychological researchers, and most importantly, individuals with BD. As illus-

trated in Figure 5.10(a), MoodRhythm helps patients track the five main behav-

iors (getting out of bed, starting one’s day, first social contact, having dinner,

going to bed) used in the standard version of the SRM. Users can also add cus-

tom activities that help anchor their behavioral rhythms, set and track daily

routine-related targets, and record notes.

Reviews with patients and therapists indicated that most BD patients

completed their paper-based SRM entries in batches at the end of the day.

MoodRhythm’s goal was to support momentary use of the app to increase pa-

180

(a) Self-report screen (b) Real-time feedback (c) Badges for adherence

Figure 5.10: Screenshots of MoodRhythm.

tient awareness of daily targets and to limit the impact of memory impairments,

poor concentration, and variations across stages of the illness. To achieve this,

several paths were implemented to streamline the recording process, including

using overt notifications and making it possible for patients to quickly record

SRM events directly from the smartphone notification panel. Another way to

promote momentary self-assessment could be to allow MoodRhythm users to

record information directly from the lock screen — a lightweight self-tracking

strategy HCI researchers are currently exploring, as I mentioned in Section 2.2.1.

The app uses color indicators to provide glanceable feedback for the current

and past days about how well an individual is hitting behavioral targets and

maintaining routine regularity. It also provides weekly feedback using color as

well as natural language summaries, as seen in Figure 5.10(b). Another piece

of clinical knowledge that the design of MoodRhythm takes into consideration

is that individuals with BD have a higher sensitivity to rewards. MoodRhythm

therefore rewards adherence with a variety of badges, as seen in Figure 5.10(c).

181

This feature excited all patients, based on their participatory design feedback

(e.g., “I like the rewards”; “Yes, I LOVE badges”; “I got the badges. It gives me personal

satisfaction that I have completed something. And, that’s good. It keeps motivating.”)

During a four week clinical pilot of MoodRhythm conducted by collabora-

tors at the Western Psychiatric Institute and Clinic (WPIC) with seven individu-

als with BD, the app received very high usability scores from participants, who

particularly appreciated the convenience of recording activities with the smart-

phone medium as well as the way the app provided feedback in real time [2].

5.4.2 Iteratively Refining Design Guidelines

With colleagues, I have conducted both large scale surveys (N=552) [354] and

small scale interviews (N=10) [313] to inform an ongoing design process aimed

at regularly checking in with user needs (e.g., what BD-relevant indicators

would be helpful to build into MoodRhythm to further extend the SRM, does

MoodRhythm fit into existing self-monitoring practices, and how can its design

choices address the currently experienced challenges of self-tracking while be-

ing sure to preserve the perceived benefits). Through these studies, a number

of insights emerged related to what people are tracking, how, and why.

What Are People Tracking?

Participants reported recording a range of indicators, with mood, sleep, fi-

nances, exercise, and sociability being among the most common (see Figure

5.11). I also found that these self-tracking practices could vary and evolve

over time, often in parallel with an individual’s phase of illness. Currently,

182

%

45

0

5

10

15

20

25

30

35

40

Mood Sleep Finance Exercise Sociability

43.4%45.1%

19.3%21.5% 21.9%

Figure 5.11: Percentages of respondents (y axis) who report tracking spec-ified indicators (x axis).

MoodRhythm does provide mood, sleep, and activity monitoring, though these

findings indicate finances would be desirable to support next, as excessive

spending during manic episodes is known to cause financial repercussions

[517]. Further, nearly half of participants in the large survey study reported

that they also track various other items that are relevant to their condition, in-

cluding medication, side effects, and doctor appointments as well as personal

triggers and manifestations of symptoms such as caffeine and alcohol intake,

pain levels, appetite, libido, suicidal ideation, and self-harm. 20% of survey

participants noted tracking items like chores, pet care, leisure time, and recipes

that are seemingly less health-relevant but that they explained help structure

daily behaviors, which can improve symptoms. Such findings reinforce our de-

sign decision to provide support for tracking custom activities and are also a

reminder of the idiosyncratic nature of symptom triggers and manifestations —

further motivating the need for technology-based solutions to move in person-

alized directions.

183

How Are People Tracking?

To track such variables, participants in both studies reported using paper-

based formats (e.g., journals, calendars, sticky notes) as well as digital tools

(e.g., apps, wearables, spreadsheets), along with mental notes, and (less com-

monly) feedback from other people (see Figure 5.12).

36.2%

35.6%

21.8%

6.4%

33.2%

31.8%

32.3%

2.7%

24.9%

53.9%

14.3%

6.9%

18.4%

54.7%

26.9%33.8%

36.5%

29.2%

Social FeedbackMentallyTechnologyPaper

100

0

10

20

30

40

5060

70

80

90

%

Mood Sleep Finance Exercise Sociability

0.5%

Figure 5.12: Percentages (y axis) of various methods (key) used to trackspecified indicator (x axis).

Sometimes, participants reported elaborate tracking setups as necessary to

accommodate personal tracking habits in ways technologies do not currently

support. Figure 5.13 shows an example from our smaller scale study — an Excel

spreadsheet that includes a simple macro to display personalized messages of

encouragement based on patterns in the person’s data. This custom tool gave

this participant a way to record the things most important to her. It also felt

more privacy-preserving to create her own self-tracking tool — another crucial

factor to keep in mind when making design decisions, given the stigma some

BD patients attach to their illness.

184

Figure 5.13: Example of one participant’s custom tracking setup that cap-tures personally meaningful variables, assesses daily status,and delivers messages of encouragement. Diary entries havebeen blurred to protect the participant’s privacy.

Why Are People Tracking?

The majority of participants in both studies expressed that their self-

monitoring practices are beneficial in managing their BD, for example by im-

proving interactions with their clinicians, by offering opportunities for intro-

spection or identity-building, and by helping them gain a greater sense of self-

compassion and acceptance. Participants also explained how the ability to rec-

ognize patterns or pre-cursors to symptoms let them take a more direct role in

their own treatment as well as learn personal coping strategies that worked for

avoiding or recovering from mood shifts. Though their clinicians usually in-

troduced them to self-tracking by way of the paper-based SRM, participants de-

scribed advantages to technology-based self-monitoring tools they had adopted

independently, including that tools make data capture less burdensome, pro-

mote adherence, and provide visual feedback that encourages accountability.

185

Still, our participants encountered shortcomings with existing technologies,

including a lack of support in capturing BD-specific indicators (e.g., well-known

manic or depressive prodromes, mixed moods, etc.) or capturing at a level of

granularity sufficient for BD management (e.g., reporting multiple moods daily)

as well as problems with the usability of interfaces that make them too cumber-

some to use, especially when challenged with symptoms.

Based on these findings, I have derived a set of design implications that

MoodRhythm’s ongoing development continues to pursue. I offer these as

guidelines in Table 5.1 for how tools can be designed to support positive aspects

of self-monitoring yet overcome extant challenges, in ways that meet patients’

expressed needs and are more condition-tailored than generic tools.

Design Implications Needs Addressed Representative QuotationsDeploy software across platforms,devices, and operating systems Pervasive accessibility • “The easier it is to access the program the more likely I am to use it.”

• “I like typing on my work computer but use my iPad and iPhone at home.”

Deliver proactive notificationsPromotes adherence toself-monitoring and behavioralregularity

• “I also just found eMoods for my smartphone. I am just now starting to like it but I need toset an alarm to do it.”

• “I am so chaotic I find it difficult to keep track of anything without help and prompting.”Synthesize data and highlightpatterns

Increases self-awareness andreflection

• “It has helped me see general patterns and to recognize personal triggers. And the moreaware I am of the symptoms, the more I can do proactively.”

Provide encouraging messages andrewards or (after non-compliance)flexibility and forgiveness

Provides experiences of mastery andcultivates self-efficacy andself-compassion

• “My first few episodes I felt intensely guilty about failure. It was this intense guilt thatmade me feel suicidal. Recognizing symptoms of depression has allowed me to be muchmore forgiving during episodes.”

Integrate with clinical care viadoctor-view interfaces, digestiblesummary reports, and modifiablesettings

Facilitates improved acceptance,transmission, and interpretation ofinformation by treatment teams

• “I find that the reports succinctly provide my doctors with a more accurate picture over timethan what I can recall at any given time. It also helps me to create a dialogue with myproviders other than the fact that I don’t feel well (mentally). It has also helped myproviders to see symptoms and patterns that I wouldn’t have thought to mention in short,15 minute appointments.”

• “I tried to share optimism [tracking app] but my Dr was confused by the graphs.”

Provide BD-oriented functionality Allows tracking of indicatorssignificant to BD management

• “I have not been able to find an app that I really like enough to use. One problem withcharting apps is they don’t allow you to chart more than one mood a day. If you have rapidcycling the app is useless.”

Allow customization

Supports idiosyncratic circumstances,preferences, and goals including howindividuals’ conditions, managementpractices, and needs evolve over time

• “Very tedious. Would prefer to customize the computer program to track routine,socializing, etc.”

• “I used the app Optimism for about 4 months...but it only gave feedback/patterns on a fewelements. I then switched to an elaborate excel spreadsheet that provided betterfeedback/patterns, but it ran off my laptop & wasn’t ‘handy’/convenient for tracking whenI have time. Now my day is HIGHLY structured & my mood very stable. I now track in myhead, have daily google calendar reminders, keep a running list to monitor elements, andhave alarms that help with sleeping, eating, etc.”

Implement user-friendly features Alleviates hurdles to tracking,including during mood episodes

• “I used to use a mood app I found on my phone but it was confusing so now I just use thenotes section on my iPhone or an actual paper journal.”

• “I find many of the mood tracking apps overly complex and overly rigid.”

Passively monitor and intervene Reduces user burdens and supportscontinuous capture of data

• “I used to use a calendar on my wall (for tracking), but I had a long mood episode of morethan a month and quit tracking.”

Table 5.1: Guidelines for designing self-monitoring technologies to support men-tal health management.

186

5.4.3 Implications for Preemptive Interventions

A finding prominent among all the engagements with individuals with BD is

their high receptivity to ideas of ”intelligent” technology-based tracking sys-

tems that passively mine behaviors, automatically detect and predict affective

changes, and report feedback about potential symptom onset and appropriate

coping strategies. An electronic version of the SRM still faces many of the same

limitations as the paper form — the burden of remembering to track is still

upon the user; and for individuals with BD, the reliability of self-assessment

(especially during mania) is questionable for any self-report based instrument.

Many of the participants expressed a desire to have tools relieve such self-

management difficulties by proactively providing reminders to self-track or

passively monitoring behaviors without any explicit user input required. By

crisscrossing from ideation to review and back again, this subsection unpacks

this promising design direction.

As mentioned, smartphone sensing capabilities do appear well-suited to the

automatic detection of many key parameters of bipolar disorder that are objec-

tively observable and do not require patients to actively reflect on an internal

state. Based on criteria regarding cognitive and behavioral manifestations of

manic and depressive episodes as well as participants’ own descriptions about

ways in which their technology-mediated activities vary with symptoms, indi-

cators of symptom onset could be passively collected using commonly available

smartphone sensors or usage logging in the following ways: excessive/reduced

activity from accelerometer and geolocation data; increased/reduced sleep from

light sensor data, app use, and social media patterns; and increased/reduced

social activity from microphone, geolocation, and social media data.

187

To verify these potential measures and identify others, collaborators and I

conducted another survey study with 87 individuals with BD [314] to further

inform how a person’s amount, timing, and types of technology use may ex-

hibit measurable differences during manic and depressive periods as compared

to balanced periods. Table 5.2 provides a categorization of the specific mani-

festations of mood shifts that resulted from my qualitative analysis of responses

and that could be used to guide the design of MoodRhythm’s (or other systems’)

automated intervention strategies.

Indicator Manic Manifestations Depressive ManifestationsComputer-mediated communication (e.g.,email, phone, text messages, Facebookmessages, tweets, blog posts)

Repetitively (re)reading and sending messages — to the point it maybe construed as spam — or writing an excess of online content

Avoiding email or dodgingphone calls and texts

Social networking (e.g., Facebook, Twitter) Checking social news feeds repeatedly Avoidance of social media

Web searching and browsing Obsession with performing web searches or managing and rapidlyswitching among multiple browser windows and tabs

Non-use or “idle” use (e.g.,reading websites)

Streaming media (e.g., Netflix, Hulu) “Binge watching” Non-use or “zombie- like”watching

E-commerce Compulsively online shopping Non-use

Gambling and gamingObsessively gambling or playing games on computers or phones forhours — especially social games, high-action games, or multiplegames at once

Non-use or solitary, “calming”games (e.g., solitaire)

Digital calendars Excessively booking activities Diminished useTyping and audio Faster, more careless use (e.g., more typos or more garbled speech) Slower use

Technology-mediated risky behavior

Increased visitation of dating or pornography sites, sending X-ratedphotos, using more inappropriate and aggressive language in writtencontent, or more risk-oriented web searches (e.g., to find tattooparlors or research exuberant vacations)

None reported

Use timing and frequency Late night use: excessively checking phone notifications; or paranoidchecking partners’ emails, social media accounts, or cell phone logs) Diminished use overall

Table 5.2: Variations in technology use identified as characteristic of maniaand depression.

Related research has had similar success inferring mood for individuals with

BD using sensor data from their phones (e.g., the Monarca system [175]), which

suggests the feasibility of using mobile systems for symptom detection and pre-

diction and, in turn, preemptive care and targeted interventions.

However, MoodRhythm differs from such previous work in two key ways:

(1) the foundations of MoodRhythm’s design (the Social Zeitgeber hypoth-

esis, the SRM, and IPSRT therapy) provide a theoretical and clinical basis

188

for the data collected and feedback supplied, as opposed to previous efforts,

which are not influenced by clinically-established characteristics of BD, plus

(2) MoodRhythm’s participatory design process engaged individuals integrally

throughout development — including in long-term, in-situ use of the tool —

in order to ensure its likely adoption, ecological validity, and ability to support

real-life needs. In these ways, MoodRhythm’s design work has been informed

by the chronobiology literature on BD, clinical knowledge about validated treat-

ment strategies, and participatory engagement with potential users.

Earlier sections of this chapter exhibited the same domain-driven design

ethos in other contexts. Altogether, a primary intention of this chapter has been

to put forth and demonstrate an evidence-based approach to design that fore-

grounds a deep understanding of the scientific underpinnings of a targeted as-

pect of health alongside a sensitive consideration of the lived experiences, extant

practices, and expressed needs of users.

189

CHAPTER 6

GENERAL DISCUSSION AND CONCLUSION

In the final chapter of this dissertation, I first look inward, summarizing previ-

ous chapters as well as contributions. I also reflect on my research through sev-

eral discussion points. In particular, I consider the other side of the coin — chal-

lenging this dissertation’s selected assumptions, arguments, and approaches by

surfacing tradeoffs important for system designers to consider. Technology can

be the solution to our health problems! But what if technology is the culprit of those

problems? System-driven approaches relieve user burdens and enable insights

invisible to the unaided human eye! But what if those manual, unaided practices

have personal value? Personally tailored interventions are more effective! But

when does a snug fit become constricting? Data-driven methods are the nemesis

of domain-driven modeling! But could they become friends, in a combined effort

that compensates for each other’s limitations? As part of this contemplation, I not

only acknowledge this dissertation’s limitations, but I offer up concrete strate-

gies for balancing these tensions in order to minimize the risks and maximize

the benefits of future PHI systems.

Then shifting my gaze further outward, I discuss opportunities for future

work to build on the contributions of this dissertation. I first describe how PHI

systems can move beyond the single-user model that HCI has largely focused

on to date, in order to accommodate more socially-oriented health management

practices that extend beyond the individual. I also consider ways in which the

domain-driven framework I have presented in this dissertation could be applied

in areas beyond health, identifying well-suited domains and using my own re-

search to illustrate a concrete example. Finally, I outline several other fertile

areas to pursue going forward and leave the reader with concluding remarks.

190

6.1 Summary of Chapters and Contributions

In Chapter 1, I set the stage for this dissertation. Motivated by the need for

novel solutions to help address a modern crisis of chronic disease, I explained

why personal technology shows promise as an effective strategy for supporting

self-driven health management. In particular, I identified its ability to enable

the delivery of care in a broadly-accessible and cost-effective manner, in large

part due to its increasing ubiquity, technological capability, and user receptivity.

In Chapter 2, I described these motivating ingredients in more detail. I first

overviewed how medicine, disease, treatment, and technology have evolved

and entwined over the last ∼200 years, from the roots of modern medicine to to-

day’s “age of behavior change”. In describing the expanding role of technology

in supporting personal health, I provided a synopsis of the terminology, defini-

tions, and relationships among the prominent work in this area (e.g., “eHealth”,

which encompasses “mHealth”, which intersects with “behavior change tech-

nology”). Because my work comes from an HCI foundation, I particularly fo-

cused on health-related technologies grown in that field (e.g., “persuasive com-

puting”, “personal informatics”, and “quantified self”).

Drawing inspiration from the characteristics of these tools, I then con-

structed a tractable definition of the sort of technology I aim to advance with

this dissertation research: Personal Health Informatics (PHI), a class of tools that

support personal management of healthy behavior via three key mechanisms —

personal data collection, analysis of individual characteristics relevant to a par-

ticular health outcome, and feedback to help a person gain self-knowledge and

potentially change or maintain behavior accordingly. I next mapped the design

191

space of these tools, reviewed how existing systems embody such character-

istics, and provided guidance about how to make choices within that design

space based on the goals of a system and the anticipated needs of its users.

Next, I motivated a central argument of this dissertation, that domain knowl-

edge can help drive this PHI development process in order to fully capitalize

on the potential of these technologies. Specifically, I gave a working definition

of domain knowledge (e.g., theoretical constructs, empirical evidence, practi-

tioner expertise, user perspectives, etc.) and explained the benefits of domain-

informed PHI along with the problematic aspects of more domain-disconnected

approaches. Finally, this chapter primed the remainder of the dissertation by

providing a high-level outline of my framework for domain-driven PHI devel-

opment, with components for how to go about (a) domain inquiry, (b) domain-

driven data collection and analysis to assess health, and (c) domain-aware de-

sign processes to build user-facing tools for supplying health-related feedback.

Each of the following chapters then unpacked those components in more detail.

Chapter 3 explained the process of domain inquiry, which involves select-

ing a problem area; assessing the merit, feasibility, and appropriateness of a

PHI solution; and identifying salient domains from which knowledge is gath-

ered to inform the subsequent stages of development. To demonstrate this in

practice, I used a case study from my own research, describing the value in tak-

ing a chronobiology-driven approach to develop PHI technologies that support

sleep, daily performance, and emotional wellness. To motivate that work and

simultaneously provide a knowledge resource that can support others’ research

in that space, I provided an overview of the field of chronobiology and how

concepts like circadian rhythms, chronotype, and circadian disruption relate to

192

the aspects of health I was aiming to support. Based on this knowledge, I then

devised a soft sensing health assessment approach, scoped to studying sleep-

related circadian disruption as well as cognitive performance in young adults.

In Chapter 4, I undertook the analytic plan that the inquiry process had

informed. Using my case study to demonstrate that domain knowledge can

guide how to operationalize constructs, define variables, and conduct analyses,

I presented the methodology for two experiments exploring how soft sensed

data could provide informative signals about sleep and daily functioning. This

chapter then contributed the empirical findings from these experiments, which

illustrated how phenomena well-known in a domain of study can be detected

and assessed by analyzing a person’s technology usage patterns.

In the first experiment, I focused on examining sleep: diagnosing sleep-

related circadian disruptions, measuring how inadequate sleep relates to

changes in cognition and mood the following day, and overall exploring how

social-sensor data can reflect these trends as part of passive health monitor-

ing. In the second experiment, I dug deeper into daily functioning: building

on chronobiological foundations about cognitive performance rhythms in order

to explore and interpret a number of relationships among alertness, chronotype,

sleep, and smartphone app use — data I found could be similarly leveraged for

passive monitoring. Together, these experiments substantiated how a domain-

driven approach to assessment helps ensure the most significant health determi-

nants, indicators, and outcomes are modeled; conserves computational costs by

concentrating analyses on those targets; aids the interpretation of findings; and

contributes novel methods and scientific insights back to the informing domain.

193

Chapter 5 built on the findings from these experiments, together with

knowledge drawn from the chronobiology literature and interactions with

users, in order to plan, ideate, build, and evaluate designs for technology capa-

ble of providing feedback and interventions that target domain-relevant behav-

iors and health indicators. Continuing to work in the context of my case study,

this chapter contributed chronobiology-aware design artifacts of various fideli-

ties produced by moving through the components of the framework’s domain-

driven design cycle. Specifically, these designs included guidelines for person-

alized sleep support tools that stabilize circadian disruptions, mockups and pro-

totypes for biologically-friendly productivity technology that can increase per-

sonal and interpersonal awareness of rhythms and assist in scheduling, and a

deployed smartphone app that helps manage bipolar disorder by minimizing

circadian misalignments that can fuel the condition’s symptoms.

Which brings us to the current chapter, Chapter 6, where I now reflect on the

research presented in this dissertation along with opportunities for future work.

6.2 Discussion

This dissertation generally promotes a technological, automated, personalized,

domain-driven agenda. However, as with nearly anything, there are promises

as well as pitfalls associated with each of these strategies. In this section, I take

time to ponder the limitations of my research and engage in a sort of tradeoff

analysis. I hope this discussion serves as the starting point of a conversation

I believe is important to have about how systems can be made more aware of

and sensitive to such tradeoffs. Each point of discussion and the ideas I share are

194

guided by my experiences in developing the PHI development framework and

in employing it to conduct the case study I have presented in this dissertation.

6.2.1 Technology as a Double-Edged Sword

One of the central claims of this dissertation is that technology has the potential

to transform health care for the better. I have been a PHI promoter, generally

making the implicit assumption that the more technology is used, the deeper

it is embedded into daily life, and the greater individuals’ adherence to their

PHI systems, the better. I have good reasons for thinking this way — the posi-

tive aspects of technology use have been well documented, and the efficacy of a

PHI system does depend in many ways on heavy usage. In particular, rich data

streams of technology-mediated behavior can act as a window into a person’s

daily life, increasing our fundamental understanding of the factors impacting a

given aspect of health. Further, an abundance of data can also improve the ro-

bustness of modeling techniques to assess those factors and related health out-

comes. In addition, an individual has a better chance of seeing and acting upon

feedback and interventions if they are delivered through a medium with which

that person frequently engages, which is especially important for information

that is context- or time-sensitive.

What I have not discussed (and what the HCI community in general tends

to discuss less — some, but less [272]) are the occasions when technology is not

the solution to the problem but rather, it is the problem. I can provide examples

from two contexts in my case study: sleep and mental health.

195

Despite recommendations to limit technology use before bed, multiple stud-

ies find that nightly usage regularly occurs and can contribute to sleep prob-

lems in several ways [68, 415, 485]: by delaying the onset of sleep; by alerting

the nervous system (e.g., from exposure to light or mentally-stimulating con-

tent), which makes it difficult to fall asleep or achieve restorative sleep; and by

interrupting sleep (e.g., when notifications arrive during the night) [193, 194].

Such light exposure and time-shifted sleep can also disturb our body’s circa-

dian rhythms. I observed similar issues regarding sleep disruption stemming

from technology (over)use or dependence in my own research.

From the interviews conducted with 29 university students during the case

study’s Experiments 1 and 2 (described in Section 4.1), I found that technology

usage bookended sleep for nearly all those individuals. Activities such as check-

ing email, texts, and social media; playing games; and watching videos typically

capped off the day for over three-quarters of participants, and all but one per-

son reported sleeping with their phones on or next to their beds. During these

interviews, the majority of participants described that they set their phones to

silent overnight since it would otherwise produce sleep interruptions; however,

even phantom notifications would sometimes awaken them.

From my work with individuals with bipolar disorder (BD), I similarly

found technology use to be a common disruptor to sleep, which is especially

troublesome in this context given that sleep instability can trigger mood symp-

toms for those with BD [202]. In the N=87 survey about technology usage habits

(described in Section 5.4.3), nearly three-quarters of respondents reported keep-

ing their smartphones bedside at night, 67% of respondents reported losing

sleep due to late-night use at least occasionally, and 45% reported always or of-

196

ten staying online longer than they intend or saying “just a few more minutes”

when online, including at bedtime. In open-ended responses, participants de-

scribed having difficulty falling asleep even once they put the technology down.

One individual’s comment articulated well the interlocking issues of addiction

and sleep disruption: “I often feel anxious about not being connected, and I spend too

much time online, especially on Facebook. I used to use Twitter and LiveJournal every

day, but I’ve managed to cut back on those drastically, and I don’t overshare nearly as

much. But I will definitely forego sleep just so I can check my email and Facebook.”

From this same survey, I also found that technology can have an agitating

effect or even provoke symptoms for individuals with BD. 41% of respondents

reported that technology could trigger their depressive or manic episodes either

by causing social distress or by exposing them to upsetting content. First, just

as offline social anxiety and isolation are known precursors of depression, in-

dividuals described that feeling socially excluded online is similarly disruptive

to their emotional balance. Social media in particular was noted as a common

source of such feelings, often causing a sense of being ignored or left out, jeal-

ousy, or missed opportunity, any or all of which would contribute to feelings of

loneliness and hopelessness and ultimately lead to or amplify their depression.

Second, respondents explained that negative content such as inflammatory or

offensive posts and disturbing news stories could produce anxiety and para-

noia and in turn, trigger or reinforce their depressive symptoms. Respondents

reported that their mania could also be triggered — particularly by exposure to

pornography or posts with sexual innuendo, and frustrating experiences with

technology such as software crashes were also noted as agitating. Most promi-

nently, participants reported that spending too much time with technology is

overstimulating and can fuel a spiral into mania. Further, technology can not

197

only act as a gateway to symptoms for these individuals, but it can also make

the ramifications of episodes more severe. For example, one participant ex-

plained that during a manic shopping spree, there is a limit to how many items

can be carried, stores can be visited, and hence overall purchases can be made

in the real world — but online, there are no such “protections” and the financial

consequences can be devastating. Finally, I suspect some of the emotional and

behavioral patterns tracking tools monitor could themselves become triggers if

reported insensitively or even reported at all. For instance, instead of leading

to useful behavioral changes, knowing about a social “off week” might fuel ru-

mination on social inferiority, depending on how it is presented to the user, her

mood status, and her current ability to act on such feedback.

Thus as technology continues to permeate into people’s personal lives, it

is imperative to further explore the potentially negative aspects of usage, es-

pecially heavy usage. Emerging research is beginning to express a concern

over an unhealthy connection it has found many people are developing with

their phones in particular. One thread of such work focuses on whether mobile

phones pose a hazard to physical health, for instance due to ergonomic issues

or radiation exposure or due to safety risks in the context of driving. Another

area of research is more concerned with the psychological risks of excessive use,

which studies from around the world have found can include anxiety, depres-

sion, guilt, stress, and worry, often due to the wearing strain of a perceived

expectation to be constantly available [319, 484].

Herein lies the tension. On one hand, attachment to our personal technology

may be precisely what gives power to PHI solutions for sleep support and men-

tal health management, given the data frequent interactions generate, the reli-

198

able and continuous monitoring made possible by such granular data, and the

often-attended digital medium for delivering personalized feedback in a timely

manner. On the other hand, prior work and my own studies have discovered

ways in which technology can actually be the culprit of the health problems it is

aiming to correct, for instance by disrupting sleep or exacerbating mental health

symptoms. One participant with BD nicely articulated this tension as follows:

“Technology is a double-edged sword. Sometimes it’s a trigger, sometimes it’s a caution

sign, sometimes it’s a tool, and sometimes it can save your life.”

The question is then: There are a variety of positive aspects to regularly using

technology, but it has its dark sides; how can system designers optimize the benefits

while protecting against the risks? I see an opportunity for “protective” features

whose priority is to dampen the detrimental impacts of problematic technology

use or perhaps preemptively discourage usage patterns that may harm health.

An example of protective sleep technology is software that automatically

dims a device’s emitted light at night. As mentioned, several such tools have re-

cently come on the market (e.g., Night Shift on the iPhone, Twilight on Android,

and f.lux on computers and tablets). In the future, a smart home could protect

sleep as the night winds down by similarly adapting home lighting conditions,

along with additional environmental factors that help regulate sleep (e.g., tem-

perature, similar to the chronobiology-aware home heating system described in

Section 5.2.2). Meanwhile, a bedroom’s sleep-protective sound system could be-

gin playing pink noise, which a recent study finds can induce stable sleep and

improve sleep quality [549]. Delicate delivery of such environmental changes

would be key, for instance through subtle, gradual cues that send the mind and

body a strong signal to sleep while ensuring the home remains comfortable.

199

In the BD context, a protective system might adapt to manic behaviors by

artificially diminishing the performance of a person’s device in order to slow

down interactions (e.g., to buffer against a destructive spending spree by more

slowly loading search results, opening tabs, etc). Design work would have

to proceed extremely carefully though, given my aforementioned finding that

frustrating interactions with technology can themselves fuel symptoms. In ad-

dition, considering the idiosyncratic nature of BD episodes’ behavioral man-

ifestations, a protective system might offer a range of customizable interven-

tions. Then, given the impairment effects that these mood episodes can have

on decision-making, the user could choose from these options during a period

of wellness and provide authorization that the intervention be deployed when

symptoms appear.

Such ideas seem more sensitive and humane than coarse approaches like

simply blocking usage past a certain time at night to guard sleep or shutting

off one’s device when manic patterns are detected. Still, they highlight another

tightrope to walk: developing protective systems that are both effective and

ethical so that protection does not, in practice, begin to more closely resemble

domination or manipulation. This challenge is likely to be further compounded

by the fact that some designers may see no moral dilemma here, instead actively

arguing that such control is acceptable or even desirable if the ends are benefi-

cial enough to justify the means. (Personally, I hope the offspring of protective

systems is not a generation of “For Your Own Good” tools).

Providing user-adjustable settings and manual overrides is one way to help

preserve user control. Or, some protective technologies might adopt an “ad-

vance directive” approach [90] similar to the idea above, where an individual

200

in a healthy state provides consent regarding the level of control she is comfort-

able with a protective technology exerting during periods of infirmity. In other

cases, the implication may simply be to not design [38]. That is, a protective de-

sign implication might be that the best course of action is actually to remove

technology in part or in full from the health management equation.

6.2.2 Manual and Passive Modes

This issue of autonomy relates to another design tension I encountered in my

research: balancing the tradeoffs between user-driven versus system-driven ap-

proaches. As explained in Section 2.2, each component of a PHI system (capture,

analysis, intervention) can be designed to be more or less “participatory” [453],

where a system that is more participatory puts manual control in the hands of

the user, while a less participatory system gives more responsibility to the sys-

tem. Each end of this spectrum comes with advantages and drawbacks.

As overviewed earlier, the main benefits of user-driven approaches relate to

the idea that deliberate, effortful health management activities can afford oppor-

tunities for personal reflection, insight, and growth. Such experiences can culti-

vate mindfulness, self-knowledge, fulfillment, and empowerment. This can be

especially true for users with special requirements, like those managing serious

mental illness, for whom the nature of the condition can interfere with identity

construction and self-tracking can help in building a stronger sense of self. I

found this to be the case in my work on bipolar disorder (BD).

BD can have a considerable impact on individual psychosocial development

and in particular on identity development. The initial onset of BD often oc-

201

curs during the teenage years [392] — a period of significant biological, cog-

nitive, emotional, and social development; plus further identity confusion can

stem from contradictions in experiences of the self across different BD mood

states. As a result, individuals with BD often struggle with difficulties in self-

acceptance due to an inability to integrate these different experiences and for-

mulate a stable sense of identity [221]. In my research, I found that individuals

with BD often integrated manual self-tracking into personal and therapeutic

practices as a means of building such identity. In the N=552 survey (described

in Section 5.4.2), respondents expressed that more manual forms of tracking and

sensemaking made them feel that they were taking a more direct role in man-

aging their own health, which in turn helped them develop a more internalized

locus of control — a perception of agency that they described as instrumental to

recovery. They appreciated that this made them more “active patients” who do

not “simply fill prescriptions”. Individuals explained that the control they could

exert during self-tracking and the intentionality of that process also kept them

calmer since it made their lives feel more structured, manageable, and purpose-

ful. In addition, my findings suggested that not only did manual self-tracking

boost confidence and self-efficacy but that the opportunities for introspection it

afforded were able to foster individuals’ self-compassion and acceptance, which

in turn helped them be kinder to themselves and make more nurturing and un-

derstanding decisions about their mental health.

At the same time, these individuals also remarked on concerns about the

potential fallibility of manual tracking, especially when they felt their ability to

self-assess was compromised during mood episodes. They also discussed how

recall could be difficult not only cognitively but also psychologically, especially

if the recalled experiences were negative or traumatic, as this would actually

202

draw their attention to and accentuate that distress. Considering a move away

from manual approaches, nearly all participants in my studies had high recep-

tivity toward more “intelligent” tracking systems that passively mine behaviors

in order to automatically detect and predict affective changes. They also felt

such passively sensed data could be more “honest” and less susceptible to per-

sonal biases, and a number of participants believed sensor-based systems would

be able to provide insights that they as humans would find difficult to discover

on their own. Still, several people acknowledged that while today’s devices

may be able to enhance human capabilities when it comes to noticing patterns

and making predictions, there is still a limit to the precision they can achieve.

A number of individuals conveyed an awareness of this tension between

agency and automaticity, which is as follows: System-driven sensing can relieve

some of the difficulties associated with user burden and adherence as well as improve

tracking accuracy and pattern recognition, yet such passive approaches diminish inten-

tionality and opportunities for self reflection and can also interfere with agency building

and identity development. Especially as PHI systems move toward more auto-

mated forms of health tracking, it is important to consider this give-and-take.

My work suggests there is an opportunity to design systems that strike an

effective balance between manual and automated activities in a way that fos-

ters self-reflection while still relieving user burdens and remaining reliable. A

number of designers have suggested a hybrid approach, and several passive

sensing systems do allow users to contribute self-reported information. Beyond

this, I think there is room to develop systems capable of moving back and forth

between manual and automated “modes”. For example, a PHI system might fa-

vor manual mode during early stages of use or post-diagnosis in order to build

203

a foundation of agency, self-efficacy, and intimate understanding of one’s con-

dition. During certain periods, however, the more automated mode would be

activated (e.g., during mood episodes in the case of a PHI tool for BD). Then

progressively with time, a system could transition toward more frequent appli-

cation of the automated mode, as a user becomes more familiar and adept with

self-management and would prefer more of a passive monitor that keeps an eye

on her health without requiring much hands-on input. Just as people’s health

management practices, preferences, and needs change over time, I believe there

is value in creating more agile tools that can similarly adapt and evolve.

6.2.3 Avoiding Over-Personalization of Tailored Experiences

As described earlier, recent decades have seen the health domain move toward a

more individual-centric paradigm, with increased personalization of diagnosis,

treatment, and cure. Technology’s recruitment into the battle against chronic

disease has similarly enabled more direct and personally-tailored delivery of

care; and in Section 2.2.3, I specified personalization as one of the key dimen-

sions in the design space of personal health informatics.

I did note that non-personalized approaches could suffice in some cases. For

example, a few studies on increasing physical activity [71] or reducing weight

[229] have found that a generic intervention is better than no intervention. Or,

the personalized strategy may simply not be cost-effective. For instance, before

smartphones and wearables came on the scene, the idea of providing personal-

ized sleep care on a broad scale was considered highly impractical given the spe-

cialized, expensive, and far less accessible equipment required for individually-

204

accurate diagnosis and intervention [423]. Understandably, public health orga-

nizations therefore favored the specification of generic sleep recommendations

over leaving people with no support at all.

In general, however, I espoused the benefits of highly personalized ap-

proaches and reviewed the various cognitive, psychological, demographic, or

contextual factors that can vary between and within individuals and sway

whether or not an intervention is effective. To further champion the person-

alization strategy, I also warned about instances when non-personalized feed-

back could be outright harmful, such as when biologically-generic sleep advice

might introduce circadian misalignment if it does not fit the recipient’s chrono-

type, when a badly tuned intervention might trigger a relapse of symptoms for

an individual with bipolar disorder, or when dealing with vulnerable popula-

tions in general [136, 161, 337].

Still, it is worth contemplating the potential downsides to very tailored ex-

periences. Many of the conversations about this concern are in the context of

search results and recommendation systems, including most recently, personal

news feeds on social media sites like Facebook. Here I focus on two problematic

aspects of over-personalization that are salient in the context of PHI.

One issue is related to a user’s perceptions of systems that feel “too personal-

ized”. This can make a user feel uncomfortable and reject a system as “creepy”

[250], especially if the user does not understand how it is making such spe-

cific personal recommendations [493]. A second concern is that personalization

might lock people into feedback loops that reinforce their baseline attributes,

attitudes, and behaviors (or the system’s model of those user characteristics)

[278]. Related to the idea of “filter bubbles” or “echo chambers”, these digital

205

comfort zones can constrain the range of information to which an individual is

exposed. Feedback that may very well be useful and well-received might be

deemphasized or filtered out entirely by ranking algorithms that judge it as too

dissimilar to a user’s profile, perhaps because that person has not seen, ranked

highly, or accepted (i.e., acted upon) that feedback in the past. This loop can

become reinforcing and self-fulfilling: the user becomes more like the image the

system has of her, as she continues to absorb the information she is provided

and adopt the behaviors she is recommended.

This sort of over-personalization has also been described as a form of bias,

with algorithms perpetuating either (a) the beliefs that their creators knowingly

or unknowingly bake into them from the start (e.g., about what is “worth” mea-

suring”) [360] or (b) the statistical patterns they learn over time — patterns that

are potentially discriminatory in that they reinforce existing, systematic health

disparities that often align with differential outcomes elsewhere in society re-

lated to age, gender, race, sexual orientation, and socioeconomic status [48].

In fact, ”implicit bias” is a well-known problem in health care, where clini-

cians’ subconscious attitudes or stereotypes can affect their understanding and

decisions in ways that perpetuate health care disparities. Consider obesity as

one example, where researchers have identified a reluctance of doctors to look

beyond an obese patient’s weight when making diagnoses or suggesting treat-

ment options, even if the pathway to reported symptoms (e.g., shortness of

breath, pain, etc.) is an underlying condition entirely unrelated to how much

the person weighs [393]. In seeking out domain knowledge in the form of

clinical expertise, a PHI developer who spoke to doctors with such an implicit

bias might construct features in a way that specified higher importance scores

206

(i.e., feature weights, no pun intended) to those related to a user’s weight in-

formation. Or, if that system’s algorithms were trained using medical records

of patients with such implicitly biased doctors, the models might automatically

pick up these patterns and similarly promote weight-focused diagnoses and

treatments. Either way, a well-intentioned domain-driven system could end up

propagating such bias.

Altogether, the tension is then: While personalization is associated with numer-

ous benefits, how can a system avoid over-personalization, where feedback locks users

into potentially creepy, constrained, or even discriminatory experiences? One idea is

to design PHI algorithms that foster risk-averse serendipity. In the context of

information retrieval, a system’s measure of accuracy is usually based on sim-

ilarity — that is, the degree to which a piece of content delivered to a user is

similar to what that person has already enjoyed (e.g., seen and rated highly)

in the past. However, this is what locks individuals into cocoons that can dra-

matically reduce their chances of seeing something completely different. The

notion of serendipity (well-discussed in the recommender systems community

[264]) refers to a system helping a user break free from these similarity clusters

by supplying content that is diverse, novel, and unexpected.

For example, a less serendipitous PHI tool to support smoking cessation

might deliver feedback related to a standard set of cessation strategies such as

personalized encouragement messages or reminders about adhering to one’s

nicotine replacement therapy. Instead, a system designed to promote more di-

verse experiences might also provide fitness-based interventions (that the less

serendipitous tool likely would have demoted as less relevant than traditional

interventions in this context). Such an approach might take further inspiration

207

from health interventions employing “cognitive dissonance” (which generally

see positive outcomes [172]) — i.e., motivating behavior change by helping indi-

viduals come to recognize the difference between their current status and envi-

sioned self-image. For instance, the serendipitous smoking cessation tool might

choose to suggest fitness activities that would initially be extremely challeng-

ing for a heavy smoker, in order to accentuate dissonance and motivate change.

More generally, by promoting exposure to fresh and unexpected feedback in

such ways, a serendipitous PHI system could help to challenge and expand a

user’s comfort zone in order to encourage positive progress.

Various algorithmic approaches could be employed to produce these sorts

of serendipitous feedback deliveries (e.g., after clustering results, using an item-

voting scheme that selects items from different similarity clusters proportionally

to their relevance scores [480]). Moreover, an important aspect of a serendipi-

tous PHI system is its “risk-averse” nature, which builds in a measure of pro-

tection in order to guard against the delivery of content that is so mismatched

with a user’s profile that it could risk dangerous outcomes. Domain knowledge

could help in tuning the thresholds for a system’s risk parameter (e.g., based

on the vulnerability of a user, the severity of her condition, and the health haz-

ards linked to that illness). For example, a PHI tool for managing BD would

be extremely risk-averse considering the potential for misguided interventions

to be life-threatening in this context. Further, given the fact that consistency in

one’s routines is immensely helpful in improving BD symptoms, designing for

excessive novelty, unexpectedness, and irregularity would likely be undesirable

in this case. On the other hand, a fitness PHI tool might consider the perils of

divergent feedback to be more minimal, tuning its parameter in a way to in-

stead minimize the chances of user dissatisfaction or abandonment — although

208

within limits. For instance, recommending that a novice, unfit user (e.g., the

heavy smoker mentioned above) do extreme interval training for the sake of

serendipity or cognitive dissonance could indeed be unsafe.

In addition, PHI interfaces can be designed so that not just the content but

also the presentation of feedback facilitates serendipity. For example, search en-

gines sometimes avoid ranked lists in favor of grid views or organize results into

category views that promote exploratory information discovery. Design choices

in terms of how feedback is presented as well as the user’s control over that pre-

sentation can also help mitigate problematic aspects of personalization related

both to bias as well as a user’s discomfort with feeling monitored. First, provid-

ing transparency or context about why a particular piece of feedback is being

delivered can increase user satisfaction, as it helps a person understand how ex-

actly her behaviors came to influence the feedback she is receiving [458]. Given

that there are various ways that personalization can fail, such transparency can

help a system fail gracefully. Transparency can also diminish a sense of creepi-

ness, in large part because transparency builds trust — something particularly

important in a health context, where users might have special concerns about

confidentiality, privacy, sensitivity, and stigma. By exposing how the system

works, transparency can additionally provide users with a sense of control.

In fact, the ability to control personalization mechanisms can be crucial to

their acceptance [493]. A PHI system can give users control in various ways,

for example by providing a means of deferring behavioral recommendations

or through options to turn personalization on and off. The ability to turn off

personalization can help assuage privacy concerns as well as combat bias; a

variation might be to provide a setting that lets a user see feedback personal-

209

ized for a different profile (e.g., “what would I see if the system thought I was a

different gender?”). Further, design choices like the name of a system or how

its functionality is framed can change users’ receptivity to the information it

provides; for instance, a user might be more accepting of a PHI fitness “coach”

that supplies novel and unexpected (and sometimes imperfect) feedback than

the identical system branded as a PHI fitness “recommender system” [40].

6.2.4 A Dual (Not Duel) Domain & Data Driven Approach

Another core tenet of this dissertation is the benefit of taking a domain-driven

strategy when developing PHI technology. Indeed, I spent a considerable

amount of effort describing the value of domain-driven PHI (e.g., see Section

2.3.2) and the trouble with domain-disconnected efforts (e.g., see Section 2.3.4).

I particularly took issue with more data-driven, “black box” approaches where

models are guided by patterns noticeable in the data rather than on theory, em-

pirical evidence, or other validated knowledge, arguing that such models could

be inaccurate, irrelevant, misleading, or even disrespectful of human complex-

ity. In this Discussion’s previous subsection, I further pointed out that purely

data-driven models might reinforce pre-existing biases in potentially unfair, dis-

criminatory, and harmful ways. In contrast, I argued that domain knowledge

could thoughtfully inform all aspects of system design from data collection, to

analysis, to intervention, to evaluation. By guiding the important health de-

terminants to target, avoiding the computational costs of modeling unneeded

features, aiding the interpretation of outcomes, and enabling the refinement of

existing theories, I do believe a domain-driven approach can drive the develop-

ment of systems highly effective at improving people’s health.

210

That being said, domain-driven approaches can have limitations. First, do-

main knowledge may be biased, incomplete, or simply unavailable. Especially

when working in an understudied or novel area (e.g., relatively uncommon

health conditions), pre-existing scientific literature is likely to be sparse. In such

cases, one can try to generate domain knowledge, for instance by engaging with

the anticipated users of a system to determine practices or needs; however, lack

of access to such individuals may be similarly challenged, especially if the pop-

ulation is small, remote, or marginalized.

In other cases, knowledge might exist but its relevance is questionable. For

example, empirical evidence might have arose from a context very different

from that of the current PHI project (e.g., many years ago, with a very differ-

ent population, or under otherwise substantially dissimilar circumstances). In

an increasingly diverse world of evolving health conditions, evidence can lag

behind; and in today’s rapidly changing technological climate, cutting-edge re-

search about user characteristics or usage tendencies may become obsolete even

after a short period of time.

Furthermore, there will always be a gap between the health needs and id-

iosyncratic symptom manifestations of any one particular individual with ag-

gregate findings (e.g., from a large-scale randomized controlled trial, or RCT)

— if any such RCT has even been conducted. Just as I mentioned that some per-

sonalization strategies are cost-ineffective, the time and resources required to

conduct RCTs for some conditions preclude them from ever being realistically

conducted. For example, it has been estimated that it would require 127 RCTs

involving 63,500 patients over 286 years to produce the evidence necessary to

inform clinical decisions about Alzheimer’s [442].

211

In addition, there is a well-known publication bias against reporting null re-

sults, regardless of the quality of study design [143], which not only means that

relevant knowledge may exist but be unavailable, but there could be a bias in-

herent to relying on the information that is available. Further, the flip side may

be possible too: flawed results (e.g., from poorly designed or executed studies)

do get published — unreliable knowledge that future researchers would mistak-

enly deem as a trustworthy foundation to build upon. While a well-seasoned

researcher may be able to identify such red herrings, this may not always be the

case, which relates to my next and final point.

Lastly, domain knowledge can be applied inappropriately. This disserta-

tion’s framework, as I mentioned when I introduced it, is intended to serve as

a way of thinking, a methodological process guide rather than a binding set of

cookbook recipes. A domain-driven researcher should use the best available

evidence to make informed decisions — but this process will still involve per-

sonal judgment to some extent (e.g., in order to determine what that evidence

might be, whether it is in fact applicable, and if so, how to apply it). However,

this takes time, effort, and practice. If a research team does not have the nec-

essary skills, interest, or patience to rigorously scour, assess, and apply domain

knowledge, then a shallow application can be equally or more problematic than

a process absent of domain knowledge entirely. Researchers taking a domain-

driven strategy but expending minimal effort are prone to confirmation bias —

noticing, incorporating, and interpreting only the conveniently-acquirable evi-

dence that confirms their pre-existing beliefs, previous experiences, or personal

judgments, rather than alternative and potentially more relevant information.

212

For example, a common pitfall of HCI researchers applying the transtheoret-

ical model to develop behavior change systems is to consider only the model’s

stages of change but not its constructs related to decisional balance or self-

efficacy. Not integrating these aspects of the theory diminishes the potency of

the full conceptual framework for designing a system, plus it makes it more

difficult to evaluate outcomes of the system, for instance to explain why the

system is or is not effective for some users (e.g., those individuals with or with-

out strong self-efficacy) [206]. Or, as I mentioned in Chapter 2, reviews find that

sometimes systems mention drawing on theory but supply few details or ratio-

nalizations [259, 382]. As I pointed out at the end of Section 2.3.3, this leaves

it unclear in many cases whether those theories were actually the most suitable

— or simply the most familiar and in vogue, given a cascade of prior work that

had already been drawing on them.

Therefore to resolve this tension, we must consider: Are there worlds in which

domain-driven and data-driven methods can peacefully co-exist — or even further, lead

to PHI solutions more favorable than either could achieve independently?

I believe combining these two approaches, domain-driven and data-driven,

provides a promising research direction: guiding practical, data-driven meth-

ods with responsibly-gathered knowledge from both the health and HCI do-

mains (e.g., knowledge about the health condition being managed; needs, val-

ues, preferences, and practices of patients/users; advantages and limitations of

available treatment options and relevant PHI tools; and applicable computa-

tional and interaction principles).

213

This process would be one of iterative exploration and experimentation,

punctuated by critical reflection about what has been discovered, worked,

failed, and how to proceed. Speaking back to the limiting case where domain

knowledge is lacking (e.g., unavailable, incomplete, or biased), data-driven ap-

proaches can help to address such challenges.

In particular, data-driven identification of patterns can be useful in con-

structing hypotheses, which can then be evaluated and, over time, solidified

into verified theories themselves. In addition, such patterns might have been

imperceptible via human-driven appraisals alone. This means data-driven ap-

proaches can help researchers generate hypotheses and, eventually, domain

knowledge that we would have otherwise found difficult to construct. Further,

going back to the bias issue, these previously unrecognized patterns might cor-

respond to societal inequalities currently in our “blind spots” that data-driven

computation helps make noticeable.

Altogether, I look forward to a future where data and domain knowledge can

live side-by-side in a symbiotic relationship that helps to advance PHI solutions

and our current health challenges.

214

6.3 Opportunities for Future Work

My overarching and long-term research agenda is directed at supporting well-

being through positive interactions with technology. Throughout the course of

developing a framework for domain-driven development and conducting my

case study research, I have identified several additional areas I believe it would

be worthwhile to continue exploring. In this section, I discuss these opportuni-

ties for future work that I see as particularly compelling.

6.3.1 Moving From Personal to Collective Informatics

In this dissertation, I reviewed a number of tools developed in recent years to

support health management, and I presented domain-driven guidelines I ar-

gued could positively impact the way we create future generations of such tech-

nologies. I particularly emphasized a focus on the individual — relaying visions

of people at the center of their own care, using technologies that support self-

driven health management via personally-tailored experiences. I purposefully

included the word “personal” in my conception of “personal health informat-

ics” to stress this focus. Most extant systems and implications share the same

model and have been developed and deployed from a single-user perspective.

However, health management practices are frequently embedded in social

contexts. From data capture to sensemaking, individuals enact a range of social

practices. For example, in my work with individuals with bipolar disorder, I

found that social feedback was sometimes used as a form of condition monitor-

ing, whereby individuals would periodically report their mood to loved ones or

215

would depend on trusted connections to act as a sort of “human sensor” who

would pay attention for warning signs of a mood episode [354].

Engagement with collected data (whether captured solo or socially) can ex-

tend beyond the individual as well. For instance, individuals can use per-

sonal records to build understanding among family and friends, share data with

groups of peers (e.g., individuals with the same condition) for social support, or

transmit data to caregivers to facilitate treatment oversight. I found examples

of this in research I have done on PHI for pain management, where interviews I

conducted with individuals experiencing chronic pain revealed that they would

sometimes use their pain logs to gain empathy from their family members (e.g.,

by using personal data to “prove” they were experiencing severe pain) or to in-

crease credibility when discussing treatments with doctors (e.g., by using data

as “hard evidence” to substantiate intuitions about their pain’s fluctuations or

their medication’s efficacy).

Likewise, in the N=552 survey with individuals with bipolar disorder (de-

scribed in Section 5.4.2), about two-thirds of respondents reported using self-

tracked data with doctors, psychiatrists, or therapists as a way to open and

maintain lines of communication. Specifically, individuals explained that this

personal data helped them relay symptomatic patterns in a more aggregate and

accurate manner than they could have done without the assistance of technol-

ogy. They also mentioned that data enabled them to more accurately recount

behaviors and events, especially if significant time had passed since a prior ap-

pointment. Additionally, similar to the pain patients I met, these individuals

with bipolar disorder felt that data gave them defensible evidence for discus-

sions about treatment efficacy or medication adjustments.

216

I believe there is a considerable opportunity to design computing infrastruc-

tures and interfaces that can better support such collective practices across all

phases of health management. Some technologies do include social features,

but they are typically geared toward comparing performance with other users

(e.g., through leaderboards or competitions like Fitbit’s “Who can get the most

steps this weekend?” challenge) or providing group discussion spaces (e.g., on-

line health fora to seek information or social support). Peripheral public dis-

plays support interpersonal awareness to some extent (e.g., the designs I pre-

sented in Section 5.3.1 to promote social reflection about an individual’s per-

sonal circadian rhythms), but such work is still nascent in the context of health

management. Going forward, novel collective informatics technologies could

significantly enhance even deeper social engagement with the collection and

sensemaking of others’ health data, in the ways my own research experiences

have demonstrated individuals and members of their support networks desire.

As with the design work I presented in Chapter 5, user-centered participa-

tory design processes could be undertaken to (1) more fully understand col-

lective informatics practices, including who are the involved parties (e.g., care

providers, family members, close friends, extended networks), how they engage

with each other, and how technologies can facilitate these interactions and to (2)

develop innovative collaborative computing architectures, data representations,

interaction principles, and interface designs that meet these social needs. Such

developments could help in scaffolding social support networks that play im-

portant roles in health management, increase the accuracy and utility of tracked

personal data (e.g., by enabling a trusted circle of stakeholders to participate in

its collection, verification, and analysis), and provide caregivers with new types

of information that might support more effective treatment.

217

There are also a number of critical issues in this space that must be identified,

documented, and designed for. For example, applications for collective forms

of data capture, sharing, and sensemaking must be sensitive to data manage-

ment (e.g., data ownership, privacy, and access) as well as how to protect the

autonomy of an individual when a group of people come together through a

mutual interest in his or her wellness. Overall, such responsibly designed so-

ciotechnical systems could lead to a number of positive impacts at individual,

group, and societal levels.

6.3.2 Applying the Framework in Areas Beyond Health

While this dissertation has demonstrated the utility of a domain-driven ap-

proach in developing PHI — personal health informatics — an important ques-

tion remains regarding generalizability: Can a domain-driven technology develop-

ment framework be applied in domains beyond those related to health?

As described in Section 2.2, PHI is essentially a subclass of personal infor-

matics tools — the “H” is added precisely to distinguish PHI from the broader

definition of personal informatics that can encompass a wider range of personal

data. Seminal personal informatics literature (e.g., the publication that coined

the term) has identified a number of domains where such tools have been or

could be applied. Many are related to some aspect of physical, cognitive, or

emotional health, as I have explored in this dissertation. Others that stand out

include sustainability, information seeking, and online contribution. Examin-

ing the generalizability of the domain-driven framework in the context of these

domains therefore seems like a reasonable thing to do.

218

To begin, academic and industrial HCI researchers have developed various

tools to promote sustainable behaviors (e.g., for energy consumption, water us-

age, transportation, and waste disposal). One line of work studies the design

of “eco-feedback” technology, which uses personal devices (e.g., smartphone

apps) or home-integrations (e.g., sensors and lights on faucets) to measure and

increase environmentally-friendly behaviors. For example, LCD displays like

the Energy Orb have been used to display daily energy usage levels in a descrip-

tive, peripheral way, while other systems like Kill-A-Watt or Energy Detective

have more prescriptively provided persuasive prompts to reduce usage during

peak hours when energy costs are high [205]. Similarly, the Shower Calendar is

a prototype system that displays personalized water consumption information

via a screen or projection located in the bathroom in order to foster awareness,

family competition, and ultimately reduction of water use [276]. Systems like

the Electricity Portal provide a more societal-level (in this case, city-wide) por-

trait of energy consumption that uses household-level feedback, incentives, and

social comparisons to promote conservation [150].

I would argue that such technologies are very well-suited for a domain-

driven development strategy, as seen by walking through the components of

the framework. First, it is possible to identify a rich body of knowledge that

provides insights into why people do or do not tend to engage in environ-

mentally responsible behavior (e.g., from individual characteristics like altru-

ism or fiscal concerns to group level considerations like societal norms) from

domains including education, economics, philosophy, psychology, and sociol-

ogy [173, 174]. Further, while PHI systems target health behaviors and systems

like Energy Detective target conservation behaviors, both are rooted in behavior

change [69]. This makes many of the domain-driven development stages quite

219

similar: drawing from behavioral theories (e.g., goal setting, social compari-

son, and positive/negative reinforcement have been employed in this context

of sustainability technology [472]) in order to inform how to collect, model, and

provide feedback about those behaviors, ultimately in order to increase oppor-

tunities for reflection and positive change.

The next application area, information seeking, is less analogous to behavior

change yet a domain-driven approach still seems applicable. While most in-

formation retrieval systems use an inductive approach (i.e., devising user mod-

els by analyzing patterns in users’ interaction data) [292], a few have shown

it is possible to take a more deductive, domain-driven approach. Specifically,

such work has utilized established behavioral theories (e.g., information forag-

ing theory or economic utility theory) in order to guide their modeling of users’

search strategies and information goals and inform the development of more

personalized search systems (e.g., for images [292] or work-related information

like emails or documents [67]).

As a final example, I have had success at applying the domain-driven frame-

work in the context of online contribution. Attracting new members to on-

line communities and encouraging substantive engagement are open and com-

pelling problems at the intersection of multiple disciplines. My inquiry into

social psychology theories as well as empirical studies related to community

participation helped me identify a link between effective contribution and the

psychological construct of self-efficacy [26, 215, 516, 543]. In addition, inquiry

into the psycholinguistics literature informed me about text analysis methods I

could apply in order to measure various personal attributes that the social psy-

chology theories suggested as relevant to self-efficacy. This groundwork moti-

220

vated me to continue on a domain-driven approach to assessing and designing

around this personal trait.

In essence, my analytic plan was to model self-efficacy using linguistic fea-

tures of an individual’s passively mined text-based data. I therefore constructed

various semantic, syntactic, and stylistic features intended to represent salient

characteristics of self-efficacy according to my gathered domain knowledge

(e.g., to operationalize attributes such as self-confidence, self-regulation, execu-

tive skill, critical thinking, and social competency) [27, 211].

I then collected data by acquiring two existing experimental datasets that

were well suited to the task, would let me explore my domain-driven approach

relatively cheaply, and would also help me assess its generalizability across

more than one social context. In each experiment, the participants had con-

tributed comments in an ad-hoc online community environment modeled after

RegulationRoom [321], an established website designed to facilitate the con-

tribution of feedback about proposed regulatory policies. For both datasets,

a variety of participant information was also available, including self-efficacy

measured using a validated survey [87].

Before moving onto the assessment phase of the framework (i.e., attempt-

ing to assess participants’ self-efficacy from their text-based data), I addition-

ally used domain knowledge to verify a key assumption: that high self-efficacy

actually translates into “effective” contributing behavior. Here, I used a defi-

nition of domain knowledge that included organizational theories, experimen-

tal evidence (e.g., about predictors of answer quality in Question & Answer

sites), and publicly-posted community guidelines (e.g., Wikipedia policies) in

order to define generally accepted metrics of effective contribution: comment

221

volume, length, and quality. Then comparing participants with strong versus

weak self-efficacy (where the threshold was informed by prior literature [394]),

I confirmed that individuals with strong self-efficacy did post statistically sig-

nificantly more, longer, and higher quality comments.

Next, for both datasets, I compared individuals with strong versus weak

self-efficacy and found statistically significant differences for nearly all my lin-

guistic features. I additionally examined whether these features could predict

self-efficacy, treating the task as a binary classification problem using the same

strong versus weak self-efficacy groups. Using 10-fold cross-validation with a

C4.5 decision tree algorithm, I found that my linguistic features were able to

predict self-efficacy with reasonably good accuracy, precision, and recall and

far outperformed a baseline model that returned the majority class.

I could have taken a data-driven approach to analysis, for instance by in-

cluding numerous additional features (e.g., n-gram frequencies) and then using

feature selection or other statistical techniques to prune. Instead, I focused on a

grounded operationalization of psychological traits, constructing features that

domain knowledge suggested would be more relevant to the task from the start.

To evaluate whether this domain-driven approach actually boosted predic-

tive power, I performed the same binary self-efficacy classification task again,

this time using more data-driven feature sets available from three recent rel-

evant studies [186, 307, 544]. In fact, I found that my domain-driven set de-

livered the best performance, as it incorporated informative features that were

not captured by these other sets’ dictionary-based categories. My intention is

not to imply these studies are methodologically arbitrary but rather that they

are representative of data-driven approaches that apply off-the-shelf text ana-

222

lytics instead of developing custom features to represent constructs of interest.

Encouragingly, these results supported the idea that a computational strategy

rooted in domain knowledge can enhance a phase of assessment (in a domain

outside health).

Lastly, I was able to identify several ways these findings could help cultivate

more fruitful communities through a domain-aware design process. For ex-

ample, community recruitment could be enhanced by targeting self-efficacious

individuals more likely to be willing to engage and do so in valued ways. Or,

an intelligent task routing approach [110] that leverages self-efficacy profiling

could improve the chances that contribution tasks would be funneled to indi-

viduals more likely to be willing and able to do them. Such an approach might

even help to iteratively foster self-efficacy, given that the personal fulfillment

and social recognition of completing tasks could provide members with mas-

tery experiences, which is an effective way to develop a stronger sense of self-

efficacy [27]. Finally, the increased understanding this work contributed about

the relationship between self-efficacy and online contributing behavior could in-

form the design of moderator protocols to better direct attention or personalize

engagements with users in ways suitable to their psychological characteristics.

Overall, my work indicated the generalizability of the domain-driven frame-

work to online contribution. In particular, by applying the framework in this

context, I was able to show how a suite of domain-driven features has use-

ful prediction power, including compared to commonly used feature construc-

tion strategies that apply off-the-shelf dictionary-based methods without much

eye to theory — bolstering the notion that a domain-driven approach may gen-

uinely enhance assessment performance and, in turn, design opportunities.

223

6.3.3 Open Challenges of Data, Analysis, and Design

In addition to broadening PHI beyond a single-user model and expanding the

framework to applications beyond health, a number of other open challenges

exist that warrant further investigation. I focus on three issues that each relate

to one of the three components of PHI technology: data concerns, analysis con-

siderations, and possibilities for novel forms of feedback that engage users with

PHI systems in new and meaningful ways.

Entire dissertations could be devoted to various facets of these topics. Nev-

ertheless, while the following paragraphs cannot fully address all the myriad

considerations, I believe there is value in at least recognizing these as important

talking points, articulating their links to my particular dissertation, and offering

concrete directions for the HCI community to pursue going forward.

Data Management — Ownership, Privacy, and Security

As increasingly diverse and granular personal data streams continue to be

generated, either knowingly by individuals interested in tracking that informa-

tion or as a passive byproduct of one’s interactions with technology, it is impor-

tant to contemplate a number of concerns related to the responsible manage-

ment of that data — especially considering health data is highly personal and

potentially sensitive, stigmatic, and exploitable.

For instance, a move toward personal informatics paradigms that better sup-

port collective engagement, as described earlier in Section 6.3.1, will throw is-

sues of data ownership, privacy, and access into sharp relief. In addition, data

collection that is passive, streaming, and “always on” can hinder individuals’

224

ability to practice intentional self-disclosure and raises issues of privacy rights

and data curation. An important step is determining the various motives, ex-

pectations, and concerns users would have regarding what data would be ac-

ceptable for systems to handle and how. There are numerous opportunities for

research to examine ways to go about establishing legal standards and user-

friendly terms of service; increasing systems’ transparency in how data is col-

lected, stored, protected, brokered, or otherwise managed; and more deliber-

ately providing usable mechanisms for opting in or out of certain aspects of

digital health services. In pursuing these directions, some scholars have sug-

gested that personal health data could be reimagined as a child, treating data

with the same rights and responsibilities as would be given to a vulnerable,

precious loved one [368].

The smartphone app I have been involved in developing to help people

manage bipolar disorder (MoodRhythm, introduced in Section 5.4) builds in

privacy-preserving mechanisms in several ways. For example, the system does

not record audio but rather processes it in real-time on the smartphone in order

to only extract and store features (e.g., spectral content, loudness) that are use-

ful for detecting the presence of a human voice but insufficient to reconstruct

speech content [537]. Using these privacy-sensitive audio features and proba-

bilistic inference techniques, MoodRhythm is able to estimate whether a user is

engaging in a healthy level of social interaction. In addition, accelerometer data

is used to classify and securely store a user’s activity state as a binary active ver-

sus sedentary. Attempting to model specific activities would be more privacy

invasive and computationally expensive — yet not necessarily more useful in

detecting symptoms than MoodRhythm’s high-level, more privacy-preserving

representation. In fact, a domain-driven approach that understands the specific

225

data that must be targeted and at what minimum level of granularity neces-

sary to support management might better protect users’ privacy, be more well-

received by users, and more easily comply with future standards.

The Elusiveness of Causality

When analyzing personal data or behaviors to assess health, observed re-

lationships can be complex and multifaceted. In particular, I encountered an

issue of elusive causality throughout my case study on sleep, performance, and

emotional wellness. In Experiments 1 and 2 (described in Chapter 4), poor sleep

could manifest in technology use, technology use could result in poor sleep, and

both could have been modulating or modulated by a number of other factors. In

many participants, I observed what seemed to be a cycle of disruption wherein

they would get insufficient sleep, cyberloaf the next day due to problems with

attention, and ultimately again lose sleep that night by staying up to compen-

sate for such unproductive time. Observed links between sleep and mood were

similarly difficult to unpack in those studies, as negative mood may both reflect

and cause poor sleep, and both negative mood and poor sleep may be indicative

of other underlying factors such as depression or stress.

In my studies on bipolar disorder, I ran into a similar chicken-and-egg ques-

tion between technology use and condition symptoms. Many individuals dis-

cussed their uncertainty about whether particular patterns of use would gen-

erally trigger and fuel bipolar episodes or whether being manic or depressed

would lead to distinct levels and types of technology use (e.g., “I also start to

notice that I lose more and more sleep because I’ve been up late, usually online. Is

losing sleep triggering episodes, or is the episode triggering my staying up late?”). I

226

also found that this interlaced relationship between usage and symptoms could

become a positive feedback loop. For instance, somewhat elevated energy and

mood levels could lead an individual to initiate conversations on a dating site

or play games longer than usual — behaviors that would then increase stimula-

tion, heighten energy and mood further, and eventually spiral into a full-blown

manic episode. As another example I commonly observed, an individual expe-

riencing stress would lose sleep as a result, increasingly use technology during

periods of this insomnia, be more susceptible to negative posts online, experi-

ence stress yet again, and eventually fall into a phase of depression.

While it was beyond the scope of this dissertation to fully disentangle such

intricacies, they are important to look for and attempt to unravel during PHI

analyses, so as to avoid making mistaken causal assumptions as well as to de-

termine what behaviors to actually target for intervention.

Novel Feedback Formats

In Section 2.2.3, I identified the ways in which PHI systems commonly con-

vey feedback to individuals (e.g., using text, charts, visual metaphors, sounds,

or vibrations). Being able to understand and derive value from such informa-

tion requires a new form of literacy on the part of the end-user as well as new

communication strategies on the part of PHI system developers. Moving be-

yond conventional approaches to data visualization may be particularly benefi-

cial in some contexts, such as PHI tools for vulnerable populations with special

requirements. For instance, while feedback can support healthy self-awareness,

I found that many individuals with bipolar disorder struggle with traditional vi-

sualizations, which sometimes lead to hyper self-scrutiny, unrealistic normative

227

expectations of health or identity, or a distressing clash between the smoothed

curves often emphasized by standard graphs and the sharply erratic fluctua-

tions they associated with their own moods.

Recently, HCI researchers have begun exploring novel ways to engage

users in their health information, for instance through storytelling experiences

[179, 487] or tools that allow users to build custom visualizations themselves

[21]. In my research, I am particularly excited about personal informatics

games: tools that provide gameful and playful approaches to data capture, self-

reflection, and behavioral intervention [357]. As an example, I am developing

a game called Stress Fighter that incorporates biofeedback passively collected

through off-the-shelf wearables in order to deliver stress-relief interventions.

Gameplay dynamics are based on the classic arcade game Street Fighter, with

attributes of the opponent boss character corresponding to the player’s sensed

stress levels that day. Encouraging full body movement while experiencing

the intervention, Stress Fighter continues to capture and incorporate real-time

biofeedback during gameplay since physical exertion can also help tackle stress.

My next step is exploring how games can be used as part of cognitive or psy-

chological therapy, inspired by games like EyeSpy [119], where searching for the

approving face in a crowd of frowns is used to help recondition the mindsets of

people with low self-esteem, or Play Attention [467], where neurofeedback is

used to control game elements and improve ADHD symptoms. In particular, I

am interested in the potential role of video games in nonpharmacological treat-

ments for managing mental illnesses like bipolar disorder, for instance to help

combat periods of depression or act as a safe outlet during manic episodes.

228

6.4 Concluding Remarks

The rapid evolution and dissemination of personal technologies have created

unprecedented opportunities for enhancing health on a broad scale. This dis-

sertation contributes to a growing area of HCI research aimed at realizing this

potential by developing systems that capture personal data, use this data to as-

sess health, and provide tailored feedback that empowers self-management.

Specifically, I have shown the value of a domain-driven approach that draws

on diverse sources (e.g., scientific literature, empirical evidence, practitioner ex-

pertise, and user perspectives) in order to build a solid foundation on which

development decisions can be made. I believe that an approach grounded in

such domain knowledge can better target significant health determinants for as-

sessment and intervention, provide designers with a deep understanding and

empathy for the role of technology in a given health context, and, in my experi-

ence, lead to individuals’ downright enthusiasm in using these technologies.

Integral to this framework is a consideration for the context of application

and the needs of users. In light of the tensions and open challenges I have dis-

cussed, I would like to underscore the value in taking a “do no harm” approach

that carefully considers the promises and pitfalls of technology-based solutions,

which have the potential to positively contribute to health self-management in

meaningful ways when responsibly and compassionately developed.

Having only scratched the surface of opportunities for PHI tools, I see a vast

design space to continue exploring going forward. I hope this dissertation pro-

vides a roadmap that helps researchers traverse a development path toward

technologies that support effective, safe, and empowering experiences.

229

BIBLIOGRAPHY

[1] Khaled Abdel-Kader, Manisha Jhamb, Lee Anne Mandich, JonathanYabes, Robert M Keene, Scott Beach, Daniel J Buysse, and Mark L Unruh.Ecological momentary assessment of fatigue, sleepiness, and exhaustionin ESKD. BMC nephrology, 15(1):1, 2014.

[2] Saeed Abdullah, Mark Matthews, Ellen Frank, Gavin Doherty, Geri Gay,and Tanzeem Choudhury. Automatic detection of social rhythms inbipolar disorder. Journal of the American Medical Informatics Association,23(3):538–543, 2016.

[3] Saeed Abdullah, Mark Matthews, Elizabeth L Murnane, Geri Gay, andTanzeem Choudhury. Towards circadian computing: early to bed andearly to rise makes some of us unhealthy and sleep deprived. In Proceed-ings of the 2014 ACM International Joint Conference on Pervasive and Ubiqui-tous Computing, pages 673–684. ACM, 2014.

[4] Saeed Abdullah, Elizabeth L Murnane, Jean MR Costa, and TanzeemChoudhury. Collective smile: Measuring societal happiness from geolo-cated images. In Proceedings of the 18th ACM Conference on Computer Sup-ported Cooperative Work & Social Computing, pages 361–374. ACM, 2015.

[5] Lorien C Abroms, Nalini Padmanabhan, Lalida Thaweethai, and ToddPhillips. iPhone apps for smoking cessation: a content analysis. Americanjournal of preventive medicine, 40(3):279–285, 2011.

[6] Ana Adan, Simon N Archer, Maria Paz Hidalgo, Lee Di Milia, VincenzoNatale, and Christoph Randler. Circadian typology: A comprehensivereview. Chronobiology international, 29(9):1153–1175, 2012.

[7] Inaki Merino Albaina, Thomas Visser, Charles APG van der Mast, andMartijn H Vastenburg. Flowie: A persuasive virtual coach to motivate el-derly individuals to walk. In 2009 3rd International Conference on PervasiveComputing Technologies for Healthcare, pages 1–7. IEEE, 2009.

[8] Urs Albrecht and Jurgen A Ripperger. Clock genes. In Encyclopedia ofNeuroscience, pages 759–762. Springer, 2009.

[9] Gordon Willard Allport. The use of personal documents in psychologicalscience. Social Science Research Council Bulletin, 1942.

230

[10] Lawrence K Altman. Who goes first?: the story of self-experimentation inmedicine. University of California Press, 1998.

[11] Sanjay Krishna Anbalagan, George Grinstein, and Shweta Purushe. Per-sonal informatics weave your numbers. In Contemporary Computing andInformatics (IC3I), 2014 International Conference on, pages 1211–1214. IEEE,2014.

[12] S Ancoli-Israel, R Cole, C Alessi, M Chambers, W Moorcroft, and C Pol-lak. The role of actigraphy in the study of sleep and circadian rhythms.american academy of sleep medicine review paper. Sleep, 26(3):342–392,2003.

[13] Gerard F Anderson. Chronic care: making the case for ongoing care. RobertWood Johnson Foundation, 2010.

[14] John AE Anderson, Karen L Campbell, Tarek Amer, Cheryl L Grady, andLynn Hasher. Timing is everything: Age differences in the cognitive con-trol network are modulated by time of day. Psychology and aging, 29(3):648,2014.

[15] Jacob Anhøj and Claus Møldrup. Feasibility of collecting diary data fromasthma patients through mobile phones and SMS (short message service):response rate analysis and focus group evaluation from a pilot study. Jour-nal of Medical Internet Research, 6(4):e42, 2004.

[16] Masashi Arakawa, Kazuhiko Taira, Hideki Tanaka, Kenji Yamakawa,Hiroki Toguchi, Hatuko Kadekaru, Yukari Yamamoto, Eiko Uezu, andShuichiro Shirakawa. A survey of junior high school students’ sleep habitand lifestyle in Okinawa. Psychiatry and clinical neurosciences, 55(3):211–212, 2001.

[17] Deanna M Arble, Joseph Bass, Aaron D Laposky, Martha H Vitaterna, andFred W Turek. Circadian timing of food intake contributes to weight gain.Obesity, 17(11):2100–2102, 2009.

[18] Amelia M Arria and Robert L DuPont. Nonmedical prescription stimulantuse among college students: why we need to do something and what weneed to do. Journal of addictive diseases, 29(4):417–426, 2010.

[19] Jurgen Aschoff. Circadian Rhythms in Man. Science, 148:1427–1432, 1965.

231

[20] Jurgen Aschoff. Circadian rhythms: Influences of internal and externalfactors on the period measured in constant conditions. Zeitschrift furTierpsychologie, 49(3):225–249, 1979.

[21] Bon Adriel Aseniero, Sheelagh Carpendale, and Anthony Tang. DeepPersonalization in Tools for Reflection. In CHI 2012 Workshop on PersonalInformatics in Practice: Improving Quality of Life Through Data, 2011.

[22] Najib T Ayas, David P White, JoAnn E Manson, Meir J Stampfer, Frank ESpeizer, Atul Malhotra, and Frank B Hu. A prospective study of sleep du-ration and coronary heart disease in women. Archives of internal medicine,163(2):205–209, 2003.

[23] Kristen MJ Azar, Lenard I Lesser, Brian Y Laing, Janna Stephens, Magi SAurora, Lora E Burke, and Latha P Palaniappan. Mobile applications forweight management: theory-based content analysis. American journal ofpreventive medicine, 45(5):583–589, 2013.

[24] Yin Bai, Bin Xu, Yuanchao Ma, Guodong Sun, and Yu Zhao. Will you havea good sleep tonight?: sleep quality prediction with mobile phone. InProceedings of the 7th International Conference on Body Area Networks, pages124–130. ICST (Institute for Computer Sciences, Social-Informatics andTelecommunications Engineering), 2012.

[25] Brian P Bailey and Shamsi T Iqbal. Understanding changes in mentalworkload during execution of goal-directed tasks and its application forinterruption management. ACM Transactions on Computer-Human Interac-tion (TOCHI), 14(4):21, 2008.

[26] Albert Bandura. Self-efficacy mechanism in human agency. American psy-chologist, 37(2):122, 1982.

[27] Albert Bandura. Social foundations of thought and action: A social cognitivetheory. Prentice-Hall, Inc, 1986.

[28] Albert Bandura. Self-regulation of motivation through anticipatory andself-reactive mechanisms. In Perspectives on motivation: Nebraska sympo-sium on motivation, volume 38, pages 69–164, 1991.

[29] Albert Bandura. Self-efficacy: The exercise of control, 1997.

[30] Larissa K Barber and David C Munz. Consistent-sufficient sleep predicts

232

improvements in self-regulatory performance and psychological strain.Stress and Health, 27(4):314–324, 2011.

[31] Jakob E Bardram, Mads Frost, Karoly Szanto, Maria Faurholt-Jepsen, MajVinberg, and Lars Vedel Kessing. Designing mobile health technology forbipolar disorder: a field trial of the MONARCA system. In Proceedingsof the SIGCHI Conference on Human Factors in Computing Systems, pages2627–2636. ACM, 2013.

[32] Ana Barion and Phyllis C Zee. A clinical approach to circadian rhythmsleep disorders. Sleep Medicine, 8(6):566–577, 2007.

[33] Lisa Feldman Barrett and Daniel J Barrett. An Introduction to Computer-ized Experience Sampling in Psychology. Social Science Computer Review,19(2):175–185, 2001.

[34] Ofir Barzilay and Victor L Brailovsky. On domain knowledge and fea-ture selection using a support vector machine. Pattern Recognition Letters,20(5):475–484, 1999.

[35] Mathias Basner, Daniel Mollicone, and David F Dinges. Validity and sen-sitivity of a brief psychomotor vigilance test (PVT-B) to total and partialsleep deprivation. Acta Astronautica, 69(11):949–959, 2011.

[36] Jared S Bauer, Sunny Consolvo, Benjamin Greenstein, Jonathan Schooler,Eric Wu, Nathaniel F Watson, and Julie Kientz. ShutEye: encouragingawareness of healthy sleep recommendations with a mobile, peripheraldisplay. In Proceedings of the SIGCHI Conference on Human Factors in Com-puting Systems, pages 1401–1410. ACM, 2012.

[37] Roy F Baumeister, Mark Muraven, and Dianne M Tice. Ego depletion:A resource model of volition, self-regulation, and controlled processing.Social cognition, 18(2):130, 2000.

[38] Eric PS Baumer and M Silberman. When the implication is not to design(technology). In Proceedings of the SIGCHI Conference on Human Factors inComputing Systems, pages 2271–2274. ACM, 2011.

[39] Claire Baxter and T Reilly. Influence of time of day on all-out swimming.British Journal of Sports Medicine, 17(2):122–127, 1983.

[40] Victoria Bellotti, Bo Begole, Ed H Chi, Nicolas Ducheneaut, Ji Fang, Ellen

233

Isaacs, Tracy King, Mark W Newman, Kurt Partridge, Bob Price, PaulRasmussen, Michael Roberts, Diane J Schiano, and Alan Walendowski.Activity-based serendipitous recommendations with the Magitti mobileleisure guide. In Proceedings of the SIGCHI Conference on Human Factors inComputing Systems, pages 1157–1166. ACM, 2008.

[41] Dror Ben-Zeev, Kristin E Davis, Susan Kaiser, Izabela Krzsos, andRobert E Drake. Mobile technologies among people with serious men-tal illness: opportunities for future services. Administration and Policy inMental Health and Mental Health Services Research, 40(4):340–343, 2013.

[42] Francesco Benedetti. Antidepressant chronotherapeutics for bipolar de-pression. Dialogues Clin Neurosci, 14(4):401–11, 2012.

[43] Francesco Benedetti, Sara Dallaspezia, Cristina Colombo, Adele Pirovano,Elena Marino, and Enrico Smeraldi. A length polymorphism in the circa-dian clock gene Per3 influences age at onset of bipolar disorder. Neuro-science Letters, 445(2):184–187, 2008.

[44] Frank Bentley and Konrad Tollmar. The power of mobile notifications toincrease wellbeing logging behavior. In Proceedings of the SIGCHI Confer-ence on Human Factors in Computing Systems, pages 1095–1098. ACM, 2013.

[45] Frank Bentley, Konrad Tollmar, Peter Stephenson, Laura Levy, BrianJones, Scott Robertson, Ed Price, Richard Catrambone, and Jeff Wilson.Health Mashups: Presenting statistical patterns between wellbeing dataand context in natural language to promote behavior change. ACM Trans-actions on Computer-Human Interaction (TOCHI), 20(5):30, 2013.

[46] Shlomo Berkovsky, Jill Freyne, and Harri Oinas-Kukkonen. Influencingindividually: fusing personalization and persuasion. ACM Transactionson Interactive Intelligent Systems (TiiS), 2(2):9, 2012.

[47] CH Bjornsson. Lasbarhet (readability) liber, 1968.

[48] Irene V Blair, John F Steiner, and Edward P Havranek. Unconscious (im-plicit) bias and health disparities: where do we go from here. Permanente,15(2):71–78, 2011.

[49] Katharina Blatter and Christian Cajochen. Circadian rhythms in cogni-tive performance: methodological constraints, protocols, theoretical un-derpinnings. Physiology & Behavior, 90(2):196–208, 2007.

234

[50] Jan Blom. Personalization: A taxonomy. In Proceedings of the SIGCHI Con-ference on Human Factors in Computing Systems, pages 313–314. ACM, 2000.

[51] Jan O Blom and Andrew F Monk. Theory of personalization of ap-pearance: why users personalize their pcs and mobile phones. Human-Computer Interaction, 18(3):193–228, 2003.

[52] David E Bloom, Elizabeth Cafiero, Eva Jane-Llopis, Shafika Abrahams-Gessel, Lakshmi Reddy Bloom, Sana Fathima, Andrea B Feigl, TomGaziano, Ali Hamandi, Mona Mowafi, et al. The global economic burdenof noncommunicable diseases. Technical report, Program on the GlobalDemography of Aging, 2012.

[53] Ian Bogost. Persuasive games: exploitationware. Gamasutra, 3, 2011.

[54] Matthias Bohmer, Brent Hecht, Johannes Schoning, Antonio Kruger, andGernot Bauer. Falling asleep with Angry Birds, Facebook and Kindle: alarge scale study on mobile application usage. In Proceedings of the 13thInternational Conference on Human Computer Interaction with Mobile Devicesand Services, pages 47–56. ACM, 2011.

[55] Niall Bolger, Angelina Davis, and Eshkol Rafaeli. Diary methods: Cap-turing life as it is lived. Annual Review of Psychology, 54(1):579–616, 2003.

[56] Michael H Bonnet. Sleep deprivation. Principles and Practice of SleepMedicine, 2:50–67, 2000.

[57] Michael H Bonnet and Donna L Arand. We are chronically sleep deprived.Sleep, 18:908–911, 1995.

[58] Richard R Bootzin and Mindy Engle-Friedman. The assessment of insom-nia. Behavioral Assessment, 3(2):107–126, 1981.

[59] Alexander A Borbely. A two process model of sleep regulation. Humanneurobiology, 1982.

[60] Mohamed Boubekri, Ivy N Cheung, Kathryn J Reid, Nai-Wen Kuo, Chia-Hui Wang, and Phyllis C Zee. Impact of Windows and Daylight Exposureon Overall Health and Sleep Quality of Office Workers-A Case-ControlPilot Study 2. Disclosure, 19:20, 2014.

235

[61] Richard E Boyatzis. Transforming qualitative information: Thematic analysisand code development. Sage, 1998.

[62] Edward W Boyer, Rich Fletcher, Richard J Fay, David Smelson, DouglasZiedonis, and Rosalind W Picard. Preliminary efforts directed towardthe detection of craving of illicit substances: the iheal project. Journal ofMedical Toxicology, 8(1):5–9, 2012.

[63] MM Bradley and PJ Lang. Affective Norms for English Words (ANEW):Technical manual and affective ratings. Gainesville, FL: Center for Researchin Psychophysiology, University of Florida, 1999.

[64] Braun Research Center. Trends in Consumer Mobility Report, 2015.

[65] Naomi Breslau, Thomas Roth, Leon Rosenthal, and Patricia Andreski.Sleep disturbance and psychiatric disorders: a longitudinal epidemiolog-ical study of young adults. Biological psychiatry, 39(6):411–418, 1996.

[66] Barry Brown, Moira McGregor, and Donald McMillan. 100 days of iphoneuse: understanding the details of mobile device use. In Proceedings ofthe 16th international conference on Human-computer interaction with mobiledevices & services, pages 223–232. ACM, 2014.

[67] Scott M Brown, Eugene Santos Jr, and Sheila B Banks. Utility theory-baseduser models for intelligent interface agents. In Conference of the CanadianSociety for Computational Studies of Intelligence, pages 378–392. Springer,1998.

[68] Geir Scott Brunborg, Rune Aune Mentzoni, Helge Molde, Helga Myrseth,Knut Joachim Mr Skouverøe, Bjørn Bjorvatn, and Stle Pallesen. The rela-tionship between media use in the bedroom, sleep habits and symptomsof insomnia. Journal of sleep research, 20(4):569–575, 2011.

[69] Hronn Brynjarsdottir, Maria Hkansson, James Pierce, Eric Baumer, CarlDiSalvo, and Phoebe Sengers. Sustainably unpersuaded: how persuasionnarrows our vision of sustainability. In Proceedings of the SIGCHI Confer-ence on Human Factors in Computing Systems, pages 947–956. ACM, 2012.

[70] Fiona C Bull, CL Holt, MW Kreuter, EM Clark, and D Scharff. Under-standing the effects of printed health education materials: which featureslead to which outcomes? Journal of Health Communication, 6(3):265–280,2001.

236

[71] Fiona C Bull, Matthew W Kreuter, and Darcell P Scharff. Effects of tai-lored, personalized and general health messages on physical activity. Pa-tient education and counseling, 36(2):181–192, 1999.

[72] Helen J Burgess, Katherine M Sharkey, and Charmane I Eastman. Brightlight, dark and melatonin can promote circadian adaptation in night shiftworkers. Sleep Medicine Reviews, 6(5):407–420, 2002.

[73] Matthias Burisch. Test length and validity revisited. European Journal ofPersonality, 11(4):303–315, 1997.

[74] Nancy J Burke, Galen Joseph, Rena J Pasick, and Judith C Barker. The-orizing social context: Rethinking behavioral theory. Health Education &Behavior, 36(5 suppl):55S–70S, 2009.

[75] Marc Busch, Johann Schrammel, and Manfred Tscheligi. Personalized per-suasive technology–development and validation of scales for measuringpersuadability. In Persuasive Technology, pages 33–38. Springer, 2013.

[76] William F Bynum. Science and the Practice of Medicine in the NineteenthCentury. Cambridge University Press, 1994.

[77] John T Cacioppo and Richard E Petty. The need for cognition. Journal ofPersonality and Social Psychology, 42(1):116–131, 1982.

[78] R Sherlock Campbell and James W Pennebaker. The secret life of pro-nouns flexibility in writing style and physical health. Psychological science,14(1):60–65, 2003.

[79] Julie Carrier and Timothy H Monk. Circadian Rhythms of Performance:New Trends. Chronobiology International, 17(6):719–732, 2000.

[80] Erin A Carroll, Mary Czerwinski, Asta Roseway, Ashish Kapoor, PaulJohns, Kael Rowan, and MC Schraefel. Food and mood: Just-in-time sup-port for emotional eating. In Humaine Association Conference on AffectiveComputing and Intelligent Interaction (ACII), pages 252–257. IEEE, 2013.

[81] Mary A Carskadon, Christine Acebo, and Oskar G Jenni. Regulationof adolescent sleep: implications for behavior. Annals of the New YorkAcademy of Sciences, 1021(1):276–291, 2004.

[82] Centers for Disease Control and Prevention. The power of prevention:

237

Chronic disease — the public health challenge of the 21st century. USDepartment of Health and Human Services, 2009.

[83] Tarani Chandola, Jane E Ferrie, Aleksander Perski, Tasnime Akbaraly, andMichael G Marmot. The effect of short sleep duration on coronary heartdisease risk is greatest among those with sleep disturbance: a prospectivestudy from the Whitehall II cohort. Sleep, 33(6):739–744, 2010.

[84] Anne-Marie Chang, Daniel Aeschbach, Jeanne F Duffy, and Charles ACzeisler. Evening use of light-emitting eReaders negatively affects sleep,circadian timing, and next-morning alertness. Proceedings of the NationalAcademy of Sciences, 112(4):1232–1237, 2015.

[85] Samir Chatterjee, Jongbok Byun, Akshay Pottathil, Miles N Moore,Kaushik Dutta, and Harry Qi Xie. Persuasive sensing: a novel in-homemonitoring technology to assist elderly adult diabetic patients. In Interna-tional Conference on Persuasive Technology, pages 31–42. Springer, 2012.

[86] Chih-Yen Chen, Yung-Hsiang Chen, Chun-Fu Lin, Chun-Jen Weng, andHung-Chun Chien. A review of ubiquitous mobile sensing based onsmartphones. International Journal of Automation and Smart Technology,4(1):13–19, 2014.

[87] Gilad Chen, Stanley M Gully, and Dov Eden. Validation of a new generalself-efficacy scale. Organizational Research Methods, 4(1):62–83, 2001.

[88] Mei-Yen Chen, Edward K Wang, and Yi-Jong Jeng. Adequate sleep amongadolescents is positively associated with health status and health-relatedbehaviors. BMC Public Health, 6(1):1, 2006.

[89] Zhenyu Chen, Mu Lin, Fanglin Chen, Nicholas D Lane, Giuseppe Car-done, Rui Wang, Tianxing Li, Yiqiang Chen, Tanzeem Choudhury, andAndrew T Campbell. Unobtrusive sleep monitoring using smartphones.In International Conference on Pervasive Computing Technologies for Healthcareand Workshops, pages 145–152. IEEE, 2013.

[90] James F Childress. Dying patients: who’s in control. Law, Medicine, &Health Care, 17:227, 1989.

[91] Kwangwook Cho. Chronic ‘jet lag’ produces temporal lobe atrophy andspatial cognitive deficits. Nature Neuroscience, 4(6):567–568, 2001.

238

[92] Eun Kyoung Choe, Sunny Consolvo, Nathaniel F Watson, and Julie AKientz. Opportunities for computing technologies to support healthysleep behaviors. In Proceedings of the SIGCHI Conference on Human Factorsin Computing Systems, pages 3053–3062. ACM, 2011.

[93] Eun Kyoung Choe, Bongshin Lee, Matthew Kay, Wanda Pratt, and Julie AKientz. SleepTight: low-burden, self-monitoring technology for captur-ing and reflecting on sleep behaviors. In Proceedings of the 2015 ACM In-ternational Joint Conference on Pervasive and Ubiquitous Computing, pages121–132. ACM, 2015.

[94] Eun Kyoung Choe, Nicole B Lee, Bongshin Lee, Wanda Pratt, and Julie AKientz. Understanding quantified-selfers’ practices in collecting and ex-ploring personal data. In Proceedings of the 32nd annual ACM conference onHuman factors in computing systems, pages 1143–1152. ACM, 2014.

[95] Tanzeem Choudhury, Gaetano Borriello, Sunny Consolvo, Dirk Haehnel,Beverly Harrison, Bruce Hemingway, Jeffrey Hightower, Karl Koscher,Anthony LaMarca, and James A Landay. The mobile sensing platform: Anembedded activity recognition system. IEEE Pervasive Computing, 7(2):32–41, 2008.

[96] Clayton M Christensen, Jerome H Grossman, and Jason Hwang. The inno-vator’s prescription: A disruptive solution for health care. McGraw-Hill,2009.

[97] Chia-Fang Chung, Kristin Dew, Allison Cole, Jasmine Zia, James Fogarty,Julie A Kientz, and Sean A Munson. Boundary negotiating artifacts in per-sonal informatics: Patient-provider collaboration with patient-generateddata. In Proceedings of the 19th ACM Conference on Computer-Supported Co-operative Work & Social Computing, pages 770–786. ACM, 2016.

[98] Ruth C Clark and Richard E Mayer. E-learning and the science of instruction:Proven guidelines for consumers and designers of multimedia learning. JohnWiley & Sons, 2016.

[99] James Clawson, Jessica A Pater, Andrew D Miller, Elizabeth D Mynatt,and Lena Mamykina. No longer wearing: investigating the abandonmentof personal health-tracking technologies on craigslist. In Proceedings of the2015 ACM International Joint Conference on Pervasive and Ubiquitous Com-puting, pages 647–658. ACM, 2015.

[100] Sheldon Cohen, William J Doyle, Cuneyt M Alper, Denise Janicki-Deverts,

239

and Ronald B Turner. Sleep habits and susceptibility to the common cold.Archives of internal medicine, 169(1):62–67, 2009.

[101] Sheldon Cohen and Sarah D Pressman. Positive affect and health. CurrentDirections in Psychological Science, 15(3):122–125, 2006.

[102] Nathalie Colineau and Cecile Paris. Motivating reflection about healthwithin the family: the use of goal setting and tailored feedback. UserModeling and User-Adapted Interaction, 21(4-5):341–376, 2011.

[103] Harvey R Colten and Bruce M Altevogt. Sleep disorders and sleep depriva-tion: an unmet public health problem. National Academies Press, 2006.

[104] Kay H Connelly, Anne M Faber, Yvonne Rogers, Katie A Siek, and TammyToscos. Mobile applications that empower people to monitor their per-sonal health. e & i Elektrotechnik und Informationstechnik, 123(4):124–128,2006.

[105] Sunny Consolvo. Designing and evaluating a persuasive technology to encour-age lifestyle behavior change. ProQuest, 2008.

[106] Sunny Consolvo, Katherine Everitt, Ian Smith, and James A Landay. De-sign requirements for technologies that encourage physical activity. InProceedings of the SIGCHI conference on Human Factors in computing systems,pages 457–466. ACM, 2006.

[107] Sunny Consolvo, David W McDonald, and James A Landay. Theory-driven design strategies for technologies that support behavior changein everyday life. In CHI, pages 405–414, 2009.

[108] Sunny Consolvo, David W McDonald, Tammy Toscos, Mike Y Chen, JonFroehlich, Beverly Harrison, Predrag Klasnja, Anthony LaMarca, LouisLeGrand, Ryan Libby, Ian Smith, and James A Landay. Activity sensingin the wild: a field trial of UbiFit garden. In Proceedings of the SIGCHIConference on Human Factors in Computing Systems, pages 1797–1806. ACM,2008.

[109] Angel Correa, Tania Lara, and Juan Antonio Madrid. Influence of circa-dian typology and time of day on temporal preparation. Timing & TimePerception, 1(2):217–238, 2013.

[110] Dan Cosley, Dan Frankowski, Loren Terveen, and John Riedl. SuggestBot:

240

using intelligent task routing to help people find work in wikipedia. InProceedings of the 12th international conference on Intelligent user interfaces,pages 32–41. ACM, 2007.

[111] Paul T Costa and Robert R McCrae. Revised Neo Personality Inventory(NEO-PI-R) and NEO Five-Factor Inventory (NEO-FFI). Psychological As-sessment Resources, 1992.

[112] Steven S Coughlin. Recall bias in epidemiologic studies. Journal of clinicalepidemiology, 43(1):87–91, 1990.

[113] Stephanie J Crowley, Christine Acebo, and Mary A Carskadon. Sleep,circadian rhythms, and delayed phase in adolescence. Sleep Medicine,8(6):602–612, 2007.

[114] Mihaly Csikszentmihalyi and Reed Larson. Validity and reliability of theexperience-sampling method. In Flow and the foundations of positive psy-chology, pages 35–54. Springer, 2014.

[115] Giuseppe Curcio, Michele Ferrara, and Luigi De Gennaro. Sleep loss,learning capacity and academic performance. Sleep Medicine Reviews,10(5):323–337, 2006.

[116] Charles A Czeisler, Jeanne F Duffy, Theresa L Shanahan, Emery N Brown,Jude F Mitchell, David W Rimmer, Joseph M Ronda, Edward J Silva,James S Allan, Jonathan S Emens, Derk-Jan Dijk, and Richard E Kronauer.Stability, precision, and near-24-hour period of the human circadian pace-maker. Science, 284(5423):2177–2181, 1999.

[117] Serge Daan and Martha Merrow. External time–internal time. Journal ofbiological rhythms, 17(2):107–109, 2002.

[118] James M Dabbs, Rebecca Strong, and Rhonda Milun. Exploring themind of testosterone: A beeper study. Journal of Research in Personality,31(4):577–587, 1997.

[119] Stephane D Dandeneau and Mark W Baldwin. The inhibition of sociallyrejecting information among people with high versus low self-esteem:The role of attentional bias and the effects of bias reduction training. Jour-nal of Social and Clinical Psychology, 23(4):584–603, 2004.

[120] Munmun De Choudhury. Opportunities of social media in health and

241

well-being. XRDS: Crossroads, The ACM Magazine for Students, 21(2):23–27, 2014.

[121] Munmun De Choudhury, Scott Counts, and Eric Horvitz. Major lifechanges and behavioral markers in social media: case of childbirth. InProceedings of the 2013 conference on Computer supported cooperative work,pages 1431–1442. ACM, 2013.

[122] Munmun De Choudhury, Scott Counts, Eric J Horvitz, and Aaron Hoff.Characterizing and predicting postpartum depression from shared Face-book data. In Proceedings of the 17th ACM conference on Computer SupportedCooperative Work & Social Computing, pages 626–638. ACM, 2014.

[123] Munmun De Choudhury, Michael Gamon, Scott Counts, and EricHorvitz. Predicting depression via social media. In ICWSM, page 2, 2013.

[124] Rodrigo De Oliveira and Nuria Oliver. TripleBeat: enhancing exerciseperformance with persuasion. In Proceedings of the 10th international con-ference on Human computer interaction with mobile devices and services, pages255–264. ACM, 2008.

[125] Vincenzo Della Mea. What is e-health: the death of telemedicine? Journalof Medical Internet Research, 3(2):e22, 2001.

[126] Catherine Delmas, Christine Vandamme, and Donna Spalding Andreolle.Science and Empire in the Nineteenth Century: A Journey of Imperial Conquestand Scientific Progress. Cambridge Scholars Publishing, 2010.

[127] William C Dement and Christopher Vaughan. The promise of sleep: A pi-oneer in sleep medicine explores the vital connection between health, happiness,and a good night’s sleep. Dell Publishing Co, 1999.

[128] Anind K Dey. Understanding and using context. Personal and ubiquitouscomputing, 5(1):4–7, 2001.

[129] Carlo C DiClemente, James O Prochaska, Scott K Fairhurst, Wayne FVelicer, Mary M Velasquez, and Joseph S Rossi. The process of smokingcessation: an analysis of precontemplation, contemplation, and prepara-tion stages of change. Journal of consulting and clinical psychology, 59(2):295,1991.

[130] Nancy L Digdon and Andrew J Howell. College students who have an

242

eveningness preference report lower self-control and greater procrastina-tion. Chronobiology international, 25(6):1029–1046, 2008.

[131] John DiNardo. Natural experiments and quasi-natural experiments. InMicroeconometrics, pages 139–153. Springer, 2010.

[132] David F Dinges. An overview of sleepiness and accidents. Journal of sleepresearch, 4(s2):4–14, 1995.

[133] David F Dinges, Frances Pack, Katherine Williams, Kelly A Gillen, John WPowell, Geoffrey E Ott, Caitlin Aptowicz, and Allan I Pack. Cumulativesleepiness, mood disturbance and psychomotor vigilance performancedecrements during a week of sleep restricted to 4-5 hours per night. Sleep:Journal of Sleep Research & Sleep Medicine, 1997.

[134] Trinh Minh Tri Do, Jan Blom, and Daniel Gatica-Perez. Smartphone usagein the wild: a large-scale analysis of applications and context. In Pro-ceedings of the 13th international conference on multimodal interfaces, pages353–360. ACM, 2011.

[135] Martin Dodge and Rob Kitchin. ‘outlines of a world coming into exis-tence’: pervasive computing and the ethics of forgetting. Environment andplanning B: planning and design, 34(3):431–445, 2007.

[136] Gavin Doherty, David Coyle, and John Sharry. Engagement with onlinemental health interventions: an exploratory clinical study of a treatmentfor depression. In Proceedings of the SIGCHI Conference on Human Factorsin Computing Systems, pages 1421–1430. ACM, 2012.

[137] Angelika Dohr, Jeff Engler, Frank Bentley, and Richard Whalley. Glubal-loon: an unobtrusive and educational way to better understand one’s di-abetes. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing,pages 665–666. ACM, 2012.

[138] SM Doran, HPA Van Dongen, and David F Dinges. Sustained atten-tion performance during sleep deprivation: evidence of state instability.Archives italiennes de biologie, 139(3):253–267, 2001.

[139] Jillian Dorrian, Nicole Lamond, Alexandra L Holmes, Helen J Burgess,Gregory D Roach, Adam Fletcher, and Drew Dawson. The ability toself-monitor performance during a week of simulated night shifts. Sleep,26(7):871–877, 2003.

243

[140] Barry Drust, Jim Waterhouse, Greg Atkinson, Ben Edwards, and TomReilly. Circadian rhythms in sports performance—an update. Chrono-biology international, 22(1):21–44, 2005.

[141] Jeanne F Duffy, Sean W Cain, Anne-Marie Chang, Andrew JK Phillips,Mirjam Y Munch, Claude Gronfier, James K Wyatt, Derk-Jan Dijk, Ken-neth P Wright, and Charles A Czeisler. Sex difference in the near-24-hourintrinsic period of the human circadian timing system. Proceedings of theNational Academy of Sciences, 108(Supplement 3):15602–15608, 2011.

[142] M Duggan, NB Ellison, C Lampe, A Lenhart, and M Madden. Social me-dia update 2014. Pew Research Center, 2015.

[143] Phillipa J Easterbrook, Ramana Gopalan, JA Berlin, and David RMatthews. Publication bias in clinical research. The Lancet, 337(8746):867–872, 1991.

[144] Charmane I Eastman, Erin K Hoese, Shawn D Youngstedt, and LiwenLiu. Phase-shifting human circadian rhythms with exercise during thenight shift. Physiology & Behavior, 58(6):1287–1291, 1995.

[145] John D Eastwood, Alexandra Frischen, Mark J Fenske, and Daniel Smilek.The unengaged mind defining boredom in terms of attention. Perspectiveson Psychological Science, 7(5):482–495, 2012.

[146] Cindy L Ehlers, Ellen Frank, and David J Kupfer. Social zeitgebers andbiological rhythms: a unified approach to understanding the etiology ofdepression. Archives of general psychiatry, 45(10):948–952, 1988.

[147] Norman S Endler and James DA Parker. Interactionism revisited: Reflec-tions on the continuing crisis in the personality area. European Journal ofpersonality, 6(3):177–198, 1992.

[148] Daniel Epstein, Felicia Cordeiro, Elizabeth Bales, James Fogarty, and SeanMunson. Taming data complexity in lifelogs: exploring visual cuts ofpersonal informatics data. In Proceedings of the 2014 conference on Designinginteractive systems, pages 667–676. ACM, 2014.

[149] Daniel A Epstein, An Ping, James Fogarty, and Sean A Munson. A livedinformatics model of personal informatics. In Proceedings of the 2015 ACMInternational Joint Conference on Pervasive and Ubiquitous Computing, pages731–742. ACM, 2015.

244

[150] Thomas Erickson, Ming Li, Younghun Kim, Ajay Deshpande, SambitSahu, Tian Chao, Piyawadee Sukaviriya, and Milind Naphade. Thedubuque electricity portal: evaluation of a city-scale residential electric-ity consumption feedback system. In Proceedings of the SIGCHI Conferenceon Human Factors in Computing Systems, pages 1203–1212. ACM, 2013.

[151] Jaume Escofet and Salvador Bara. Reducing the circadian input from self-luminous devices using hardware filters and software applications. Light-ing Research and Technology, page 1477153515621946, 2015.

[152] Deborah Estrin. Small data, where n= me. Communications of the ACM,57(4):32–34, 2014.

[153] Benedict Evans. Mobile is Eating the World. 2014.

[154] Brian S Everitt and Christopher Palmer. Encyclopaedic companion to medicalstatistics. John Wiley & Sons, 2011.

[155] Stewart A Factor, Terence McAlarney, Juan R Sanchez-Ramos, andWilliam J Weiner. Sleep disorders and sleep effect in Parkinson’s disease.Movement disorders, 5(4):280–285, 1990.

[156] Haiyan Fan and Marshall Scott Poole. What is personalization? perspec-tives on the design and implementation of personalization in informationsystems. Journal of Organizational Computing and Electronic Commerce, 16(3-4):179–202, 2006.

[157] Michele Ferrara and Luigi De Gennaro. How much sleep do we need?Sleep Medicine Reviews, 5(2):155–179, 2001.

[158] Denzil Ferreira, Jorge Goncalves, Vassilis Kostakos, Louise Barkhuus, andAnind K Dey. Contextual experience sampling of mobile applicationmicro-usage. In Proceedings of the 16th international conference on Human-computer interaction with mobile devices & services, pages 91–100. ACM,2014.

[159] Denzil Ferreira, Vassilis Kostakos, and Anind K Dey. AWARE: mobilecontext instrumentation framework. Frontiers in ICT, 2:6, 2015.

[160] Dorothee Fischer, Celine Vetter, Christoph Oberlinner, Sven Wegener, andTill Roenneberg. A unique, fast-forwards rotating schedule with 12-h long

245

shifts prevents chronic sleep debt. Chronobiology international, 33(1):98–107, 2016.

[161] Brianna S Fjeldsoe, Alison L Marshall, and Yvette D Miller. Behaviorchange interventions delivered by mobile telephone short-message ser-vice. American journal of preventive medicine, 36(2):165–173, 2009.

[162] BJ Fogg. Computers as persuasive social actors. 2003.

[163] BJ Fogg. Mobile persuasion: 20 perspectives on the future of behavior change.Mobile Persuasion, 2007.

[164] Brian J Fogg. Persuasive technology: using computers to change what wethink and do. Ubiquity, 2002(December):5, 2002.

[165] Brian J Fogg. A behavior model for persuasive design. In Proceedings of the4th international Conference on Persuasive Technology, page 40. ACM, 2009.

[166] Daniel Foley, Sonia Ancoli-Israel, Patricia Britz, and James Walsh. Sleepdisturbances and chronic disease in older adults: results of the 2003 Na-tional Sleep Foundation Sleep in America Survey. Journal of psychosomaticresearch, 56(5):497–502, 2004.

[167] Daniel E Ford and Douglas B Kamerow. Epidemiologic study of sleepdisturbances and psychiatric disorders: an opportunity for prevention?Jama, 262(11):1479–1484, 1989.

[168] Susannah Fox and Maeve Duggan. Health online 2013, 2013.

[169] Susannah Fox and Maeve Duggan. Tracking for health. Pew Research Cen-ter’s Internet & American Life Project, 2013.

[170] Ellen Frank, David J Kupfer, Michael E Thase, Alan G Mallinger, Holly ASwartz, Andrea M Fagiolini, Victoria Grochocinski, Patricia Houck, JohnScott, Wesley Thompson, and Timothy H Monk. Two-year outcomes forinterpersonal and social rhythm therapy in individuals with bipolar I dis-order. Archives of general psychiatry, 62(9):996–1004, 2005.

[171] Ellen Frank, Holly A Swartz, and David J Kupfer. Interpersonal and so-cial rhythm therapy: managing the chaos of bipolar disorder. Biologicalpsychiatry, 48(6):593–604, 2000.

246

[172] Tanya Freijy and Emily J Kothe. Dissonance-based interventions forhealth behaviour change: A systematic review. British journal of healthpsychology, 18(2):310–337, 2013.

[173] Jon Froehlich, Leah Findlater, and James Landay. The design of eco-feedback technology. In Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems, pages 1999–2008. ACM, 2010.

[174] Jon Froehlich, Leah Findlater, Marilyn Ostergren, Solai Ramanathan, JoshPeterson, Inness Wragg, Eric Larson, Fabia Fu, Mazhengmin Bai, ShwetakPatel, and James A Landay. The design and evaluation of prototype eco-feedback displays for fixture-level water usage data. In Proceedings of theSIGCHI conference on human factors in computing systems, pages 2367–2376.ACM, 2012.

[175] Mads Frost, Afsaneh Doryab, Maria Faurholt-Jepsen, Lars Vedel Kessing,and Jakob E Bardram. Supporting disease insight through data analysis:refinements of the monarca self-assessment system. In Proceedings of the2013 ACM international joint conference on Pervasive and ubiquitous comput-ing, pages 133–142. ACM, 2013.

[176] Nicole B Gabler, Naihua Duan, Sunita Vohra, and Richard L Kravitz. N-of-1 trials in the medical literature: a systematic review. Medical care,49(8):761–768, 2011.

[177] Roland Gasser, Dominique Brodbeck, Markus Degen, Jurg Luthiger,Remo Wyss, and Serge Reichlin. Persuasiveness of a mobile lifestylecoaching application using social facilitation. In International Conferenceon Persuasive Technology, pages 27–38. Springer, 2006.

[178] Moniek Geerdink, Thijs J Walbeek, Domien GM Beersma, Vanja Hommes,and Marijke CM Gordijn. Short blue light pulses (30 min) in the morningsupport a sleep-advancing protocol in a home setting. Journal of BiologicalRhythms, 31(5):483–497, 2016.

[179] Jim Gemmell, Aleks Aris, and Roger Lueder. Telling Stories withMylifebits. In IEEE International Conference on Multimedia and Expo, pages6–9, 2005.

[180] JC Gillin. Are sleep disturbances risk factors for anxiety, depressive andaddictive disorders? Acta Psychiatrica Scandinavica, 98(s393):39–43, 1998.

247

[181] Russell E Glasgow, Sheana S Bull, John D Piette, and John F Steiner. In-teractive behavior change technology: a partial solution to the competingdemands of primary care. American journal of preventive medicine, 27(2):80–87, 2004.

[182] Russell E Glasgow, Michael G Goldstein, Judith K Ockene, and Nicolaas PPronk. Translating what we have learned into practice: principles andhypotheses for interventions addressing multiple behaviors in primarycare. American journal of preventive medicine, 27(2):88–101, 2004.

[183] N Goel, HPA Van Dongen, and DF Dinges. Circadian rhythms in sleepi-ness, alertness, and performance. Principles and practice of sleep medicine,5:445–55, 2011.

[184] Namni Goel, Mathias Basner, Hengyi Rao, and David F Dinges. Circadianrhythms, sleep deprivation, and human performance. Progress in molecularbiology and translational science, 119:155, 2013.

[185] Erving Goffman. The presentation of self in everyday life. Harmondsworth,1978.

[186] Jennifer Golbeck, Cristina Robles, and Karen Turner. Predicting person-ality with social media. In CHI’11 extended abstracts on human factors incomputing systems, pages 253–262. ACM, 2011.

[187] Scott A Golder and Michael W Macy. Diurnal and seasonal moodvary with work, sleep, and daylength across diverse cultures. Science,333(6051):1878–1881, 2011.

[188] Peter M Gollwitzer and Paschal Sheeran. Implementation intentions andgoal achievement: A meta-analysis of effects and processes. Advances inexperimental social psychology, 38:69–119, 2006.

[189] Robert Gonzalez. The relationship between bipolar disorder and biologi-cal rhythms. The Journal of clinical psychiatry, 75(4):323–331, 2014.

[190] Frederick K Goodwin and Kay Redfield Jamison. Manic-depressive ill-ness: bipolar disorders and recurrent depression, volume 1. Oxford UniversityPress, 2007.

[191] P Gorwood. Anxiety disorders and circadian rhythms. Medicographica,34:289–294, 2012.

248

[192] Daniel J Gottlieb, Naresh M Punjabi, Ann B Newman, Helaine E Resnick,Susan Redline, Carol M Baldwin, and F Javier Nieto. Association of sleeptime with diabetes mellitus and impaired glucose tolerance. Archives ofinternal medicine, 165(8):863–867, 2005.

[193] Michael Gradisar and Michelle A Short. Sleep hygiene and environment: roleof technology. PhD thesis, Oxford University Press, 2013.

[194] Michael Gradisar, Amy R Wolfson, Allison G Harvey, Lauren Hale, Rus-sell Rosenberg, and Charles A Czeisler. The sleep and technology use ofAmericans: findings from the National Sleep Foundation’s 2011 Sleep inAmerica poll. Journal of Clinical Sleep Medicine, 9(12):1291–1299, 2013.

[195] Erin Griffiths, T Scott Saponas, and AJ Brush. Health chair: implicitlysensing heart and respiratory rate. In Proceedings of the 2014 ACM Inter-national Joint Conference on Pervasive and Ubiquitous Computing, pages 661–671. ACM, 2014.

[196] William Groves. Using domain knowledge to systematically guide featureselection. In Proceedings of the Twenty-Third international joint conference onArtificial Intelligence, pages 3215–3216. Citeseer, 2013.

[197] Franz Halberg, Germaine Cornelissen, George Katinas, Elena V Syutkina,Robert B Sothern, Rina Zaslavskaya, Francine Halberg, Yoshihiko Watan-abe, Othild Schwartzkopff, Kuniaki Otsuka, et al. Transdisciplinary uni-fying implications of circadian findings in the 1950s. Journal of CircadianRhythms, 1(1):1, 2003.

[198] Sajanee Halko and Julie A Kientz. Personality and persuasive technol-ogy: an exploratory study on health-promoting mobile applications. InInternational Conference on Persuasive Technology, pages 150–161. Springer,2010.

[199] Alina Hang, Alexander De Luca, Jonas Hartmann, and Heinrich Huss-mann. Oh app, where art thou?: on app launching habits of smartphoneusers. In Proceedings of the 15th international conference on Human-computerinteraction with mobile devices and services, pages 392–395. ACM, 2013.

[200] Tian Hao, Guoliang Xing, and Gang Zhou. isleep: unobtrusive sleep qual-ity monitoring using smartphones. In Proceedings of the 11th ACM Confer-ence on Embedded Networked Sensor Systems, page 4. ACM, 2013.

249

[201] Dr Peter Harrop, Raghu Das, and G Chansin. Wearable technology 2014–2024. Report, IDTechEx, 2014.

[202] Allison G Harvey. Sleep and circadian rhythms in bipolar disorder: seek-ing synchrony, harmony, and regulation. American Journal of Psychiatry,2008.

[203] Lynn Hasher, David Goldstein, and Cynthia P May. It’s about time: Cir-cadian rhythms, memory, & aging. 2005.

[204] Jennifer A Haythornthwaite, Lynette A Menefee, Leslie J Heinberg, andMichael R Clark. Pain coping strategies predict perceived control overpain. Pain, 77(1):33–39, 1998.

[205] Helen Ai He, Saul Greenberg, and Elaine M Huang. One size does notfit all: applying the transtheoretical model to energy feedback technol-ogy design. In Proceedings of the SIGCHI Conference on Human Factors inComputing Systems, pages 927–936. ACM, 2010.

[206] Eric B Hekler, Predrag Klasnja, Jon E Froehlich, and Matthew P Buman.Mind the theoretical gap: interpreting, using, and developing behavioraltheory in hci research. In CHI, pages 3307–3316, 2013.

[207] Helene Hembrooke and Geri Gay. The laptop and the lecture: The effectsof multitasking in learning environments. Journal of computing in highereducation, 15(1):46–64, 2003.

[208] Shelley D Hershner and Ronald D Chervin. Causes and consequences ofsleepiness among college students. Nat Sci Sleep, 6:73–84, 2014.

[209] Pauline Heslop, George Davey Smith, Chris Metcalfe, John Macleod, andCarole Hart. Sleep duration and mortality: the effect of short or long sleepduration on cardiovascular and all-cause mortality in working men andwomen. Sleep Medicine, 3(4):305–314, 2002.

[210] Max Hirshkowitz, Kaitlyn Whiton, Steven M Albert, Cathy Alessi,Oliviero Bruni, Lydia DonCarlos, Nancy Hazen, John Herman, Eliot SKatz, Leila Kheirandish-Gozal, David Neubauer, Anne E O’Donnell,Maurice Ohayon, John Peever, Robert Rawding, Ramesh C Sachdeva,Belinda Setters, Vitiello Michael V, J Catesby Ware, and Paula J Adams-Hillard. National Sleep Foundation’s sleep time duration recommenda-tions: methodology and results summary. Sleep Health, 1(1):40–43, 2015.

250

[211] John L Holland and Joan E Holland. Vocational indecision: More evidenceand speculation. Journal of Counseling Psychology, 24(5):404, 1977.

[212] James A Horne and Olov Ostberg. Individual differences in human circa-dian rhythms. Biological Psychology, 5(3):179–190, 1977.

[213] Jim A Horne and Olov Ostberg. A self-assessment questionnaire to de-termine morningness-eveningness in human circadian rhythms. Interna-tional journal of chronobiology, 4(2):97–110, 1975.

[214] House-n Research Group. A Just-In-Time Intervention Technology to En-courage Healthy Behavior Using Operant Conditioning. In Proceedings ofthe 2004 ACM Conference on Ubiquitous Computing. ACM, 2004.

[215] Meng-Hsiang Hsu, Teresa L Ju, Chia-Hui Yen, and Chun-Ming Chang.Knowledge sharing behavior in virtual communities: The relationship be-tween trust, self-efficacy, and outcome expectations. International journalof human-computer studies, 65(2):153–169, 2007.

[216] Dandan Huang, Melanie Tory, Bon Adriel Aseniero, Lyn Bartram, ScottBateman, Sheelagh Carpendale, Anthony Tang, and Robert Woodbury.Personal visualization and personal visual analytics. IEEE transactions onvisualization and computer graphics, 21(3):420–433, 2015.

[217] Ke Huang, Xiang Ding, Jing Xu, Guanling Chen, and Wei Ding. Moni-toring sleep and detecting irregular nights through unconstrained smart-phone sensing. In 12th IEEE International Conference on Ubiquitous Intelli-gence and Computing. IEEE, 2015.

[218] Steven R Hursh, Daniel P Redmond, Michael L Johnson, David R Thorne,Gregory Belenky, Thomas J Balkin, William F Storm, James C Miller, andDouglas R Eddy. Fatigue models for applied research in warfighting. Avi-ation, space, and environmental medicine, 75(Supplement 1):A44–A53, 2004.

[219] Margarita P Hurtado, Elaine K Swift, and Janet M Corrigan. Envisioningthe national health care quality report. National Academies Press, 2001.

[220] Clayton J Hutto, Sarita Yardi, and Eric Gilbert. A longitudinal study offollow predictors on Twitter. In Proceedings of the sigchi conference on humanfactors in computing systems, pages 821–830. ACM, 2013.

[221] Maree L Inder, Marie T Crowe, Stephanie Moor, Suzanne E Luty, Janet D

251

Carter, and Peter R Joyce. “I actually don’t know who I am”: The impactof bipolar disorder on the development of self. Psychiatry, 71(2):123–133,2008.

[222] Stephen S Intille. Designing a home of the future. IEEE pervasive comput-ing, 1(2):76–82, 2002.

[223] Sheena S Iyengar and Mark R Lepper. When choice is demotivating: Canone desire too much of a good thing? Journal of personality and social psy-chology, 79(6):995, 2000.

[224] Sue Jamison-Powell, Conor Linehan, Laura Daley, Andrew Garbett, andShaun Lawson. I can’t get no sleep: discussing# insomnia on Twitter. InProceedings of the SIGCHI Conference on Human Factors in Computing Sys-tems, pages 1501–1510. ACM, 2012.

[225] Kasthuri Jayarajah, Meera Radhakrishnan, Steven Hoi, and Archan Misra.Candy crushing your sleep. In Adjunct Proceedings of the 2015 ACM Inter-national Joint Conference on Pervasive and Ubiquitous Computing and Proceed-ings of the 2015 ACM International Symposium on Wearable Computers, pages753–762. ACM, 2015.

[226] Oliver P John, Eileen M Donahue, and Robert L Kentle. The Big FiveInventory: Versions 4a and 54, institute of personality and social research.University of California, Berkeley, CA, 1991.

[227] Jeffrey V Johnson and Jane Lipscomb. Long working hours, occupationalhealth and the changing nature of work organization. American journal ofindustrial medicine, 49(11):921–929, 2006.

[228] Simon L Jones, Denzil Ferreira, Simo Hosio, Jorge Goncalves, and VassilisKostakos. Revisitation analysis of smartphone app use. In Proceedingsof the 2015 ACM International Joint Conference on Pervasive and UbiquitousComputing, pages 1197–1208. ACM, 2015.

[229] Nam-Seok Joo and Bom-Taeck Kim. Mobile phone short message servicemessaging for behaviour modification in a community-based weight con-trol programme in korea. Journal of Telemedicine and Telecare, 13(8):416–420,2007.

[230] Myriam Juda, Celine Vetter, and Till Roenneberg. The Munich Chrono-type Questionnaire for Shift-Workers (MCTQShift). Journal of biologicalrhythms, 28(2):130–140, 2013.

252

[231] Richard Kadison and Theresa Foy DiGeronimo. College of the overwhelmed:The campus mental health crisis and what to do about it. Jossey-Bass, 2004.

[232] Biren B Kamdar, Katherine A Kaplan, Eric J Kezirian, and William C De-ment. The impact of extended sleep on daytime alertness, vigilance, andmood. Sleep Medicine, 5(5):441–448, 2004.

[233] Warren A Kaplan. Can the ubiquitous power of mobile phones be usedto improve health outcomes in developing countries? Globalization andhealth, 2(1):1, 2006.

[234] Maurits Kaptein. Personalized persuasion in ambient intelligence. Journalof Ambient Intelligence and Smart Environments, 4(3):279–280, 2012.

[235] Maurits Kaptein and Dean Eckles. Heterogeneity in the effects of onlinepersuasion. Journal of Interactive Marketing, 26(3):176–188, 2012.

[236] Maurits Kaptein, Joyca Lacroix, and Privender Saini. Individual differ-ences in persuadability in the health promotion domain. In InternationalConference on Persuasive Technology, pages 94–105. Springer, 2010.

[237] Yamini Karanam, Leslie Filko, Lindsay Kaser, Hanan Alotaibi, ElhamMakhsoom, and Stephen Voida. Motivational affordances and personalitytypes in personal informatics. In Proceedings of the 2014 ACM InternationalJoint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication,pages 79–82. ACM, 2014.

[238] Ilia N Karatsoreos. Links between circadian rhythms and psychiatric dis-ease. Frontiers in behavioral neuroscience, 8:162, 2014.

[239] Ilia N Karatsoreos, Sarah Bhagat, Erik B Bloss, John H Morrison, andBruce S McEwen. Disruption of circadian clocks has ramifications formetabolism, brain, and behavior. Proceedings of the national Academy ofSciences, 108(4):1657–1662, 2011.

[240] Ravi Karkar, James Fogarty, Julie A Kientz, Sean A Munson, RogerVilardaga, and Jasmine Zia. Opportunities and challenges for self-experimentation in self-tracking. In Adjunct Proceedings of the 2015 ACMInternational Joint Conference on Pervasive and Ubiquitous Computing and Pro-ceedings of the 2015 ACM International Symposium on Wearable Computers,pages 991–996. ACM, 2015.

253

[241] Ravi Karkar, Jasmine Zia, Roger Vilardaga, Sonali R Mishra, JamesFogarty, Sean A Munson, and Julie A Kientz. A framework for self-experimentation in personalized health. Journal of the American MedicalInformatics Association, page ocv150, 2015.

[242] Matthew Kay, Eun Kyoung Choe, Jesse Shepherd, Benjamin Greenstein,Nathaniel Watson, Sunny Consolvo, and Julie A Kientz. Lullaby: a cap-ture & access system for understanding the sleep environment. In Proceed-ings of the 2012 ACM Conference on Ubiquitous Computing, pages 226–234.ACM, 2012.

[243] Matthew Kay, Kyle Rector, Sunny Consolvo, Ben Greenstein, Jacob OWobbrock, Nathaniel F Watson, and Julie A Kientz. PVT-touch: adapt-ing a reaction time test for touchscreen devices. In 2013 7th InternationalConference on Pervasive Computing Technologies for Healthcare and Workshops,pages 248–251. IEEE, 2013.

[244] Misha Kay, Jonathan Santos, and Marina Takane. mHealth: New hori-zons for health through mobile technologies. World Health Organization,64(7):66–71, 2011.

[245] Naz Kaya and Helen H Epps. Relationship between color and emotion:A study of college students. College student journal, 38(3):396, 2004.

[246] Alan E Kazdin. Reactive self-monitoring: the effects of response desirabil-ity, goal setting, and feedback. Journal of consulting and clinical psychology,42(5):704, 1974.

[247] Diane Kelly and Jaime Teevan. Implicit feedback for inferring user pref-erence: a bibliography. ACM SIGIR Forum, 37(2):18–28, 2003.

[248] Ronald C Kessler and Paul E Greenberg. The economic burden of anxi-ety and stress disorders. Neuropsychopharmacology: The fifth generation ofprogress, 67:982–92, 2002.

[249] Katherine M Keyes, Julie Maslowsky, Ava Hamilton, and John Schulen-berg. The great sleep recession: Changes in sleep duration among USadolescents, 1991–2012. Pediatrics, 135(3):460–468, 2015.

[250] Vera Khovanskaya, Eric PS Baumer, Dan Cosley, Stephen Voida, and GeriGay. Everybody knows what you’re doing: a critical design approach topersonal informatics. In Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems, pages 3403–3412. ACM, 2013.

254

[251] Sunyoung Kim, Julie A Kientz, Shwetak N Patel, and Gregory D Abowd.Are you sleeping?: sharing portrayed sleeping status within a social net-work. In Proceedings of the 2008 ACM conference on Computer SupportedCooperative Work, pages 619–628. ACM, 2008.

[252] Abby C King. Behavioral medicine in the 21st century: transforming “theRoad Less Traveled” into the “American Way of Life”. Annals of behavioralmedicine, 47(1):71–78, 2014.

[253] Abby C King, Deborah Toobert, David Ahn, Ken Resnicow, Mace Co-day, Deborah Riebe, Carol E Garber, Shannon Hurtz, Jessica Morton, andJames F Sallis. Perceived environments as physical activity correlates andmoderators of intervention in five studies. American Journal of Health Pro-motion, 21(1):24–35, 2006.

[254] Martha Anne Kitzrow. The mental health needs of today’s college stu-dents: Challenges and recommendations. NASPA journal, 41(1):167–181,2003.

[255] Predrag Klasnja and Wanda Pratt. Healthcare in the pocket: mapping thespace of mobile-phone health interventions. Journal of biomedical informat-ics, 45(1):184–198, 2012.

[256] Urban Knutsson, Jovanna Dahlgren, Claude Marcus, Sten Rosberg,Mikael BronnegaŁrd, Pontus Stierna, and Kerstin Albertsson-Wikland.Circadian cortisol rhythms in healthy boys and girls: Relationship withage, growth, body composition, and pubertal development 1. The Journalof Clinical Endocrinology & Metabolism, 82(2):536–540, 1997.

[257] PT Ko, Julie A Kientz, Eun Kyoung Choe, Matthew Kay, Carol A Landis,and Nathaniel F Watson. Consumer sleep technologies: a review of thelandscape. Journal of Clinical Sleep Medicine, 11:1455–1461, 2015.

[258] Annika Kohl, Winfried Rief, and Julia Anna Glombiewski. Acceptance,cognitive restructuring, and distraction as coping strategies for acute pain.The journal of pain, 14(3):305–315, 2013.

[259] Artie Konrad, Victoria Bellotti, Nicole Crenshaw, Simon Tucker, Les Nel-son, Honglu Du, Peter Pirolli, and Steve Whittaker. Finding the adaptivesweet spot: Balancing compliance and achievement in automated stressreduction. In Proceedings of the 33rd Annual ACM Conference on HumanFactors in Computing Systems, pages 3829–3838. ACM, 2015.

255

[260] Judy Kopp. Self-monitoring: A literature review of research and practice.24(4):8–20, 1988.

[261] Ilkka Korhonen, Juha Parkka, and Mark Van Gils. Health monitoring inthe home of the future. IEEE Engineering in medicine and biology magazine,22(3):66–73, 2003.

[262] William J Korotitsch and Rosemery O Nelson-Gray. An overview of self-monitoring research in assessment and treatment. Psychological Assess-ment, 11(4):415, 1999.

[263] Raghavendra Kotikalapudi, S Chellappan, F Montgomery, D Wunsch,and K Lutzen. Associating depressive symptoms in college students withinternet usage using real internet data. IEEE Technology and Society Maga-zine, 31(4):73–80, 2012.

[264] Denis Kotkov, Shuaiqiang Wang, and Jari Veijalainen. A survey ofserendipity in recommender systems. Knowledge-Based Systems, 111:180–192, 2016.

[265] Efthymios Kouloumpis, Theresa Wilson, and Johanna D Moore. Twittersentiment analysis: The good the bad and the omg! Icwsm, 11:538–541,2011.

[266] Leon Kreitzman and Russell Foster. The Rhythms Of Life: The BiologicalClocks That Control the Daily Lives of Every Living Thing. Profile books,2011.

[267] Ondrej Krejcar, Jakub Jirka, and Dalibor Janckulik. Use of mobile phonesas intelligent sensors for sound input analysis and sleep state detection.Sensors, 11(6):6037–6055, 2011.

[268] Gerald P Krueger. Sustained work, fatigue, sleep loss and performance:A review of the issues. Work & Stress, 3(2):129–141, 1989.

[269] Nicole Lamond and Drew Dawson. Quantifying the performance impair-ment associated with fatigue. Journal of sleep research, 8(4):255–262, 1999.

[270] Nicole Lamond, Sarah M Jay, Jillian Dorrian, Sally A Ferguson, Gregory DRoach, and Drew Dawson. The sensitivity of a palm-based psychomotorvigilance task to severe sleep loss. Behavior research methods, 40(1):347–352,2008.

256

[271] Elaine Waddington Lamont, Daniel Legault-Coutu, Nicolas Cermakian,and Diane B Boivin. The role of circadian clock genes in mental disorders.Dialogues in clinical neuroscience, 9(3):333, 2007.

[272] Klodiana Lanaj, Russell E Johnson, and Christopher M Barnes. Begin-ning the workday yet already depleted? consequences of late-night smart-phone use and sleep. Organizational Behavior and Human Decision Processes,124(1):11–23, 2014.

[273] Nicholas D Lane, Mu Lin, Mashfiqui Mohammod, Xiaochao Yang, HongLu, Giuseppe Cardone, Shahid Ali, Afsaneh Doryab, Ethan Berke, An-drew T Campbell, and Tanzeem Choudhury. BeWell: Sensing sleep, phys-ical activities and social interactions to promote wellbeing. Mobile Net-works and Applications, 19(3):345–359, 2014.

[274] Nicholas D Lane, Emiliano Miluzzo, Hong Lu, Daniel Peebles, TanzeemChoudhury, and Andrew T Campbell. A survey of mobile phone sensing.Communications Magazine, IEEE, 48(9):140–150, 2010.

[275] Nicholas D Lane, Mashfiqui Mohammod, Mu Lin, Xiaochao Yang, HongLu, Shahid Ali, Afsaneh Doryab, Ethan Berke, Tanzeem Choudhury, andAndrew Campbell. BeWell: A smartphone application to monitor, modeland promote wellbeing. In 5th international ICST conference on pervasivecomputing technologies for healthcare, pages 23–26, 2011.

[276] Matthias Laschke, Marc Hassenzahl, Sarah Diefenbach, and MariusTippkamper. With a little help from a friend: a shower calendar to savewater. In CHI’11 Extended Abstracts on Human Factors in Computing Sys-tems, pages 633–646. ACM, 2011.

[277] Shaun Lawson, Sue Jamison-Powell, Andrew Garbett, Conor Linehan, Er-ica Kucharczyk, Sanne Verbaan, Duncan A Rowland, and Kevin Morgan.Validating a mobile phone application for the everyday, unobtrusive, ob-jective measurement of sleep. In Proceedings of the SIGCHI Conference onHuman Factors in Computing Systems, pages 2497–2506. ACM, 2013.

[278] Min Kyung Lee, Junsung Kim, Jodi Forlizzi, and Sara Kiesler. Personaliza-tion revisited: a reflective approach helps people better personalize healthservices and motivates them to increase physical activity. In Proceedingsof the 2015 ACM International Joint Conference on Pervasive and UbiquitousComputing, pages 743–754. ACM, 2015.

[279] Uichin Lee, Joonwon Lee, Minsam Ko, Changhun Lee, Yuhwan Kim,

257

Subin Yang, Koji Yatani, Gahgene Gweon, Kyong-Mee Chung, and June-hwa Song. Hooked on smartphones: an exploratory study on smartphoneoveruse among college students. In Proceedings of the 32nd annual ACMconference on Human factors in computing systems, pages 2327–2336. ACM,2014.

[280] Damien Leger. The cost of sleep-related accidents: a report for the Na-tional Commission on Sleep Disorders Research. Sleep, 17(1):84–93, 1994.

[281] A Lemhart. Cell phones and American adults. Pew Research Center, 2010.

[282] Jessica Levenson and Ellen Frank. Sleep and circadian rhythm abnormal-ities in the pathophysiology of bipolar disorder. In Behavioral Neurobiologyof Bipolar Disorder and its Treatment, pages 247–262. Springer, 2010.

[283] Ian Li. Designing personal informatics applications and tools that facili-tate monitoring of behaviors. Proceedings of UIST’09, 2009.

[284] Ian Li, Anind Dey, and Jodi Forlizzi. A stage-based model of personalinformatics systems. In CHI, pages 557–566, 2010.

[285] Walter Libby. An introduction to the history of science. Houghton Mifflin,1917.

[286] Elizabeth O Lillie, Bradley Patay, Joel Diamant, Brian Issell, Eric J Topol,and Nicholas J Schork. The n-of-1 clinical trial: the ultimate strategy forindividualizing medicine? Personalized medicine, 8(2):161–173, 2011.

[287] Julian Lim and David F Dinges. Sleep deprivation and vigilant attention.Annals of the New York Academy of Sciences, 1129(1):305–322, 2008.

[288] Julian Lim and David F Dinges. A meta-analysis of the impact ofshort-term sleep deprivation on cognitive variables. Psychological bulletin,136(3):375, 2010.

[289] Vivien KG Lim. The it way of loafing on the job: cyberloafing, neutralizingand organizational justice. Journal of Organizational Behavior, 23(5):675–694, 2002.

[290] James J Lin, Lena Mamykina, Silvia Lindtner, Gregory Delajoux, andHenry B Strub. Fish’n’steps: Encouraging physical activity with an inter-

258

active computer game. In International Conference on Ubiquitous Computing,pages 261–278. Springer, 2006.

[291] Mark D Litt, Ned L Cooney, and Priscilla Morse. Ecological momentaryassessment (EMA) with treated alcoholics: methodological problems andpotential solutions. Health Psychology, 17(1):48, 1998.

[292] Haiming Liu, Dawei Song, and Paul Mulholland. Exploration of applyinga theory-based user classification model to inform personalised content-based image retrieval system design. In Proceedings of HCI Korea, pages61–68. Hanbit Media, Inc., 2016.

[293] G Livingston, B Blizard, and A Mann. Does sleep disturbance predictdepression in elderly people? A study in inner London. Br J Gen Pract,43(376):445–448, 1993.

[294] Edwin A Locke and Gary P Latham. A theory of goal setting & task perfor-mance. Prentice-Hall, Inc, 1990.

[295] Edwin A Locke and Gary P Latham. Building a practically useful theoryof goal setting and task motivation: A 35-year odyssey. American psychol-ogist, 57(9):705, 2002.

[296] Edwin A Locke, Karyll N Shaw, Lise M Saari, and Gary P Latham. Goalsetting and task performance: 1969–1980. Psychological bulletin, 90(1):125,1981.

[297] Steven W Lockley, Debra J Skene, and Josephine Arendt. Compari-son between subjective and actigraphic measurement of sleep and sleeprhythms. Journal of sleep research, 8(3):175–183, 1999.

[298] Alan D Lopez and Christopher JL Murray. The Global Burden of Disease: AComprehensive Assessment of Mortality and Disability from Diseases, Injuries,and Risk Factors in 1990 and Projected to 2020. Harvard School of Public ofPublic Health, 1996.

[299] Shane A Lowe and Gearoid OLaighin. Monitoring human health be-haviour in one’s living environment: A technological review. Medical en-gineering & physics, 36(2):147–168, 2014.

[300] Deborah Lupton. Critical perspectives on digital health technologies. So-ciology compass, 8(12):1344–1359, 2014.

259

[301] Deborah Lupton and Annemarie Jutel. ‘It’s like having a physician inyour pocket!’A critical analysis of self-diagnosis smartphone apps. SocialScience & Medicine, 133:128–135, 2015.

[302] Elizabeth J Lyons, Zakkoyya H Lewis, Brian G Mayrsohn, and Jennifer LRowland. Behavior change techniques implemented in electronic lifestyleactivity monitors: a systematic content analysis. Journal of medical Internetresearch, 16(8):e192, 2014.

[303] Haley MacLeod, Anthony Tang, and Sheelagh Carpendale. Towards per-sonal informatics tools for chronic illness management. 2013.

[304] Liliana Maggioni and Patricia A Alexander. Knowledge domains and do-main learning. International encyclopedia of education, pages 255–264, 2010.

[305] Andres Magnusson and Diane Boivin. Seasonal affective disorder: Anoverview. Chronobiology international, 20(2):189–207, 2003.

[306] Evelyn Mai and Daniel J Buysse. Insomnia: prevalence, impact, patho-genesis, differential diagnosis, and evaluation. Sleep Medicine Clinics,3(2):167–174, 2008.

[307] Francois Mairesse, Marilyn A Walker, Matthias R Mehl, and Roger KMoore. Using linguistic cues for the automatic recognition of personalityin conversation and text. Journal of artificial intelligence research, 30:457–500,2007.

[308] Pierre Maquet. The role of sleep in learning and memory. science,294(5544):1048–1052, 2001.

[309] Gloria Mark, Shamsi T Iqbal, Mary Czerwinski, and Paul Johns. BoredMondays and focused afternoons: the rhythm of attention and online ac-tivity in the workplace. In Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems, pages 3025–3034. ACM, 2014.

[310] Gloria Mark, Yiran Wang, and Melissa Niiya. Stress and multitasking ineveryday college life: an empirical study of online activity. In Proceedingsof the SIGCHI Conference on Human Factors in Computing Systems, pages41–50. ACM, 2014.

[311] Aleksandar Matic, Martin Pielot, and Nuria Oliver. Boredom-computerinteraction: boredom proneness and the use of smartphone. In Proceedings

260

of the 2015 ACM International Joint Conference on Pervasive and UbiquitousComputing, pages 837–841. ACM, 2015.

[312] Mark Matthews, Saeed Abdullah, Elizabeth Murnane, Stephen Voida,Tanzeem Choudhury, Geri Gay, and Ellen Frank. Development and eval-uation of a smartphone-based measure of social rhythms for bipolar dis-order. Assessment, 23(4):472–483, 2016.

[313] Mark Matthews, Elizabeth Murnane, and Jaime Snyder. Quantifying theChangeable Self: The role of self-tracking in coming to terms with andmanaging bipolar disorder. Human–Computer Interaction, 2017.

[314] Mark Matthews, Elizabeth Murnane, Jaime Snyder, Shion Guha, PamaraChang, Gavin Doherty, and Geri Gay. The Double-Edged Sword: A MixedMethods Study Of The Interplay Between Bipolar Disorder and Technol-ogy Use. Computers in Human Behavior, 2017.

[315] Tara Matthews, Tye Rattenbury, and Scott Carter. Defining, designing,and evaluating peripheral displays: An analysis using activity theory.Human–Computer Interaction, 22(1-2):221–261, 2007.

[316] E Mattila, I Korhonen, R Lappalainen, A Ahtinen, L Hopsu, and T Leino.Nuadu concept for personal management of lifestyle related health risks.In 2008 30th Annual International Conference of the IEEE Engineering inMedicine and Biology Society, pages 5846–5850. IEEE, 2008.

[317] Elina Mattila, Ilkka Korhonen, Jukka H Salminen, Aino Ahtinen, EsaKoskinen, Antti Sarela, Juha Parkka, and Raimo Lappalainen. Empower-ing citizens for well-being and chronic disease management with wellnessdiary. IEEE Transactions on Information Technology in Biomedicine, 14(2):456–463, 2010.

[318] Elina Mattila, Juha Parkka, Marion Hermersdorf, Jussi Kaasinen, JanneVainio, Kai Samposalo, Juho Merilahti, Juha Kolari, Minna Kulju,Raimo Lappalainen, and Ilkka Korhonen. Mobile diary for wellnessmanagement—results on usage and usability in two user studies. IEEETransactions on information technology in biomedicine, 12(4):501–512, 2008.

[319] Sharon F Matusik and Amy E Mickel. Embracing or embattled by con-verged mobile devices? users’ experiences with a contemporary connec-tivity technology. Human Relations, 64(8):1001–1030, 2011.

261

[320] Viktor Mayer-Schonberger and Kenneth Cukier. Big data: A revolution thatwill transform how we live, work, and think. Houghton Mifflin Harcourt,2013.

[321] Brian James McInnis, Elizabeth Lindley Murnane, Dmitry Epstein, DanCosley, and Gilly Leshed. One and Done: Factors affecting one-time con-tributors to ad-hoc online communities. In Proceedings of the 19th ACMConference on Computer-Supported Cooperative Work & Social Computing,pages 609–623. ACM, 2016.

[322] Kalliopi Megari. Quality of life in chronic disease patients. Health Psychol-ogy Research, 1(3), 2013.

[323] Jerome S Menet and Michael Rosbash. When brain clocks lose track oftime: cause or consequence of neuropsychiatric disorders. Current opinionin neurobiology, 21(6):849–857, 2011.

[324] George Mensah. Global and Domestic Health Priorities: Spotlight onChronic Disease. National Business Group on Health Webinar, 2006.

[325] Jerrold S Meyer and Melinda A Novak. Minireview: hair cortisol: a novelbiomarker of hypothalamic-pituitary-adrenocortical activity. Endocrinol-ogy, 153(9):4120–4127, 2012.

[326] Jack Michael. Positive and negative reinforcement, a distinction that is nolonger necessary; or a better way to talk about bad things. Behaviorism,3(1):33–44, 1975.

[327] Susan Michie and Charles Abraham. Interventions to change health be-haviours: evidence-based or evidence-inspired? Psychology & Health,19(1):29–49, 2004.

[328] Susan Michie, Stefanie Ashford, Falko F Sniehotta, Stephan U Dom-browski, Alex Bishop, and David P French. A refined taxonomy of be-haviour change techniques to help people change their physical activityand healthy eating behaviours: the CALO-RE taxonomy. Psychology &Health, 26(11):1479–1498, 2011.

[329] Susan Michie, Marie Johnston, Jill Francis, Wendy Hardeman, and Mar-tin Eccles. From theory to intervention: mapping theoretically derivedbehavioural determinants to behaviour change techniques. Applied psy-chology, 57(4):660–680, 2008.

262

[330] Mario Miguel, Valeria Clarisse de Oliveira, Danyella Pereira, and MarioPedrazzoli. Detecting Chronotype Differences Associated to Latitude: AComparison Between Horne–Ostberg and Munich Chronotype Question-naires. Annals of human biology, 41(2):107–110, 2014.

[331] David J Miklowitz, Michael W Otto, Ellen Frank, Noreen A Reilly-Harrington, Jane N Kogan, Gary S Sachs, Michael E Thase, Joseph R Cal-abrese, Lauren B Marangell, Michael J Ostacher, et al. Intensive psychoso-cial intervention enhances functioning in patients with bipolar depres-sion: results from a 9-month randomized controlled trial. American Journalof Psychiatry, 2007.

[332] Emiliano Miluzzo, Nicholas D Lane, Kristof Fodor, Ronald Peterson,Hong Lu, Mirco Musolesi, Shane B Eisenman, Xiao Zheng, and Andrew TCampbell. Sensing meets mobile social networks: the design, implemen-tation and evaluation of the cenceme application. In Proceedings of the 6thACM conference on Embedded network sensor systems, pages 337–350. ACM,2008.

[333] Jun-Ki Min, Afsaneh Doryab, Jason Wiese, Shahriyar Amini, John Zim-merman, and Jason I Hong. Toss’n’turn: smartphone as sleep and sleepquality detector. In Proceedings of the SIGCHI Conference on Human Factorsin Computing Systems, pages 477–486. ACM, 2014.

[334] Jared D Minkel, Siobhan Banks, Oo Htaik, Marisa C Moreta, Christo-pher W Jones, Eleanor L McGlinchey, Norah S Simpson, and David FDinges. Sleep deprivation and stressors: evidence for elevated nega-tive affect in response to mild stressors when sleep deprived. Emotion,12(5):1015, 2012.

[335] Dana K Mirick and Scott Davis. Melatonin as a biomarker of circadiandysregulation. Cancer Epidemiology Biomarkers & Prevention, 17(12):3306–3313, 2008.

[336] Toshihiko Miyazaki, Satoko Hashimoto, Satoru Masubuchi, Sato Honma,and Ken-Ichi Honma. Phase-advance shifts of human circadian pace-maker are accelerated by daytime physical exercise. American Journal ofPhysiology-Regulatory, Integrative and Comparative Physiology, 281(1):R197–R205, 2001.

[337] David C Mohr, Michelle Nicole Burns, Stephen M Schueller, GregoryClarke, and Michael Klinkman. Behavioral intervention technologies: ev-

263

idence review and recommendations for future research in mental health.General hospital psychiatry, 35(4):332–338, 2013.

[338] David C Mohr, Joyce Ho, Jenna Duffecy, Kelly G Baron, Kenneth ALehman, Ling Jin, and Douglas Reifler. Perceived barriers to psychologi-cal treatments and their relationship to depression. Journal of clinical psy-chology, 66(4):394–409, 2010.

[339] David C Mohr, Stephen M Schueller, Enid Montague, Michelle NicoleBurns, and Parisa Rashidi. The behavioral intervention technology model:an integrated conceptual and technological framework for ehealth andmhealth interventions. Journal of medical Internet research, 16(6):e146, 2014.

[340] TH Monk, DJ Buysse, KS Kennedy, JM Potts, JM DeGrazia, andJM Miewald. Measuring sleep habits without using a diary: The sleeptiming questionnaire (STQ). Sleep, 26(2):208–12, 2003.

[341] Timothy H Monk. The post-lunch dip in performance. Clinics in sportsmedicine, 24(2):e15–e23, 2005.

[342] Timothy H Monk, Daniel J Buysse, Charles F Reynolds, and David JKupfer. Circadian determinants of the postlunch dip in performance.Chronobiology international, 13(2):123–133, 1996.

[343] Timothy H Monk, Susan R Petrie, Amy J Hayes, and David J Kupfer. Reg-ularity of daily life in relation to personality, age, gender, sleep qualityand circadian rhythms. Journal of Sleep Research, 3(4):196–205, 1994.

[344] Timothy K Monk, Joseph F Flaherty, Ellen Frank, Kathleen Hoskinson,and David J Kupfer. The Social Rhythm Metric: An instrument to quantifythe daily rhythms of life. Journal of Nervous and Mental Disease, 1990.

[345] Daniel S Moran and Liran Mendal. Core temperature measurement.Sports Medicine, 32(14):879–885, 2002.

[346] Margaret Morris and Farzin Guilak. Mobile heart health: project high-light. IEEE Pervasive Computing, 8(2):57–61, 2009.

[347] Sai T Moturu, Inas Khayal, Nadav Aharony, Wei Pan, and Alex SandyPentland. Using social sensing to understand the links between sleep,mood, and sociability. In Privacy, Security, Risk and Trust (PASSAT) and

264

2011 IEEE Third International Conference on Social Computing (SocialCom),pages 208–214. IEEE, 2011.

[348] James E Muller, Paul L Ludmer, Stefan N Willich, Geoffrey H Tofler, GailAylmer, Irene Klangos, and Peter H Stone. Circadian variation in the fre-quency of sudden cardiac death. Circulation, 75(1):131–138, 1987.

[349] Sean A Munson and Sunny Consolvo. Exploring goal-setting, rewards,self-monitoring, and sharing to motivate physical activity. In 2012 6thInternational Conference on Pervasive Computing Technologies for Healthcare(PervasiveHealth) and Workshops, pages 25–32. IEEE, 2012.

[350] Sean A Munson, Erin Krupka, Caroline Richardson, and Paul Resnick.Effects of public commitments and accountability in a technology-supported physical activity intervention. In Proceedings of the 33rd AnnualACM Conference on Human Factors in Computing Systems, pages 1135–1144.ACM, 2015.

[351] Mark Muraven, Roy F Baumeister, and Dianne M Tice. Longitudinalimprovement of self-regulation through practice: Building self-controlstrength through repeated exercise. The Journal of social psychology,139(4):446–457, 1999.

[352] Elizabeth L Murnane, Saeed Abdullah, Mark Matthews, Tanzeem Choud-hury, and Geri Gay. Social (media) jet lag: how usage of social technologycan modulate and reflect circadian rhythms. In Proceedings of the 2015ACM International Joint Conference on Pervasive and Ubiquitous Computing,pages 843–854. ACM, 2015.

[353] Elizabeth L Murnane, Saeed Abdullah, Mark Matthews, Matthew Kay,Julie A Kientz, Tanzeem Choudhury, Geri Gay, and Dan Cosley. Mobilemanifestations of alertness: connecting biological rhythms with patternsof smartphone app use. In Proceedings of the 18th International Conferenceon Human-Computer Interaction with Mobile Devices and Services, pages 465–477. ACM, 2016.

[354] Elizabeth L Murnane, Dan Cosley, Pamara Chang, Shion Guha, EllenFrank, Geri Gay, and Mark Matthews. Self-monitoring practices, atti-tudes, and needs of individuals with bipolar disorder: implications forthe design of technologies to manage mental health. Journal of the Ameri-can Medical Informatics Association, 23(3):477–484, 2016.

[355] Elizabeth L Murnane and Scott Counts. Unraveling abstinence and re-

265

lapse: smoking cessation reflected in social media. In Proceedings of the32nd annual ACM conference on Human factors in computing systems, pages1345–1354. ACM, 2014.

[356] Elizabeth L Murnane, David Huffaker, and Gueorgi Kossinets. MobileHealth Apps: Adoption, Adherence, and Abandonment. In Adjunct Pro-ceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiq-uitous Computing and Proceedings of the 2015 ACM International Symposiumon Wearable Computers, pages 261–264. ACM, 2015.

[357] Elizabeth L Murnane, Mark Matthews, Geri Gay, and Dan Cosley. Playingwith your data: towards personal informatics driven games. In Proceed-ings of the 2016 ACM International Joint Conference on Pervasive and Ubiqui-tous Computing: Adjunct, pages 565–569. ACM, 2016.

[358] Christopher JL Murray and Alan D Lopez. Evidence-based health policy–lessons from the Global Burden of Disease Study. Science, 274(5288):740,1996.

[359] Dawn Nafus. Quantified: Biosensing technologies in everyday life. MIT Press,2016.

[360] Dawn Nafus and Jamie Sherman. Big data, big questions— this one doesnot go up to 11: the quantified self movement as an alternative big datapractice. International journal of communication, 8:11, 2014.

[361] Inbal Nahum-Shani, Shawna N Smith, Ambuj Tewari, Katie Witkiewitz,Linda M Collins, Bonnie Spring, and S Murphy. Just In Time AdaptiveInterventions (JITAIS): An organizing framework for ongoing health be-havior support. Methodology Center technical report 14-126, 2014.

[362] Rajalakshmi Nandakumar, Shyamnath Gollakota, and Nathaniel Watson.Contactless sleep apnea detection on smartphones. In Proceedings of the13th Annual International Conference on Mobile Systems, Applications, andServices, pages 45–57. ACM, 2015.

[363] National Institutes of Health. A short history of the National Institutes ofHealth, 2009.

[364] National Sleep Foundation. Technology use and sleep, 2011.

[365] National Sleep Foundation. 2013 Sleep in America Poll, 2013.

266

[366] National Sleep Foundation. 2014 Sleep Health Index, 2014.

[367] National Transportation Safety Board. Evaluation of US Department ofTransportation Efforts in the 1990s to Address Operator Fatigue. 1999.

[368] Gina Neff and Dawn Nafus. Self-Tracking. MIT Press, 2016.

[369] Hien Nguyen, Judith Masthoff, and Peter Edwards. Modelling a re-ceiver’s position to persuasive arguments. In International Conference onPersuasive Technology, pages 271–282. Springer, 2007.

[370] Scott Nicholson. A user-centered theoretical framework for meaningfulgamification. Games+ Learning+ Society, 8(1):223–230, 2012.

[371] Reto Niedermann, Eva Wyss, Simon Annaheim, Agnes Psikuta, SarahDavey, and Rene Michel Rossi. Prediction of human core body temper-ature using non-invasive measurement methods. International journal ofbiometeorology, 58(1):7–15, 2014.

[372] Linda M Northrop and Paul C Clements. A framework for software prod-uct line practice, Version 5.0. Software Engineering Institute, 2005.

[373] Oded Nov and Ofer Arazy. Personality-targeted design: Theory, experi-mental procedure, and preliminary results. In Proceedings of the 2013 con-ference on Computer Supported Cooperative Work, pages 977–984. ACM, 2013.

[374] Oded Nov, Ofer Arazy, Claudia Lopez, and Peter Brusilovsky. Exploringpersonality-targeted UI design in online social participation systems. InProceedings of the SIGCHI Conference on Human Factors in Computing Sys-tems, pages 361–370. ACM, 2013.

[375] Danny O’Brien. Life hacks: Tech secrets of overprolific alpha geeks. InEmerging Technology Conference, San Diego, CA, pages 02–9, 2004.

[376] Mitsuhiro Ogawa, Toshiyo Tamura, and Tatasuo Togawa. Automatedacquisition system for routine, noninvasive monitoring of physiologicaldata. Telemedicine journal, 4(2):177–185, 1998.

[377] Halszka Oginska and Janusz Pokorski. Fatigue and mood correlates ofsleep length in three age-social groups: School children, students, andemployees. Chronobiology international, 23(6):1317–1328, 2006.

267

[378] Hans Oh, Carlos Rizo, Murray Enkin, and Alejandro Jadad. What isehealth?: a systematic review of published definitions. World Hosp HealthServ, 41(1):32–40, 2005.

[379] Fredrik Ohlin and Carl Magnus Olsson. Intelligent computing in personalinformatics: Key design considerations. In Proceedings of the 20th Interna-tional Conference on Intelligent User Interfaces, pages 263–274. ACM, 2015.

[380] Huseyin Oktay, Brian J Taylor, and David D Jensen. Causal discovery insocial media using quasi-experimental designs. In Proceedings of the FirstWorkshop on Social Media Analytics, pages 1–9. ACM, 2010.

[381] Rita Orji. Exploring the persuasiveness of behavior change support strate-gies and possible gender differences. In Proceedings of the Second Interna-tional Workshop on Behavior Change Support Systems (BCSS2014), Padova,Italy, 2014.

[382] Rita Orji, Regan L Mandryk, Julita Vassileva, and Kathrin M Gerling. Tai-loring persuasive health games to gamer type. In CHI, pages 2467–2476,2013.

[383] Antti Oulasvirta, Tye Rattenbury, Lingyi Ma, and Eeva Raita. Habitsmake smartphone use more pervasive. Personal and Ubiquitous Comput-ing, 16(1):105–114, 2012.

[384] James F Pagel and Seithikurippu R Pandi-Perumal. Primary Care SleepMedicine: A Practical Guide. Springer, 2014.

[385] An Pan, Elizabeth Devore, and Eva S Schernhammer. How shift workand a destabilized circadian system may increase risk for development ofcancer and type 2 diabetes. Circadian Medicine, pages 183–209, 2015.

[386] Katrina Panovich. Habitbot: A Twitter-based Tool to Help Form andMaintain Habits and Goals. In Adjunct proceedings of the 31st annual ACMconference on Human factors in computing systems. ACM, 2013.

[387] James M Parish. Sleep-related problems in common medical conditions.Chest Journal, 135(2):563–572, 2009.

[388] Minsu Park, Chiyoung Cha, and Meeyoung Cha. Depressive moods ofusers portrayed in twitter. In Proceedings of the ACM SIGKDD Workshop onhealthcare informatics (HI-KDD), pages 1–8, 2012.

268

[389] Sanjay R Patel, Najib T Ayas, Mark R Malhotra, David P White, Eva SSchernhammer, Frank E Speizer, Meir J Stampfer, and Frank B Hu. Aprospective study of sleep duration and mortality risk in women. Sleep,27(3):440–444, 2004.

[390] James W Pennebaker, Ryan L Boyd, Kayla Jordan, and Kate Blackburn.The development and psychometric properties of LIWC2015. UT Fac-ulty/Researcher Works, 2015.

[391] James W Pennebaker, Martha E Francis, and Roger J Booth. Linguistic in-quiry and word count: LIWC 2001. Mahway: Lawrence Erlbaum Associates,71:2001, 2001.

[392] Roy H Perlis, Sachiko Miyahara, Lauren B Marangell, Stephen R Wis-niewski, Michael Ostacher, Melissa P DelBello, Charles L Bowden, Gary SSachs, and Andrew A Nierenberg. Long-term implications of early onsetin bipolar disorder: data from the first 1000 participants in the systematictreatment enhancement program for bipolar disorder (STEP-BD). Biologi-cal psychiatry, 55(9):875–881, 2004.

[393] Sean M Phelan, DJ Burgess, MW Yeazel, WL Hellerstedt, JM Griffin, andM Ryn. Impact of weight bias and stigma on quality of care and outcomesfor patients with obesity. obesity reviews, 16(4):319–326, 2015.

[394] Jean M Phillips and Stanley M Gully. Organizational behavior: Tools forsuccess. Cengage Learning, 2011.

[395] Keith G Phillips, Ullrich Bartsch, Andrew P McCarthy, Dale M Edgar,Mark D Tricklebank, Keith A Wafford, and Matt W Jones. Decouplingof sleep-dependent cortical and hippocampal interactions in a neurode-velopmental model of schizophrenia. Neuron, 76(3):526–533, 2012.

[396] Martin Pielot, Rodrigo de Oliveira, Haewoon Kwak, and Nuria Oliver.Didn’t you see my message?: predicting attentiveness to mobile instantmessages. In Proceedings of the 32nd annual ACM conference on Human fac-tors in computing systems, pages 3319–3328. ACM, 2014.

[397] Martin Pielot, Tilman Dingler, Jose San Pedro, and Nuria Oliver. Whenattention is not scarce-detecting boredom from mobile phone usage. InProceedings of the 2015 ACM International Joint Conference on Pervasive andUbiquitous Computing, pages 825–836. ACM, 2015.

269

[398] John D Piette. Interactive behavior change technology to support diabetesself-management where do we stand? Diabetes Care, 30(10):2425–2432,2007.

[399] David T Plante and John W Winkelman. Sleep disturbance in bipolar dis-order: therapeutic implications. American Journal of Psychiatry, 165(7):830–843, 2008.

[400] John P Pollak, Phil Adams, and Geri Gay. PAM: a photographic affectmeter for frequent, in situ measurement of affect. In Proceedings of theSIGCHI conference on Human factors in computing systems, pages 725–734.ACM, 2011.

[401] Jacob Poushter. Smartphone ownership and internet usage continues toclimb in emerging economies. Pew Research Center, 2016.

[402] James O Prochaska, Carlo C DiClemente, and John C Norcross. In searchof how people change: applications to addictive behaviors. American psy-chologist, 47(9):1102, 1992.

[403] James O Prochaska and Wayne F Velicer. The transtheoretical model ofhealth behavior change. American journal of health promotion, 12(1):38–48,1997.

[404] James O Prochaska, Wayne F Velicer, Carlo C DiClemente, and JosephFava. Measuring processes of change: applications to the cessation ofsmoking. Journal of consulting and clinical psychology, 56(4):520, 1988.

[405] Halley Profita, Asta Roseway, and Mary Czerwinski. Lightwear: An ex-ploration in wearable light therapy. In Proceedings of the Ninth InternationalConference on Tangible, Embedded, and Embodied Interaction, pages 321–328.ACM, 2015.

[406] Stephen Purpura, Victoria Schwanda, Kaiton Williams, William Stubler,and Phoebe Sengers. Fit4Life: the design of a persuasive technology pro-moting healthy behavior and ideal weight. In Proceedings of the SIGCHIConference on Human Factors in Computing Systems, pages 423–432. ACM,2011.

[407] Quantified Self Labs. Quantified Self Data Analysis. http://quantifiedself.com/data-analysis/.

270

[408] Mashfiqui Rabbi, Min Hane Aung, Mi Zhang, and Tanzeem Choudhury.MyBehavior: automatic personalized health feedback from user behav-iors and preferences using smartphones. In Proceedings of the 2015 ACMInternational Joint Conference on Pervasive and Ubiquitous Computing, pages707–718. ACM, 2015.

[409] Tauhidur Rahman, Alexander T Adams, Ruth Vinisha Ravichandran,Mi Zhang, Shwetak N Patel, Julie A Kientz, and Tanzeem Choudhury.DoppleSleep: A contactless unobtrusive sleep sensing system using short-range doppler radar. In Proceedings of the 2015 ACM International Joint Con-ference on Pervasive and Ubiquitous Computing, pages 39–50. ACM, 2015.

[410] Reza Rawassizadeh, Blaine A Price, and Marian Petre. Wearables: has theage of smartwatches finally arrived? Communications of the ACM, 58(1):45–47, 2015.

[411] Rajasee S Rege, Jennifer L Allen, Eric P Drewski, and Robert S Molnar.The GroceryMate: eliciting community empathy and transforming it intopurposeful action. In CHI’08 Extended Abstracts on Human Factors in Com-puting Systems, pages 3885–3890. ACM, 2008.

[412] Kathryn J Reid, Kelly Glazer Baron, Brandon Lu, Erik Naylor, Lisa Wolfe,and Phyllis C Zee. Aerobic exercise improves self-reported sleep andquality of life in older adults with insomnia. Sleep Medicine, 11(9):934–940, 2010.

[413] Research2Guidance. mHealth App Developer Economics 2016, 2016.

[414] Sokwoo Rhee, Boo-Ho Yang, Kuowei Chang, and Hamhiko H Asada. Thering sensor: a new ambulatory wearable sensor for twenty-four hour pa-tient monitoring. In Engineering in Medicine and Biology Society, 1998. Pro-ceedings of the 20th Annual International Conference of the IEEE, volume 4,pages 1906–1909. IEEE, 1998.

[415] Victoria J Rideout, Ulla G Foehr, and Donald F Roberts. Generation M2.Media in the lives of 8- to 18-year-olds. Kaiser Foundation Research, 2010.

[416] William T Riley, Daniel E Rivera, Audie A Atienza, Wendy Nilsen, Susan-nah M Allison, and Robin Mermelstein. Health behavior models in theage of mobile interventions: are our theories up to the task? Translationalbehavioral medicine, 1(1):53–71, 2011.

271

[417] Barbara K Rimer and Karen Glanz. Theory at a glance: a guide for healthpromotion practice. 2005.

[418] Gregory D Roach, Drew Dawson, and Nicole Lamond. Can a ShorterPsychomotor Vigilance Task Be Used as a Reasonable Substitute for theTen-Minute Psychomotor Vigilance Task? Chronobiology international,23(6):1379–1387, 2006.

[419] Lisa M Robinson and Sara R Vail. An integrative review of adolescentsmoking cessation using the transtheoretical model of change. Journal ofPediatric Health Care, 26(5):336–345, 2012.

[420] Michael J Roche, Aaron L Pincus, Amanda L Rebar, David E Conroy,and Nilam Ram. Enriching psychological assessment using a person-specific analysis of interpersonal processes in daily life. Assessment, page1073191114540320, 2014.

[421] T Roenneberg and M Merrow. Entrainment of the human circadian clock.In Cold Spring Harbor symposia on quantitative biology, volume 72, pages293–299. Cold Spring Harbor Laboratory Press, 2007.

[422] Till Roenneberg. Internal time: Chronotypes, social jet lag, and why you’re sotired. Harvard Press, 2012.

[423] Till Roenneberg. Chronobiology: the human sleep project. Nature,498(7455):427–428, 2013.

[424] Till Roenneberg, Karla V Allebrandt, Martha Merrow, and Celine Vetter.Social jetlag and obesity. Current Biology, 22(10):939–943, 2012.

[425] Till Roenneberg, Lena K Keller, Dorothee Fischer, Joana L Matera, CelineVetter, and Eva C Winnebeck. Human activity and rest in situ. Methods inenzymology, 552:257–283, 2015.

[426] Till Roenneberg, Tim Kuehnle, Myriam Juda, Thomas Kantermann, KarlaAllebrandt, Marijke Gordijn, and Martha Merrow. Epidemiology of thehuman circadian clock. Sleep Medicine Reviews, 11(6):429–438, 2007.

[427] Till Roenneberg, Tim Kuehnle, Peter P Pramstaller, Jan Ricken, MiriamHavel, Angelika Guth, and Martha Merrow. A marker for the end of ado-lescence. Current Biology, 14(24):R1038–R1039, 2004.

272

[428] Till Roenneberg, Anna Wirz-Justice, and Martha Merrow. Life betweenclocks: daily temporal patterns of human chronotypes. Journal of BiologicalRhythms, 18(1):80–90, 2003.

[429] Stephanie E Roepke and Jeanne F Duffy. Differential impact of chronotypeon weekday and weekend sleep timing and duration. Nature and scienceof sleep, 2010(2):213, 2010.

[430] Mahsan Rofouei, Mike Sinclair, Ray Bittner, Tom Blank, Nick Saw, GeraldDeJean, and Jeff Heffron. A non-invasive wearable neck-cuff system forreal-time sleep monitoring. In 2011 International Conference on Body SensorNetworks, pages 156–161. IEEE, 2011.

[431] John Rooksby, Mattias Rost, Alistair Morrison, and Matthew ChalmersChalmers. Personal tracking as lived informatics. In Proceedings of the32nd annual ACM conference on Human factors in computing systems, pages1163–1172. ACM, 2014.

[432] Mark R Rosekind, Kevin B Gregory, Melissa M Mallis, Summer L Brandt,Brian Seal, and Debra Lerner. The cost of poor sleep: workplace produc-tivity loss and associated costs. Journal of Occupational and EnvironmentalMedicine, 52(1):91–98, 2010.

[433] S Trent Rosenbloom. Person-generated health and wellness data forhealth care. Journal of the American Medical Informatics Association,23(3):438–439, 2016.

[434] Asta Roseway, Yuliya Lutchyn, Paul Johns, Elizabeth Mynatt, and MaryCzerwinski. Biocrystal: An ambient tool for emotion and communica-tion. International Journal of Mobile Human Computer Interaction (IJMHCI),7(3):20–41, 2015.

[435] Michael Royer, Noel H Ballentine, Paul J Eslinger, Kevin Houser, RichardMistrick, Richard Behr, and Kirk Rakos. Light therapy for seniors in longterm care. Journal of the American Medical Directors Association, 13(2):100–102, 2012.

[436] Stephanie Rude, Eva-Maria Gortner, and James Pennebaker. Languageuse of depressed and depression-vulnerable college students. Cognition &Emotion, 18(8):1121–1133, 2004.

[437] Polly Ryan and Diane Ruth Lauver. The efficacy of tailored interventions.Journal of Nursing Scholarship, 34(4):331–337, 2002.

273

[438] Akane Sano. Got Sleep? In Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems. ACM, 2014.

[439] Niilo Saranummi, Donna Spruijt-Metz, Stephen S Intille, Ilkka Korho-nen, Wendy J Nilsen, and Misha Pavel. Moving the science of behavioralchange into the 21st century. IEEE pulse, 4(6):32–33, 2013.

[440] Hillol Sarker, Moushumi Sharmin, Amin Ahsan Ali, Md Mahbubur Rah-man, Rummana Bari, Syed Monowar Hossain, and Santosh Kumar. As-sessing the availability of users to engage in just-in-time intervention inthe natural environment. In Proceedings of the 2014 ACM International JointConference on Pervasive and Ubiquitous Computing, pages 909–920. ACM,2014.

[441] Corina Sas, Steve Whittaker, Steven Dow, Jodi Forlizzi, and John Zimmer-man. Generating implications for design through design research. In Pro-ceedings of the 32nd annual ACM conference on Human factors in computingsystems, pages 1971–1980. ACM, 2014.

[442] Jeffrey L Saver and Mary Kalafut. Combination therapies and the theo-retical limits of evidence-based medicine. Neuroepidemiology, 20(2):57–64,2001.

[443] Bill Schilit, Norman Adams, and Roy Want. Context-aware computingapplications. In Mobile Computing Systems and Applications, 1994. WMCSA1994. First Workshop on, pages 85–90. IEEE, 1994.

[444] A Schmeiser-Rieder, G Kapfhammer, J Bolitschek, B Holzinger, A Skrobal,M Kunze, H Lechner, B Saletu, and J Zeitlhofer. Self reported prevalenceand treatment of sleep disorders in Austria. Journal of epidemiology andcommunity health, 49(6):645, 1995.

[445] Christina Schmidt, Fabienne Collette, Christian Cajochen, and PhilippePeigneux. A time to think: circadian rhythms in human cognition. Cogni-tive neuropsychology, 24(7):755–789, 2007.

[446] Hanna Schneider, Kilian Moser, Andreas Butz, and Florian Alt. Under-standing the Mechanics of Persuasive System Design: A Mixed-MethodTheory-driven Analysis of Freeletics. In Proceedings of the 2016 CHI Con-ference on Human Factors in Computing Systems, pages 309–320. ACM, 2016.

[447] Nicholas J Schork. Personalized medicine: time for one-person trials. Na-ture, 520(7549):609–11, 2015.

274

[448] Analyne M Schroeder and Christopher S Colwell. How to fix a brokenclock. Trends in pharmacological sciences, 34(11):605–619, 2013.

[449] Steven A Schroeder. We can do better—improving the health of the Amer-ican people. New England Journal of Medicine, 357(12):1221–1228, 2007.

[450] Gary E Schwartz and Stephen M Weiss. Behavioral medicine revisited:An amended definition. Journal of Behavioral Medicine, 1(3):249–251, 1978.

[451] Leah M Scolere, Eric PS Baumer, Lindsay Reynolds, and Geri Gay. Build-ing mood, building community: Usage patterns of an interactive art in-stallation. In Proceedings of the 19th International Conference on SupportingGroup Work, pages 201–212. ACM, 2016.

[452] Saul Shiffman. Real-time self-report of momentary states in the naturalenvironment: Computerized ecological momentary assessment. The sci-ence of self-report: Implications for research and practice, pages 277–296, 2000.

[453] Katie Shilton. Participatory personal data: An emerging research chal-lenge for the information sciences. Journal of the American Society for Infor-mation Science and Technology, 63(10):1905–1915, 2012.

[454] Choonsung Shin, Jin-Hyuk Hong, and Anind K Dey. Understanding andprediction of mobile application usage for smart phones. In Proceedings ofthe 2012 ACM Conference on Ubiquitous Computing, pages 173–182. ACM,2012.

[455] Alireza Sahami Shirazi, James Clawson, Yashar Hassanpour, Moham-mad J Tourian, Albrecht Schmidt, Ed H Chi, Marko Borazio, and KristofVan Laerhoven. Already up? using mobile phones to track & share sleepbehavior. International Journal of Human-Computer Studies, 71(9):878–888,2013.

[456] Itamar Simonson. Determinants of customers’ responses to customizedoffers: Conceptual framework and research propositions. Journal of Mar-keting, 69(1):32–45, 2005.

[457] Charles T Simpkin, Oskar G Jenni, Mary A Carskadon, Kenneth P Wright,Lameese D Akacem, Katherine G Garlo, and Monique K LeBourgeois.Chronotype is associated with the timing of the circadian clock and sleepin toddlers. Journal of sleep research, 23(4):397–405, 2014.

275

[458] Rashmi Sinha and Kirsten Swearingen. The role of transparency in recom-mender systems. In CHI’02 extended abstracts on Human factors in computingsystems, pages 830–831. ACM, 2002.

[459] Burrhus Frederic Skinner. Science and human behavior. Simon and Schuster,1953.

[460] Harvey A Skinner, Oonagh Maley, and Cameron D Norman. Develop-ing internet-based ehealth promotion programs. Health Promotion Practice,7(4):406–417, 2006.

[461] Aaron Smith. US smartphone use in 2015. Pew Research Center, 2015.

[462] Aaron Smith and Dana Page. The Smartphone Difference. 2015.

[463] Jaime Snyder, Mark Matthews, Jacqueline Chien, Pamara F Chang, EmilySun, Saeed Abdullah, and Geri Gay. MoodLight: Exploring personal andsocial implications of ambient display of biosensor data. In Proceedings ofthe 18th ACM Conference on Computer Supported Cooperative Work & SocialComputing, pages 143–153. ACM, 2015.

[464] Isabella Soreca, Ellen Frank, and David J Kupfer. The phenomenology ofbipolar disorder: what drives the high rate of medical burden and deter-mines long-term prognosis? Depression and anxiety, 26(1):73–82, 2009.

[465] Bonnie Spring, Marientina Gotsis, Ana Paiva, and Donna Spruijt-Metz.Healthy apps: mobile devices for continuous monitoring and interven-tion. IEEE pulse, 4(6):34–40, 2013.

[466] Katarzyna Stawarz, Anna L Cox, and Ann Blandford. Beyond self-tracking and reminders: designing smartphone apps that support habitformation. In Proceedings of the 33rd Annual ACM Conference on HumanFactors in Computing Systems, pages 2653–2662. ACM, 2015.

[467] Naomi J Steiner, Elizabeth C Frenette, Kirsten M Rene, Robert T Brennan,and Ellen C Perrin. In-school neurofeedback training for ADHD: sus-tained improvements from a randomized control trial. Pediatrics, pagespeds–2013, 2014.

[468] Andrew Steptoe, Katie O’Donnell, Michael Marmot, and Jane Wardle.Positive affect, psychological well-being, and good sleep. Journal of psy-chosomatic research, 64(4):409–415, 2008.

276

[469] Richard G Stevens, David E Blask, George C Brainard, Johnni Hansen,Steven W Lockley, Ignacio Provencio, Mark S Rea, and Leslie Reinlib.The role of environmental lighting and circadian disruption in cancer andother diseases. Environmental health perspectives, pages 1357–1362, 2007.

[470] Arthur A Stone and Saul Shiffman. Ecological momentary assessment(EMA) in behavorial medicine. Annals of Behavioral Medicine, 1994.

[471] Arthur A Stone, Saul Shiffman, Joseph E Schwartz, Joan E Broderick, andMichael R Hufford. Patient compliance with paper and electronic diaries.Controlled clinical trials, 24(2):182–199, 2003.

[472] Valerie Sugarman and Edward Lank. Designing persuasive technologyto manage peak electricity demand in Ontario homes. In Proceedings ofthe 33rd Annual ACM Conference on Human Factors in Computing Systems,pages 1975–1984. ACM, 2015.

[473] JaYoung Sung, Rebecca E Grinter, and Henrik I Christensen. Pimp MyRoomba: Designing for Personalization. In Proceedings of the SIGCHI Con-ference on Human Factors in Computing Systems, pages 193–196. ACM, 2009.

[474] Mervyn Susser. Epidemiology in the United States after World War II: theevolution of technique. Columbia University, 1985.

[475] Melanie Swan. Health 2050: The realization of personalized medicinethrough crowdsourcing, the quantified self, and the participatory biociti-zen. Journal of Personalized Medicine, 2(3):93–118, 2012.

[476] Melanie Swan. The Quantified Self: Fundamental disruption in big datascience and biological discovery. Big Data, 1(2):85–99, 2013.

[477] Leslie M Swanson, J Arnedt, Mark R Rosekind, Gregory Belenky,Thomas J Balkin, and Christopher Drake. Sleep disorders and work per-formance: findings from the 2008 National Sleep Foundation Sleep inAmerica poll. Journal of sleep research, 20(3):487–494, 2011.

[478] Richard F Taflinger. Taking advantage. Kendall Hunt Publishing Company,2010.

[479] Toshiyo Tamura, Tatsuo Togawa, Mitsuhiro Ogawa, and Mikiko Yoda.Fully automated health monitoring system in the home. Medical engineer-ing & physics, 20(8):573–579, 1998.

277

[480] Maria Taramigkou, Efthimios Bothos, Dimitris Apostolou, and GregorisMentzas. Fostering serendipity in online information systems. In En-gineering, Technology and Innovation (ICE) & IEEE International TechnologyManagement Conference, 2013 International Conference on, pages 1–10. IEEE,2013.

[481] Daniel J Taylor and Adam D Bramoweth. Patterns and consequences ofinadequate sleep in college students: substance use and motor vehicleaccidents. Journal of Adolescent Health, 46(6):610–612, 2010.

[482] Kenneth P Tercyak, Anisha A Abraham, Amanda L Graham, Lara D Wil-son, and Leslie R Walker. Association of multiple behavioral risk factorswith adolescents’ willingness to engage in ehealth promotion. Journal ofPediatric Psychology, 34(5):457–469, 2009.

[483] Jiuan Su Terman, Michael Terman, Ee-Sing Lo, and Thomas B Cooper.Circadian time of morning light administration and therapeutic responsein winter depression. Archives of General Psychiatry, 58(1):69–75, 2001.

[484] Sara Thomee, Lotta Dellve, Annika Harenstam, and Mats Hagberg. Per-ceived connections between information and communication technologyuse and mental symptoms among young adults-a qualitative study. BMCPublic Health, 10(1):1, 2010.

[485] Sara Thomee, Annika Harenstam, and Mats Hagberg. Mobile phone useand stress, sleep disturbances, and symptoms of depression among youngadults-a prospective cohort study. BMC public health, 11(1):1, 2011.

[486] Bill Thornton, Alyson Faires, Maija Robbins, and Eric Rollins. The merepresence of a cell phone may be distracting. Social Psychology, 2014.

[487] Alice Thudt, Sheelagh Carpendale, and Dominikus Baur. Autobiographi-cal visualizations: challenges in personal storytelling. 2014.

[488] Konrad Tollmar, Frank Bentley, and Cristobal Viedma. Mobile HealthMashups: Making sense of multiple streams of wellbeing and contextualdata for presentation on a mobile device. In 2012 6th International Confer-ence on Pervasive Computing Technologies for Healthcare (PervasiveHealth) andWorkshops, pages 65–72. IEEE, 2012.

[489] Eric J Topol. The creative destruction of medicine: How the digital revolutionwill create better health care. Basic Books, 2012.

278

[490] Khai N Truong, Thariq Shihipar, and Daniel J Wigdor. Slide to X: un-locking the potential of smartphone unlocking. In Proceedings of the 32ndannual ACM conference on Human factors in computing systems, pages 3635–3644. ACM, 2014.

[491] Dennis C Turk and Ronald Melzack. Handbook of pain assessment. GuilfordPress, 2011.

[492] Alexander Tuzhilin. Personalization: The state of the art and future direc-tions. Business computing, 3(3), 2009.

[493] Blase Ur, Pedro Giovanni Leon, Lorrie Faith Cranor, Richard Shay, andYang Wang. Smart, useful, scary, creepy: perceptions of online behavioraladvertising. In proceedings of the eighth symposium on usable privacy andsecurity, page 4. ACM, 2012.

[494] Gaetano Valenza, Mimma Nardelli, Antonio Lanata, Claudio Gentili,Gilles Bertschy, Rita Paradiso, and Enzo Pasquale Scilingo. Wearablemonitoring for mood recognition in bipolar disorder based on history-dependent long-term heart rate variability analysis. IEEE Journal ofBiomedical and Health Informatics, 18(5):1625–1635, 2014.

[495] Jan Van den Bulck. Text messaging as a cause of sleep interruption in ado-lescents, evidence from a cross-sectional study. Journal of Sleep Research,12(3):263–263, 2003.

[496] Fred van Gelderen. A brief history of radiology. In Understanding X-Rays,pages 597–602. Springer, 2004.

[497] Eus JW Van Someren, Cees Lijzenga, Majid Mirmiran, and Dick F Swaab.Long-term fitness training improves the circadian rest-activity rhythm inhealthy elderly males. Journal of biological rhythms, 12(2):146–156, 1997.

[498] Jan P Vandenbroucke. A short note on the history of the randomizedcontrolled trial. Journal of chronic diseases, 40(10):985–987, 1987.

[499] Eus JW VanSomeren. More than a marker: interaction between the circa-dian regulation of temperature and sleep, age-related changes, and treat-ment possibilities. Chronobiology international, 17(3):313–354, 2000.

[500] Wayne F Velicer, Colleen A Redding, Xiaowu Sun, and James O

279

Prochaska. Demographic variables, smoking variables, and outcomeacross five studies. Health Psychology, 26(3):278, 2007.

[501] Lois M Verbrugge. Health diaries. Medical care, 18(1):73–95, 1980.

[502] Joris C Verster, Seithikurippu R Pandi-Perumal, and David L Streiner.Sleep and quality of life in clinical medicine. Springer, 2008.

[503] Celine Vetter, Dorothee Fischer, Joana L Matera, and Till Roenneberg.Aligning work and circadian time in shift workers improves sleep andreduces circadian disruption. Current Biology, 25(7):907–911, 2015.

[504] Celine Vetter, Myriam Juda, and Till Roenneberg. The influence of in-ternal time, time awake, and sleep duration on cognitive performance inshiftworkers. Chronobiology international, 29(8):1127–1138, 2012.

[505] Jacqueline M Vink, Jacqueline M Vink, Alexia S Groot, Gerard A Kerkhof,and Dorret I Boomsma. Genetic analysis of morningness and eveningness.Chronobiology international, 18(5):809–822, 2001.

[506] Antoine U Viola, Lynette M James, Luc JM Schlangen, and Derk-Jan Dijk.Blue-enriched white light in the workplace improves self-reported alert-ness, performance and sleep quality. Scandinavian journal of work, environ-ment & health, pages 297–306, 2008.

[507] David T Wagner, Christopher M Barnes, Vivien KG Lim, and D LanceFerris. Lost sleep and cyberloafing: Evidence from the laboratory anda daylight saving time quasi-experiment. Journal of Applied Psychology,97(5):1068, 2012.

[508] Edward H Wagner, Brian T Austin, and Michael Von Korff. Organizingcare for patients with chronic illness. The Milbank Quarterly, pages 511–544, 1996.

[509] Ullrich Wagner, Steffen Gais, Hilde Haider, Rolf Verleger, and Jan Born.Sleep inspires insight. Nature, 427(6972):352–355, 2004.

[510] Olivia J Walch, Amy Cochran, and Daniel B Forger. A global quantifi-cation of “normal” sleep schedules using smartphone data. Science Ad-vances, 2(5):e1501705, 2016.

[511] Ryan J Walker, Zachary D Kribs, Andrew N Christopher, Oren R Shewach,

280

and Mareike B Wieth. Age, the Big Five, and time-of-day preference: Amediational model. Personality and Individual Differences, 56:170–174, 2014.

[512] James K Walsh. Clinical and socioeconomic correlates of insomnia. TheJournal of clinical psychiatry, 65(suppl 8):13–19, 2004.

[513] Monica L Wang, Lori Pbert, and Stephenie C Lemon. Influence of family,friend and coworker social support and social undermining on weightgain prevention among adults. Obesity, 22(9):1973–1980, 2014.

[514] Rui Wang, Fanglin Chen, Zhenyu Chen, Tianxing Li, Gabriella Harari,Stefanie Tignor, Xia Zhou, Dror Ben-Zeev, and Andrew T Campbell. Stu-dentLife: assessing mental health, academic performance and behavioraltrends of college students using smartphones. In Proceedings of the 2014ACM International Joint Conference on Pervasive and Ubiquitous Computing,pages 3–14. ACM, 2014.

[515] Yi Wang, Jialiu Lin, Murali Annavaram, Quinn A Jacobson, Jason Hong,Bhaskar Krishnamachari, and Norman Sadeh. A framework of energy ef-ficient mobile sensing for automatic user state recognition. In Proceedingsof the 7th international conference on Mobile systems, applications, and services,pages 179–192. ACM, 2009.

[516] Youcheng Wang and Daniel R Fesenmaier. Assessing motivation of con-tribution in online communities: An empirical investigation of an onlinetravel community. Electronic markets, 13(1):33–45, 2003.

[517] Crystal Rose Watson. The effects of mental health on spending patterns. AlliantInternational University, 2013.

[518] Thomas Webb, Judith Joseph, Lucy Yardley, and Susan Michie. Usingthe internet to promote health behavior change: a systematic review andmeta-analysis of the impact of theoretical basis, use of behavior changetechniques, and mode of delivery on efficacy. Journal of medical Internetresearch, 12(1):e4, 2010.

[519] Robin Weis. PersonifiedSelf: Crying, 2016. http://www.robinwe.is/explorations/cry.html.

[520] Ligia Westrich and Jeffrey Sprouse. Circadian rhythm dysregulation inbipolar disorder. Current opinion in investigational drugs (London, England:2000), 11(7):779–787, 2010.

281

[521] David P White. Monitoring peripheral arterial tone (PAT) to diagnosesleep apnea in the home. J Clin Sleep Med, 4(1):73, 2008.

[522] Evelyn P Whitlock, C Tracy Orleans, Nola Pender, and Janet Allan. Eval-uating primary care behavioral counseling interventions: An evidence-based approach. American journal of preventive medicine, 22(4):267–284,2002.

[523] Jennifer R Whitson. Gaming the Quantified Self. Surveillance & Society,11(1/2):163, 2013.

[524] M Wichers. The dynamic nature of depression: a new micro-level per-spective of mental disorder that meets current challenges. Psychologicalmedicine, 44(07):1349–1360, 2014.

[525] Gregory L Willis, Cleo Moore, and Stuart M Armstrong. A historical justi-fication for and retrospective analysis of the systematic application of lighttherapy in parkinson’s disease. Reviews in the Neurosciences, 23(2):199–226,2012.

[526] Anna Wirz-Justice. Diurnal variation of depressive symptoms. Dialoguesin clinical neuroscience, 10(3):337, 2008.

[527] Anna Wirz-Justice, Michael Terman, Dan A Oren, Frederick K Goodwin,Daniel F Kripke, Peter C Whybrow, Katherine L Wisner, Joseph C Wu,Raymond W Lam, Mathias Berger, et al. Brightening depression. Science,303(5657):467–469, 2004.

[528] Marc Wittmann, Jenny Dinich, Martha Merrow, and Till Roenneberg. So-cial jetlag: misalignment of biological and social time. Chronobiology inter-national, 23(1-2):497–509, 2006.

[529] Amy R Wolfson and Mary A Carskadon. Understanding adolescents’sleep patterns and school performance: a critical appraisal. Sleep MedicineReviews, 7(6):491–506, 2003.

[530] Wendy Wood and David T Neal. A new look at habits and the habit-goalinterface. Psychological review, 114(4):843, 2007.

[531] World Health Organization. The World Health report 2002: reducing risks,promoting healthy life. 2002.

282

[532] World Health Organization. Investing in mental health. Geneva: WorldHealth Organization, 2003.

[533] World Health Organization. Constitution of the World Health Organiza-tion. Basic documents, 45, 2006. http://www.who.int/governance/eb/who_constitution_en.pdf.

[534] World Health Organization. Global action plan for the prevention and controlof noncommunicable diseases 2013-2020. 2013.

[535] World Health Organization. Global status report on noncommunicable dis-eases 2014. 2014.

[536] Ying-Hui Wu and Dick F Swaab. Disturbance and strategies for reacti-vation of the circadian rhythm system in aging and alzheimer’s disease.Sleep Medicine, 8(6):623–636, 2007.

[537] Danny Wyatt, Tanzeem Choudhury, Jeff Bilmes, and James A Kitts. Infer-ring colocation and conversation networks from privacy-sensitive audiowith implications for computational social science. ACM Transactions onIntelligent Systems and Technology (TIST), 2(1):7, 2011.

[538] Fatos Xhafa. Advanced Technological Solutions for E-Health and DementiaPatient Monitoring. IGI Global, 2015.

[539] Xiaojiang Xu, Anthony J Karis, Mark J Buller, and William R Santee. Rela-tionship between core temperature, skin temperature, and heat flux dur-ing exercise in heat. European journal of applied physiology, 113(9):2381–2389, 2013.

[540] H Klar Yaggi, Andre B Araujo, and John B McKinlay. Sleep duration as arisk factor for the development of type 2 diabetes. Diabetes care, 29(3):657–661, 2006.

[541] Tingxin Yan, David Chu, Deepak Ganesan, Aman Kansal, and Jie Liu.Fast app launching for mobile devices using predictive user context. InProceedings of the 10th international conference on Mobile systems, applications,and services, pages 113–126. ACM, 2012.

[542] Amy Yang, Abraham A Palmer, and Harriet de Wit. Genetics of caffeineconsumption and responses to caffeine. Psychopharmacology, 211(3):245–257, 2010.

283

[543] Heng-Li Yang and Cheng-Yu Lai. Motivations of Wikipedia content con-tributors. Computers in human behavior, 26(6):1377–1383, 2010.

[544] Tal Yarkoni. Personality in 100,000 words: A large-scale analysis of per-sonality and word use among bloggers. Journal of research in personality,44(3):363–373, 2010.

[545] Gandhi Yetish, Hillard Kaplan, Michael Gurven, Brian Wood, HermanPontzer, Paul R Manger, Charles Wilson, Ronald McGregor, and Jerome MSiegel. Natural sleep and its seasonal variations in three pre-industrialsocieties. Current Biology, 25(21):2862–2868, 2015.

[546] Ting Yu, Simeon J Simoff, and Donald Stokes. Incorporating prior do-main knowledge into a kernel based feature selection algorithm. In Pacific-Asia Conference on Knowledge Discovery and Data Mining, pages 1064–1071.Springer, 2007.

[547] Phyllis C Zee, Hrayr Attarian, and Aleksandar Videnovic. Circadianrhythm abnormalities. CONTINUUM: Lifelong Learning in Neurology, 19(1,Sleep Disorders):132–147, 2013.

[548] Xiaoyi Zhang, Laura R Pina, and James Fogarty. Examining Unlock Jour-naling with Diaries and Reminders for In Situ Self-Report in Health andWellness. In Proceedings of the 2016 CHI Conference on Human Factors inComputing Systems, pages 5658–5664. ACM, 2016.

[549] Junhong Zhou, Dongdong Liu, Xin Li, Jing Ma, Jue Zhang, and Jing Fang.Pink noise: Effect on complexity synchronization of brain activity andsleep consolidation. Journal of theoretical biology, 306:68–72, 2012.

284

a framework for domain-driven development of personal health informatics technologies

Documents