thesis proposal presentation
TRANSCRIPT
Automated mobile systems for multi-dimensional well-being
sensing and feedback
Mashfiqui Rabbi
People-awareComputing lab
Sensing
Sensing/inference
Physical Social Mental health
Multi-dimensional well-being
Ubicomp 2011*
StressSense
Ubicomp 2012*
Feedback
Can we do more with sensing and context aware computing?
Fogg’s behavior model
Mot
ivat
ion
Ability
Non-action region
Action region
Habit loop
MyBehavior
Weight management & fitness
Physical activity
Food consumption
Deeper look at data
Walk in my home
Avoid large burgers
Telling me to continue walking in my office
freq × avg(cal)
f(distance)
Recommender system
Clustering Multi-armed bandit
Personalization
User A
User B
All computation inside the phone
How personalized and actionable?
Seeing how long I have been stationary and the low frequency of activity made me want to make a change. [A5, diary]
I followed the suggestion about eating the low-calorie and healthy food I had in the past... I stopped getting fries at lunch since with them my meal was approaching 1000 cal. [A9, diary]
Seeing how many times I have done this walk in the past, I feel I can do it again [A3, diary]
How personalized and actionable?
They seem like good generic suggestions{the kind you would read...as tips in a health magazine or some such... It recommends me to eat stuff that I don't have at home [G4, diary]
I don't tend to look at the suggestions. It makes me want to walk the dog that I wish I had, and eat food items that I wish I could eat, etc. [G3, diary]
It doesn't take my dietary restrictions.[G5, diary]
How personalized and actionable?
MyBehavior (M = 3.4; SD = 1.2)
Nutritionist (M = 2.5; SD = 1.6)
Statistically significant: t(223) = 4.04; p < 0.0001
Are MyBehavior suggestions users doing better?
MyBehavior Nutritionist
Human-in-the-loop
Prioritize Remove
Mechanical Turk based food labeling
Notification
User-centered design
Fogg’s behavior model
Mot
ivat
ion
Ability
Non-action region
Action region
Easy to implement suggestions
Continue / Small change
Context
Easy to implement suggestions
Continue / Small change
Context
Control belief/Barrier
Repetition= Mastery
Notification
Goal
How to achieve the goal?
MyPersonalCoach
Action Long term
Unhealthy HealthyGood habits
lifestyle lifestyle
Personalized Coaching
Action MasterySelf-efficacy
HabitRepeat
same context
Repeat
Habits
Conscious HabitsReflective effort Easy and unconscious
Conscious and unconscious
• Slow• Motivational
(conscious engagement)• Serial processing
• Fast• Unconscious• Parallel processing
Dual process theory
“I've got a more graceful solution to the memory problem. I'm disciplined and organized. I use habit and routine to make my life possible.”
Process of habits
Conscious HabitsReflective effort Easy and unconscious
Notification and Habit building
Goal
How to achieve the goal?
Goal
How to achieve the goal?
ConsequenceBelief
Control belief/Barrier
Goal
How to achieve the goal?
Goal
How to achieve the goal?
Effort
Performance
Reward
Valence
Goal
How to achieve the goal?
Building sensing systems
Open, efficient and scalable systems
are necessary to support multi-dimensional context-aware well-being solutions
SAINT
Sensing And INference Toolkit
SA
INT S
erv
er
Audio Accelerometer
Bus
Activity Recognition
Your DetectorSleepSpeech/Non-speech
WearablesActivity-aware
Location sensingStress
SA
INT S
erv
er
SA
INT
Ap
ps
Audio Accelerometer
Bus
Activity Recognition
MyBehavior MoodRhythm
BeWell Your app
Your DetectorSleepSpeech/Non-speech
Inter-process communication
Inter-process communication
WearablesActivity-aware
Location sensingStress
1. Can handle up to 50 audio clients in real time.
2. Battery drag is low even with 50 audio clients (~500mW).
3. Takes ~2 hours to code a client and server side detector for new coders.
QOSSaveBattery
Batch a lotNo batch
QOSSaveBattery
Batch a lotNo batch
High priority - 1
Low priority - 0
Med. Priority – 0.5
(1-p) * 100%
QOSSaveBattery
Batch a lotNo batch
High priority
Low priority
Medium priority
Battery low (1+w)*(1-p) * 100%
Works left and timeline
MyBehavior - Long term study
Multiple baseline
time
Nutritionist MyBehavior
Long term studyNov 1 – Feb 14, Ubicomp 2015
Pilot studyJournal of Medical and Internet Research
Iterative designPervasive Health 2015, or ToCHI
MyPersonalCoach
System building and pilot testSystem building: Dec 1 - Dec31Pilot study: Feb 1 – Mid MarchPublish: Sensys 2015
Long term studyCHI 2016
SAINT – robust and open-source
Stable version with more sensing and inferenceNov 1 – Nov 30, Submit as a journal afterwards
Contribution
Mot
ivat
ion
Ability
Non-action region
Action region
freq × avg(cal)
f(distance)
Prior work
Prior work
Prior work
Small Change Small Change Or Continue
4 min
2 min2.5 min
20 min
Explore/Opportunistic
Prior work
Human-in-the-loop
Prioritize Remove