stess@work: from measuring stress to its understanding ...mpechen/publications/pubs/... · stress...
TRANSCRIPT
Stess@work: From Measuring Stress to its Understanding,Prediction and Handling with Personalized Coaching
J. BakkerTU Eindhoven
L. HolenderskiPhilips Research
[email protected]. KocielnikTU Eindhoven
M. PechenizkiyTU Eindhoven
N. SidorovaTU Eindhoven
ABSTRACTThe problem of job stress is generally recognized as one ofthe major factors leading to a spectrum of health problems.People with certain professions, like intensive care special-ists or call-center operators, and people in certain phases oftheir lives, like working parents with young children, are atincreased risk of getting overstressed. For instance, one thirdof the intensive care specialists in the Netherlands are re-ported to have (had) a burn-out. Stress management shouldstart far before the stress start causing illnesses. The cur-rent state of sensor technology allows to develop systemsmeasuring physical symptoms reflecting the stress level. Wepropose to use data mining and predictive modeling for gain-ing insight in the stress effects of the events at work and forenabling better stress management by providing timely andpersonalized coaching. In this paper we present a generalframework, allowing to achieve this goal, and discuss thelessons learnt from the conducted case study.
Categories and Subject DescriptorsH.4 [Information Systems Applications]: Miscellaneous;J.3 [Life And Medical Sciences]: Health
1. INTRODUCTIONStress at work has become a serious problem affecting manypeople of different professions, life situations, and age groups.The workplace has changed dramatically due to globaliza-tion of the economy, use of new information and communica-tions technologies, growing diversity in the workplace, andincreased mental workload. In the 2000 European Work-ing Conditions Survey (EWCS) [7], work-related stress wasfound to be the second most common work-related healthproblem across the EU. 62% of Americans say work has asignificant impact on stress levels. 54% of employees areconcerned about health problems caused by stress. One infour employees has taken a mental health day off from workto cope with stress (APA Survey 2004).
Stress can contribute to illness directly, through its physio-logical effects, or indirectly, through maladaptive health be-haviors (for example, smoking, pore eating habits or lack ofsleep) [3]. It is important to motivate people to adjust theirbehavior and life style and start using appropriate stress cop-ing strategies. So that they achieve a better stress balancefar before increased level of stress results in serious healthproblems.
Yet, the avoidance of stress in the everyday working envi-ronment is impossible. Moreover, stress might not even beobserved as problematic by the persons themselves, for highlevels of stress are often perceived by people as a norm, asignal that they do their best to achieve their goals. Thefirst necessary condition for early signalling and treatmentof stress problems is introducing inexpensive, unobtrusive,and widely available technologies for creating awareness ofthe objective level of stress and the understanding of itscauses.
Due to modern sensor technologies, objective measuring ofthe stress level becomes possible. Such symptoms as heartrate, galvanic skin response (GSR) and facial expressionsare known to be highly correlated with the level of stressa person experiences. Moreover, due to extensive use ofsocial media and electronic agendas, it becomes possible todevelop methods and tools for the analysis of correlationsbetween the increases and decreases in the stress level withthe characteristics of the events of daily lives (what, where,when, with whom, etc.).
In this paper, we propose a framework that allows for (1) mea-suring physiological reactions to stress, (2) suggesting pos-sible interpretations of its causes based on the correlationsbetween the changes in the stress level and the events thattook place, and (3) giving effective advices for preventing thestress levels of a person to escalate to unhealthy proportions.
We have conducted a pilot case study aimed at the evalua-tion of the feasibility of the proposed approach and at theidentification of likely challenges we need to address to makeit work in practice.
The rest of this paper is organized as follows. We describeour general approach, focusing on the issue of stress levelmeasuring, identification of causes of stress and stress coach-ing, in Section 2. In Section 3 we report on our pilot case
study, emphasizing the lessons learnt from the explorationof real data capturing stress at work. Finally in Section 4we give conclusions and discuss directions for further work.
2. STRESS@WORK APPROACHThere are a number of factors that are likely to cause stress,called stressors. Stress and health risks at workplace can becategorized as content- and social or organizational context-related. Those that are intrinsic to the job include long workhours, work overload, time pressure, difficult, demanding orcomplex tasks, lack of breaks, lack of variety, and poor phys-ical work conditions (limited space, inconvenient tempera-ture, limited or inappropriate lighting conditions)[6].
Yet other causes of stress at work are conflicts, limitations,and high responsibility. Also underpromotion, lack of train-ing and job insecurity are considered stressful. Continuing,managers who are critical, demanding or unsupportive cre-ate stress as well. On the other hand, job development op-portunities, positive social dimension of work and good teamworking are all important buffers against stress.
From the organizational point of view, a culture of unpaidovertime causes stress. On the other hand, cultures wheredecisions are shared, where people are informed about thestatus of the organization, reduce stress. Organizationalchanges, especially when consultation has been inadequate,are a huge source of stress. Such changes include mergers,relocations, restructuring or “downsizing”, individual con-tracts, and redundancies within the organization [6].
We aim at the automation of the identification of the stresscauses of an employee in question, as well as the identifi-cation of the common causes of stress for employees withinan organisation. Some of the causes of stress can only bediscovered by interviewing the person in question (e.g. un-derpromotion). We only aim at the causes of stress that canbe discovered by unobtrusive monitoring of the employee’sbehaviour, and therefore we focus on the causes of stressthat can be discovered from data sets including physiolog-ical measurements allowing to estimate the curve of stresslevel over a period of time and information about the eventshappened within this period of time.
We present a high-level illustration of the proposed frame-work in Figure 1. Particularly, it highlights three majorsteps of data fusion and mapping, stress detection and pre-diction with already available models and coaching in theoperational i.e. online settings. In the offline settings themodels for stress detection and prediction are discoveredfrom the historical data using state-of-the-art data miningtechniques [8] and then verified by the domain expects. Inthe rest of this section we consider the most essential partsin more detail and provide illustrative examples.
2.1 Measuring stress levelThe physiological measurements providing the bio-signalsreflecting the level of stress that we currently use are gal-vanic skin response (GSR) and accelerometer data aboutmovements. This data is gathered using a prototype de-vice [10].
Following the conclusions of [4], we assume that increases in
What, When, Where, with Whom Physical signs
Stress detection and predictionSt
CoachingStress‐
detection & prediction models
“Reschedule” “Take a break”“Prepare”
Figure 1: Three major processes of the stress@workin the operational settings: i) mapping MS Outlookrecords to sensor data, ii) detecting or predictingstress based on already discovered patterns from thehistoric data and background knowledge of stress, iii)personalized coaching
the GSR correspond to emotional arousal. Our objective isto detect the cases and the types of such arousals reflected inthe GSR changes. Then we discriminate the changes in theGSR based on the positive and negative labeling of stress.
Emotional arousal can be seen as a deviation from the “nor-mal” relaxed state, so the occurrence of arousal is a formof concept shift or drift [9]. “Stress-as-concept-shift” occurswhen a sudden startle effect of short duration is measured.However, in some cases the person feels stress long beforethe event in anticipation of it. In this case a slow drift couldbe present prior to the event. That also means that thetime at which this change occurs varies. We also need totake into account that the response of a person can varyover time due to unobserved circumstances. From the algo-rithmic perspective we can use state-of-the-art approachesfor detecting drift in time-series signal like ADWIN [1]. Po-tentially, we can also categorize the observed stress followingshape-based time-series classification paradigm [11].
2.2 Identifying causes of stressWe currently use the data of employee’s MS Outlook agendain order to acquire information about the events that tookplace during the observation period. We take care of the pri-vacy issues by anonymising all the agenda information andencoding it with a mapping from words used in the agendainto codes, with only codes being visible to the system sup-port group (or to our project members, in our pilot casestudy). Since the same words are replaced with the samecodes, we can use off-the-shelf data mining techniques in or-der to find correlations between the changes in GSR and thecharacteristics of the events that took place.
The most important characteristics of the events recorded in
TIME CONTENT LOCATION PARTICIPANTS STRESS
Thursday 5 1 4 High
Friday 3 2 12 Low
Monday 5 2 5 Medium
Wednesday 2 1 11 Low
Friday 5 1 4 High
Monday 5 3 2 Low
Tuesday 3 1 2 High
Outlook Sensor
Content = 5 2 < PARTICIPANTS < 10 Stress = High
Location = 1 Stress = High
Participants > 10 Stress = Low
Figure 2: Example of finding individual stress pat-terns with class association rules.
the agenda are what event took place (e.g. a project meeting,with the project title being included into MS Outlook record,which allows it to be compared to other events related to thisproject), when it took place (e.g. at the end of the workingday), whom the employee met at the event (which thus givesa way to signal problems in personal relations), who initiatedthe meeting (thus allowing to see the autonomy level of theemployee) (see Figure 1).
Consider a hypothetical example presented in Figure 2. Theobservations suggest that there is a correlation between thestress level and the number of participants of the meet-ing (with passive behaviour and daydreaming when beingin large groups of people as a possible reason), as well asa correlation of the stress level with the meeting location(that could be a room for meetings with clients). Auto-mated discovery of such correlations is known as associativeclassification [5, 2] in data mining.
Note that the information from the agenda allows us to iden-tify possible indications to the causes of stress that are spe-cific for an employee, like workload (overload and under-load), pace, variety, autonomy, hours of work, interpersonalissues and communication patterns, and the causes of stressthat are common to multiple employees within an organisa-tion, like a stressful project (or a difficult class for teachers),bad scheduling practices, general work overload, etc.
We discover individual and organisational stress reactionpatterns using standard data mining techniques [8]. Sincephysiological reactions to stress are highly individual, we askthe employees to tag events in their agenda reflecting theirpersonal perception of the event (e.g. boring or tense). Thistagging is used for calibrating GSR and accelerometer dataand for further interpretation of stress reaction patterns andunderstanding of the main causes of stress.
2.3 Stress coachingMaking an employee aware of his/her stress patterns andpossible interpretations of stressors can already be of greathelp, but we go beyond that and aim at providing coach-ing services based on the discovered patterns and continu-ous information stream from both the sensor device and the
Find most stressful day in
the future
Enough to reschedule?
Suggest rescheduling
Yes
Stressful enough?
Yes
No Voluntary relaxation techniques
No
Enough to prepare?
Yes Suggest preparation
Suggest distraction / relaxation
No
Find most stressful day in the past 3 days
Enough time has passed?
Stressful period is
close?
No
Suggest post-stress relaxation
No
Figure 3: A possible stress intervention algorithm.
Figure 4: Example of possible coaching by providingvisual overlay in MS Outlook.
agenda. Combining patterns with the information aboutforthcoming events and the current stress level of the em-ployees, we can make a predictive model of stress level forthe coming period of time and generate recommendations foradjustments that can prevent from escalating stress. Sim-ple examples of such recommendations are rescheduling ofcertain activities to balance workload (see Figure 4) or toavoid combining too many stressful events one after another,scheduling (and executing) a timely preparation activity wellbefore a stressful event, planning a relaxation event (whichcan be taking a coffee break with a nice colleague, a yogasession, going out in the evening, etc.). Figure 3 shows apossible intervention algorithm aimed at achieving a betterstress/relaxation balance.
Note that we can detect not only negative reactions to stres-sors but also positive reactions to relaxation events. Peopleare sometimes mistaken in their judgements about the typesof activities that help them to relax. Activities that theylike can still be stressful ones and they might be better beavoided in the extremely stressful periods. Our recommen-dations are therefore meant to be personalised based on theindividual preferences and reactions to possible relaxationoptions. We ask the user about a feedback to the recommen-dations generated by our system in order to further tune therecommendations to individual preferences.
3. PILOT CASE STUDYIn this section we present the highlights of the recently con-ducted pilot case study. Particularly we discuss the collecteddata, its exploration and the main lessons learnt from it.
3.1 Data collectionThe data was collected over the period of seven weeks fromfive employees wearing a prototype device for measurementof galvanic skin response (GSR) and movement (accelerom-eters) during working hours.
At the end of the working day, participants were download-ing the data from the device to their computers and taggedtheir daily activities in their schedule in MS Outlook withone of the following five tags: exciting – for events that werevery positive, nice – for events that were generally positive,neutral – for events that were generally not stressful, annoy-ing – for events that were a somewhat stressful, and tense –for events that were very stressful. After tagging the events,the participants were asked to answer three questions thatwere presented on a scale from very good to very bad:“Howdid you sleep last night?”, “Did you have a stressful day?”,and “How was your inner balance today?” and to provideany additional information in the free-form text about anyspecial circumstances or events that could have influencedthem, or physical measurements, or individual stress copingtechniques (e.g. jogging or careful planning of daily activi-ties).
During the seven weeks period of the data collection theparticipants were provided neither with any real-time stressrelated feedback no with an overview of the collected GSRmeasurements in order to prevent possible bias. The studyas such was observational, i.e. no stressful events or inter-ventions had been preplanned; the participating employeeswere acting as during regular working weeks.
3.2 Data explorationWe performed the visual exploration of the collected data,aligning raw GSR sensor signal with the employee calendardata. One of the most straightforward analysis that we per-formed and the only one that we report here (due to thespace restriction) is relating GSR data with the providedlabels of stress (or no stress).
The following three figures illustrate examples of annotatedGSR signals. Stressful events are marked as red lines, neu-tral events as blue, and the black lines indicate regions whereno event information was provided.
One of the patterns that is easily recognized is the short-term response to a sudden cause of stress (see Figure 5).This pattern is a single peak that has a relatively short timespan.
An important thing to know is how long a stressful event willhave an effect on the person. Since it is a state of arousal inwhich the body consumes more of its’ resources. In Figure 6the GSR values do not go down to the original level afterthe stressful event has ended. This might be an indicationthat there is no relaxation period.
Since the events are scheduled meetings, it could be that
10 11 120
0.1
0.2
0.3
0.4
0.5
0.6
time, hours
GS
R
Figure 5: A short-term stressful event.
10 11 12 13 14 15 16 17 18 19 20 210
0.5
1
1.5
2
2.5
time, hours
GS
R
Figure 6: A longer-tem stressful event not followedby an immediate relaxation phase.
the stress levels already build up at the moment the personknows of this meeting. In Figure 7 it seems as if the stressfulevent at the end of the day was already anticipated at thebeginning of the day. From these instances we can see thatthere are indeed different types of change in response.
The following case (see Figure 8) illustrates that our stresslevel detection solutions should be context-aware, i.e. be sen-sitive to other relevant information describing current activ-ity of the person and the environment (s)he is currently in.As the figure suggests, a physical activity causes sweatingthat may result in a false stress detection.
3.3 Main lessons learntThe results of this pilot case study has confirmed some of ourexpectations about the challenging related to the analysis ofGSR data. In particular, the baseline level and variance ofthe signal differ not only from one person to person but alsoe.g. from one day to another for the same person. There aremany reasons for that, including e.g. the tightness and theexact placement of the device on the wrist, and physical ac-tivities (can be traced analyzing meta-data, e.g. labels like“meeting”, and accelerometer’s data) among others. Thisfact brings additional challenges in the task of finding com-mon stress related patterns and the necessity of continuous
10 11 12 13 14 15 16 170
0.2
0.4
0.6
0.8
1
1.2
1.4
time, hours
GS
R
Figure 7: The measured GSR increases before thestressful event takes place.
09 10 11 12 13 14 15 16 170
5
10
15
20
time, hours
GS
R
End physical excercise
Start physical excercise
Figure 8: The effect of the physical exercise on GSR.
13 14 15 16 17 18 190
200
400
600
800
1000
time, hours
GS
R Outliers
Figure 9: Example of noisy signal due to the bad con-tact of sensor with the skin.
fine-tuning of the the models to an individual, current ac-tivities and possibly an environment.
There are different kinds of noise in GSR signal. Some ofthem are easy to filter out (single point outliers, see Fig-ure 9) or to decide that the signal cannot be trusted duringparticular period (cases caused e.g. by poor contact betweenthe sensor and the skin, or by too much sweating, see Fig-ure 10). However, there are many less straightforward casesrequiring algorithmic solutions to be developed.
We could see at least three distinct patterns of stress: burstsof stress, rapidly gained but long-lasting stress, and incre-mental stress. This calls for developing algorithmic ap-proaches for detecting, categorizing and predicting differentkinds of stress pattern occurrences.
4. CONCLUSION AND OUTLOOKStress at work is not uncommon for many professions. Whilesome stress is a normal part of work, excessive stress or highlevels of stress over prolonged periods of time can interferewith employee’s productivity and have serious implicationsfor the physical and emotional health of the person. Asa worker, being aware of your own stress levels is alreadyan important step towards the prevention of diseases andincrease of the productivity. Our approach can make peoplemore aware of the evolution and causes of their stress byrelating stress patterns to their daily activities. In this paperwe presented some of the lessons learnt from the case study
13 14 15 16 17 18 190
5
10
15
20
time, hours
GS
R
Figure 10: Example of noisy signal due to the exceed-ing volume of sweat near the sensor.
aimed to identify the challenges we should anticipate in thedevelopment of the stress@work coaching services.
Our further research includes conducting larger scale exper-iments targeting specific professions that have an increasedexposure to stress, particularly school teachers and healthcarers working with elderly people. There we aim not onlyat the discovery of individual stress patterns, but also at theanalysis of organization-wide patterns.An interesting prob-lem is in helping an employee to relate the occurrence andvolume of stress (s)he has with the productivity that canlead to the identification of personal “positive” and “nega-tive” stress patterns.
5. ACKNOWLEDGMENTSThis research has been supported by EIT ICT Labs, the-matic line Health & Wellbeing (http://eit.ictlabs.eu).
6. REFERENCES[1] A. Bifet and R. Gavalda. Learning from time-changing
data with adaptive windowing. In SDM. SIAM, 2007.
[2] D. Chodos and O. R. Zaıane. Arc-ui: Visualizationtool for associative classifiers. In IV, pages 296–301.IEEE Computer Society, 2008.
[3] K. Glanz and M. Schwartz. Stress, coping, and healthbehavior. Health behavior and health education:Theory, research, and practice, pages 211–236, 2008.
[4] J. Healey and R. Picard. Detecting stress duringreal-world driving tasks using physiological sensors.IEEE Trans. on Intelligent Transportation Systems,6(2):156–166, June 2005.
[5] M. J. Heravi and O. R. Zaıane. A study oninterestingness measures for associative classifiers. InS. Y. Shin, S. Ossowski, M. Schumacher, M. J.Palakal, and C.-C. Hung, editors, SAC, pages1039–1046. ACM, 2010.
[6] S. Michie. Causes and management of stress at work.Occupational and Environmental Medicine, 59(1):67,2002.
[7] P. Paoli, D. Merllie, and F. per a la Millora. ThirdEuropean survey on working conditions 2000.European Foundation for the Improvement of Livingand Working Conditions, 2001.
[8] P.-N. Tan, M. Steinbach, and V. Kumar. Introductionto Data Mining. Addison-Wesley, 2005.
[9] A. Tsymbal. The problem of concept drift: Definitionsand related work. Technical Report TCD-CS-2004-15,Trinity College Dublin, Ireland, 2004.
[10] J. Westerink, M. Ouwerkerk, G.-J. de Vries, S. deWaele, J. van den Eerenbeemd, and M. van Boven.Emotion measurement platform for daily lifesituations. In Proceedings VOLUME I, 2009International Conference on Affective Computing andIntelligent Interaction, ACII 2009, September 10-12,2009, Amsterdam, The Netherlands, pages 217–222.
[11] L. Ye and E. J. Keogh. Time series shapelets: a newprimitive for data mining. In J. F. E. IV,F. Fogelman-Soulie, P. A. Flach, and M. J. Zaki,editors, KDD, pages 947–956. ACM, 2009.
APPENDIX
Review procedure. This paper is submitted to the Sys-tems track. However, partly, it also belongs to the Analyticstrack as the paper reports on the results of the explorationof real data collected during the case study.
From the provided list of topics the paper relates most to:
• Health Information Systems and
• Datamining