Content uploaded by Mykola Pechenizkiy
Author content
All content in this area was uploaded by Mykola Pechenizkiy on Feb 12, 2015
Content may be subject to copyright.
What’s your current stress level?
Detection of stress patterns from GSR sensor data
Jorn Bakker, Mykola Pechenizkiy, Natalia Sidorova
Department of Computer Science
Eindhoven University of Technology
Eindhoven, P.O. Box 513, 5600MB,The Netherlands,
Email: {j.bakker,m.pechenizkiy,n.sidorova}@tue.nl
Abstract—The problem of job stress is generally recognized
as one of the major factors leading to a spectrum of health
problems. People with certain professions, like intensive care
specialists or call-center operators, and people in certain phases
of their lives, like working parents with young children, are
at increased risk of getting overstressed. Stress management
should start far before the stress start causing illnesses. The
current state of sensor technology allows to develop systems
measuring physical symptoms reflecting the stress level. In this
paper we (1) formulate the problem of stress identification and
categorization from the sensor data stream mining perspective,
(2) consider a reductionist approach for arousal identification as
a drift detection task, (3) highlight the major problems of dealing
with GSR data, collected from a watch-style stress measurement
device in normal (i.e. in non-lab) settings, and propose simple
approaches how to deal with them, and (4) discuss the lessons
learnt from the conducted experimental study on real GSR data
collected during the recent field study.
I. INTRODUCTION
Stress at work has become a serious problem affecting many
people of different professions, life situations, and age groups.
The workplace has changed dramatically due to globalization
of the economy, use of new information and communica-
tions technologies, growing diversity in the workplace, and
increased mental workload. In the 2000 European Working
Conditions Survey (EWCS) [12], work-related stress was
found to be the second most common work-related health
problem across the EU. 62% of Americans say work has
a significant impact on stress levels. 54% of employees are
concerned about health problems caused by stress. One in four
employees has taken a mental health day off from work to cope
with stress (APA Survey 2004).
Stress can contribute to illness directly, through its phys-
iological effects, or indirectly, through maladaptive health
behaviors (for example, smoking, poor eating habits or lack
of sleep) [4]. It is important to motivate people to adjust
their behavior and life style and start using appropriate stress
coping strategies. So that they achieve a better stress balance
far before increased level of stress results in serious health
problems.
The avoidance of stress in the everyday working environ-
ment is impossible. Still, if people are informed of their stress
levels, they become empowered for taking some preemptive
actions in order to alleviate stress [16].
What,When,Where,withWhom Physiologicalsigns
Stressdetectionandprediction
“Reschedule” “Takeabreak”“Prepare”
Coaching
StressͲ
detection&
prediction
models
Fig. 1. Stress@work in a nutshell: stress detection, prediction and coaching
There are a number of factors that are likely to cause stress
at work including but not limited to long work hours, work
overload, time pressure, difficult, demanding or complex tasks,
high responsibility, lack of breaks, conflicts, underpromotion,
lack of training, job insecurity, lack of variety, and poor phys-
ical work conditions (limited space, inconvenient temperature,
limited or inappropriate lighting conditions) [10].
In [1] we proposed the conceptual framework for managing
stress at work. One very important step in the process of stress
management is making the worker aware of the past, current or
expected stress. We aim at the automation of the identification
of the stress causes of an employee in question, as well as the
identification of the common causes of stress for employees
within an organisation. Figure 1 shows the main ideas of our
approach: We aim at making stress and stressors visible by
(1) keeping track of the calendar events and daily routine of the
worker, (2) measuring stress-related physiological signs from
the sensor data, (3) annotating these events with the sensor data
and the results of automated analysis of additional information
sources, such as sentiment classification of the incoming and
outgoing e-mails or social media messages [18] and explicit
Sympathetic system
Heartrate
Sweat production
Stress factor
Other
Other factors
Fig. 2. The reaction to stress factors is governed by the autonomous nervous
system. This path is shared with a lot of other mechanisms.
user feedback, (4) extracting the relationship between event
data and sensor data, i.e. relations between the increases and
decreases in the stress level with the characteristics of the
events of daily lives (what, where, when, with whom, etc.),
and (5) using extracted knowledge about this relationship for
personalized coaching.
In order to find this relationship, a number of subtasks
need to be done. One of the main subtasks is detecting
stress from the sensor data. Due to modern ICT and sensor
technologies, objective measuring of the stress level in non-
lab settings becomes possible. Such symptoms as voice, heart
rate, galvanic skin response (GSR) and facial expressions are
known to be highly correlated with the level of stress a person
experiences [3], [5], [7]. In this paper we focus on the use of
the GSR data (reflecting sweating) measured by a prototype
device worn at a wrist.
The direct use of the GSR measurements obtained is not that
straightforward. Partly this is caused by noise and inaccuracies
in the collected sensor data, but what is more crucial – the re-
action to various stress factors is governed by the autonomous
nervous system and this “path” to the symptomatic system is
shared with a lot of other mechanisms, such as the mechanism
of adaption to the outside temperature and humidity (Figure 2).
We have conducted a pilot case study aimed at the identi-
fication of likely challenges we need to address to make our
approach work in practice. In this paper, we focus only on
the problem of detecting changes in the stress level from the
GSR sensor data alone. We study the peculiarities of noise and
disturbances in the signal and argue the need of the related
contextual data for improving the quality of stress detection.
The rest of this paper is organized as follows. In Section II,
we formulate the problem of stress identification and cate-
gorization from the sensor data stream mining perspective.
We focus on a subproblem of arousal identification in online
settings, which we formulate as a drift detection task. We
highlight the major problems of dealing with GSR data, col-
lected from a watch-style stress measurement device in normal
(i.e. in non-lab) settings, and propose simple approaches how
to deal with them. In Section III we present the results and
lessons learnt from the conducted experimental study on real
GSR data collected during the recent pilot field study. Finally
in Section IV we give conclusions and discuss directions for
further work.
II. ACUTE STRESS IDENTIFICATION
Stress comes in three flavors:
1) Acute: stress caused by an acute short-term stress factor.
2) Episodic acute: acute stress that occurs more frequently
and/or periodically.
3) Chronic: stress caused by long-term stress factors and
can be very harmful in long run.
Most people experience acute stress during their everyday
life. It is a primal flight-or-fight response to immediate stress
factors and is not considered harmful. When the frequency
of these occurrences increase, physiological symptoms might
occur. This type of stress is associated with a very busy and
chaotic life and can be considered to be harmful when it occurs
over prolonged periods of time. The last type of stress, chronic,
is considered to be the most harmful. Prolonged periods of
stress could be caused by personal circumstances or other
long-term factors.
In our work, we want to prevent people from transferring
to the chronic category and therefore, we target the acute and
episodic acute stress. Particularly, in this paper we focus on
the identification of acute stress in order to facilitate coaching
of the episodic acute stress.
Acute stress is a mechanism that brings the body into a
state of alertness. As shown in Figure 2, it is controlled by the
autonomous nervous system. This system maintains a constant
equilibrium (also known as homeostasis). A change in this
equilibrium results in different changes in the bodily functions
(e.g. activity of digestion system).
Stress can be seen as a state of emergency that is preceded
by arousal due to an external stimulus, see Figure 3. After the
factor causing stress (the stressor) disappears, the body relaxes
and returns to a normal state.
Figure 4 shows the general case with more relationships
between the four states depicting the inner process of stress.
The problem of stress identification can be formulated in
different ways, e.g. as a traditional classification task, as one-
class classification, as event identification, and as time series
subsequence classification to name a few main options.
It should be also noticed that acute stress can also be pos-
itive (e.g. caused by an excitement or an intrinsic motivation
or an engagement in the working process), and, consequently,
staying in a normal state for too a long period without any
acute stress can be a sign of monotone uninteresting work or
poor motivation of the employee. Therefore, we would like
to perform a more detailed classification of the states in the
future.
In this paper we consider a simplified setting assuming
that a person is either in the normal state or in a stressed
state. The change between the two states can be sudden or
incremental; typically, arousal is more rapid and relaxation
Normal Aroused RelaxingStressed
Fig. 3. An example of acute stress pattern observed from GSR data and how
it can be mapped to the symbolic (time-stamped) representation of person’s
stress.
Normal
Arousing
Stressed
Relaxing
Fig. 4. Four states depicting the inner process of stress.
takes considerably longer. As we will show, different change
patterns can be observed.
A. Arousal as change detection
The principal task is to detect whether a person is stressed
at a particular moment in time or not. In other words, the
detector assigns a label “stressed” or “not stressed” based on
the observed historic data.
Detecting changes in GSR data is not as straightforward
as someone might think looking at the example in Figure 3.
Different types of noise in the data and changes in GSR data
due to other factors than stressors make it a non-trivial task. In
this section, we give illustrative examples of noise and other
factors affecting the GSR signal.
Types of noise. The quality of the GSR signal depends
primarily on the continuity of the contact between the device
and the skin of the test person. The skin conductance is
measured by two electrodes that require skin contact in order
to produce a reliable signal. However, this contact is not the
same for every person. For some people, the device fits less
well (e.g. because they dislike wearing it tight enough to
guarantee good contact, or because they have very dry skin);
due to a poor fit, we get noise in the signal (see Figure 6).
A person might also accidently touch the device (or do this
periodically in case of having such a habit), thus increasing
the pressure and influencing the GSR measurement; this also
12
0
5
10
15
20
25
Time, hours
GSR
Noise
Fig. 5. The GSR signal contains two-sided local noise peaks that are probably
caused by a physical disturbance of the contact between the skin and the
sensors, e.g. if someone has a habit to touch from time to time the watch or
the stress meter in this case.
14 15
0
5
10
15
20
Time, hours
GSR
Gaps
Fig. 6. When the fit between the skin and the sensors is not tight enough,
the contact is continuously broken. A characteristic of this behavior is the
high amount of gaps (ground value of sensor) in the signal.
creates noise in the signal in the form of gaps (see Figure 5).
Note that the skin in contact with the device contains
slightly more sweat than the skin next to the device, and when
the device is shifted on the skin, there is a resettling period of
about 15 minutes during which the skin that came in contact
with the device gets about the same level of sweat as the
skin that was in contact with the device before the shift, thus
resulting in about the same GSR (under assumption that no
change in the stress level happens in this period).
Importance of context. There are a lot of different factors
that influence the internal state of a person. Rising GSR levels
might be related to a rise in temperature or to heavy physical
work or exercises. In other words, the GSR change patterns
can be related to contexts that are mostly hidden.
11 12 13 14 15 16 17
0
0.5
1
Time, hours
GSR
Fig. 7. Prior to a stressful event (red-lined peak), the GSR level is gradually
rising. Is this rise caused by an external factor or is it due to anticipation of
the event?
11 12 13 14 15 16 17 18 19 20
0
0.5
1
1.5
2
Time, hours
GSR
Fig. 8. After a stressful event (red-lined peak) the GSR level does not return
to the level it had prior to the event. This might indicate that there is no
relaxation process.
10 11 12 13 14 15 16 17
0
0.5
1
1.5
Time, hours
GSR
Different GSR levels
Fig. 9. After a suspected stressful event the GSR level does not return to
the level it had prior to the event. This might be an indicate that there is no
relaxation process or what is more like in this case - the baseline level of
GSR corresponding to normal unstressed state changed.
One of these patterns is a steady increase of the GSR
level (see Figure 7). This might be an indication of changing
environmental factors (e.g. temperature), but it might also be a
genuine stress response. For instance, once a certain event has
been scheduled, the person might get stressed in anticipation of
the event. This is an interesting pattern for the stress detection
task.
The same holds for the patterns in Figures 8 and 9. In these
time series there is a suspected stress peak: in Figure 8 the
red part corresponds to an event tagged by the user as being
stressful, in Figure 9 there is an untagged short-term increase
in the GSR level. In both cases, the GSR level does not return
to the original baseline after passing the peaks. The question
is whether this is due to continuous stress (because of the
user being still busy with what has happened) or some other
factors.
For some series we learnt from the users’ feedback that
certain patterns were caused by environmental factors or user
activity context. In Figure 10 the person is exercising between
12:00hr and 13:00hr. The effect of the exercises is clearly
visible in the GSR time series. Moreover, due to the form
and the intensity of the picks, we can discriminate those from
genuine stress.
These context-dependent patterns will be important in the
overall stress detection task. Knowing whether a person relaxes
after a stressful event or whether he or she experiences
anticipating stress is very important. Here we do not handle
these contexts explicitly.
10 11 12 13 14 15 16 17
2
4
6
8
10
12
14
Time, hours
GSR
Excercise end
Excercise start
Fig. 10. Doing physical exercises results in a high GSR level, yet is not
related to the emotional stress.
Preprocessing
Noise Filter
(Median filter)
Aggregation
(sec min)
Discretization
(SAX)
Raw sensor
data
Change
detection
!
tc
tcurr
nm
tcurr
y
tc
y.. y
f()
y=
()y’ =y’
g()
y’’= SAX(y’’)
Fig. 11. Arousal detection approach: the GSR data is first (1) filtered,
(2) aggregated, and (3) discretized in the preprocessing phase and then passed
to a change detection technique. Each step is applied to a window of data
that is kept until a change has been detected.
B. Approach
The main task is to determine whether the observed portion
of the signal contains a change that corresponds to an arousal.
Formulating this problem as a change detection task on
univariate time series, we consider a four step approach for
arousal detection as shown in Figure 11. In the preprocessing
phase we take the raw GSR sensor data and according to the
operational settings (i.e. offline vs. online) perform its filtering,
aggregation and discretisation. The processed data is served to
a change detection technique.
The purpose of arousal detection can be twofold. The first
is to obtain labels for the supervised learning process aimed at
finding relationships between stress occurrences and external
events of factors causing stress. In this case we can perform
change detection in offline settings, i.e. the complete data
series can be used in preprocessing and detection steps. The
second purpose is to use an online detection mechanism in
online or semi-online settings as an alarm for making the user
aware of stress (and possibly asking for feedback that can be
related back to the subjective labeling process, i.e. the user
can confirm or reject the alert). Although we do not fix the
purpose of the task in this paper, we only describe an online
method that detects arousal for the point in time that might be
as much as a minute in the past.
Preprocessing. The three preprocessing steps that we use
are shown in Figure 11 and exemplified for the illustrative
purposes in Figure 12. The main objective of the preprocessing
phase is to remove noise from the GSR time series. The first
type of noise is due to poor contact between the sensors and
the skin (see Figure 6). If the contact is not sufficient, the
sensor will not measure anything. The second type of noise is
a local disturbance of the signal (see Figure 5). These local
disturbances are caused by mechanical movements (e.g. user
bumps device onto something) and should not be considered
to be actual measurements.
Noise caused by contact loss is problematic, since we cannot
be sure whether the signal can be trusted in these areas. In
these cases the frequency of the ground value (i.e. when the
sensors are not measuring anything) is a lot higher than in
a normal time series. When such periods occur in the GSR
signal, we alarm the problem and do not consider it further
in the arousal detection task. More specifically, we count
the number of occurrences of these faulty measurements and
exclude the time series if this number exceeds the number of
other points.
Noise caused by local disturbances must be filtered out
because they might be mistaken for genuine peaks. As shown
in Figures 3, one of the important parts of the arousal detection
task is to catch the transition from normal GSR levels to
aroused levels. This transition is characterized (for a typical
stress pattern) by a sudden peak in the GSR level. The filter
should filter out local disturbances while maintaining the
typical peaks. Therefore, the noise is filtered out by using
a median filter [14]. This is a filter that is used in image
processing, and it preserves edges (opposed to e.g. a moving
average) while filtering out noise.
The preprocessing step is applied to windows within the
window of kept data. Let ¯y=(ytc,...,y
tcurr )be the portion
of kept data from either the start or the last change point (ytc)
until the most recent sample (ytcurr ). The filter computes the
filtered values ¯y=f(¯y)over a moving window of size n
(n= 100 in the experiments) from ¯y1until ¯yk, where fis the
filter function and k=tcurr −tc. Each consecutive block of m
samples in ¯yis aggregated to one value, ¯y =g(¯y). In the last
preprocessing step, this data is discretised using SAX [8] into
a discrete time series from 1 to 5, SAX(¯y ). The levels can be
interpreted as being levels of stress (1: completely relaxed and
5: maximum arousal). However, they should not be interpreted
as absolute levels of arousal, but rather as a local relative
measure of arousal. Please, notice that discretisation of the
time series does not lead to an easy identification of the change
points (see Figure 13 for an illustrative example. However, the
dicretisation can help the change detector to be more accurate.
The signals are measured with a sampling frequency of
4 Herz, yet it does not make sense to expect the stress
detection to have timing requirements in the order of tenths of
seconds. For this reason, we aggregate the data to the order of
minutes. We use m= 240 in the experiments, thus after the
aggregation step 1 sample point ¯y
icorresponds to 1 minute.
In the experiments, we took ¯y
i=max(¯y
blocki). As said, in
10 11 12 13 14 15 16 17
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Time, hours
GSR
(a) Raw GSR signal
09 10 11 12 13 14 15 16 17 18
0
0.5
1
Time, hours
GSR
(b) Filtered GSR signal
050 100 150 200 250 300 350 400 450 500
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Time, minutes
GSR
(c) Aggregated GSR signal
050 100 150 200 250 300 350 400 450 500
0
1
2
3
4
5
6
7
Time, minutes
GSR (SAX)
(d) Discrete GSR data
Fig. 12. An example of GSR signal in its original form and after each of the
three individual steps in the data preprocessing: the raw GSR signal shown
in (a) is filtered using a median filter (b), then the values are aggregated to
the minute level (c), and finally they are discretised using SAX encoding (d)
to be used as an input for a change detection technique.
50 100 150 200
0
1
2
3
4
5
Time, minutes
GSR
GSR
ADWIN
SAX + MAX
ADWIN
Fig. 13. An illustration that discretising the data with SAX does not
immediately give us information about a change in arousal, e.g. by taking
a maximal value of the current window. The blue circles indicate the changes
alerted by such an approach. The red triangles indicate the change points
alerted by the ADWIN change detection method taking the SAXified time-
series as an input. ADWIN is considered below.
this setup it is important that the aggregation step is applied
after the filter in order to avoid the influence of local noise.
The discretisation using SAX is done online in a progressive
way. That is, the SAX representation is recomputed over
the historic data as new instances come within the training
window.
Change detection. Change detection in time series has been a
topic of interests in different domains. Existing approaches can
be divided into two broad groups of techniques. Techniques
from the first group are based on monitoring the evolution of
performance indicators like classification model accuracy or
some property of the data. Cumulative Sum (CUSUM), intro-
duced in [11] and recently used in [17] is one of the statistical
process monitoring mechanisms. This method monitors the
mean of the input data (that can be also any filter residual) and
gives an alarm when it is significantly different from zero, i.e.
deviates from the normal process behaviour. Other methods
rely on time series forecasting techniques such as Neural
Networks and Auto Regression functions [15] that estimate
parameter changes online based on an offline mapping.
Techniques from the second group are based on monitor-
ing distributions on two different time-windows: a reference
window summarizing past information and a window over
the most recent examples. Statistical tests based on Chernoff
bound, which decide whether samples drawn from two proba-
bility distributions are different, were studied in [6]. ADaptive
WINdowing (ADWIN) [2] that we use in our experimental
study keeps a variable-length window of recently seen data
points. It tries to keep the window of the maximal length that
is still statistically consistent with the hypothesis that there has
been no change in the mean signal value inside the window.
Thus, we consider two different approaches for change
detection. Both approaches are aimed at finding statistically
significant changes in data. The first approach that we call
here Fit is based on monitoring the model error, and the second
approach ADWIN is based on monitoring the data signal itself.
Both approaches were recently used for change detection in the
task of online prediction of the fuel mass flow in a boiler [13].
Fit: Performance monitoring-based change detection with
the non-parametric test. In this study we assume that the
general pattern of arousal resembles the curve as shown in
Figure 3. We also assume that there is no global model that
predicts the general GSR signal for a person. Instead of using a
global model in combination with statistical change detection
methods, we opt for a method that computes local models.
If we assume that the stress level of a person is stable in be-
tween changes, the changes can be detected by monitoring the
error of a locally fitted model. Given historic (preprocessed)
data, the objective is to fit a simple regression model.Based on
the observed Mean Squared Error for the incoming points, we
can apply a statistic measure (e.g. Mann Whitney U test [9])
to determine whether a significant change in the prediction
error has occurred.
Every time a new point arrives, the data is split into two sets.
The first set is a reference set that excludes the new point. The
second set is a test set that includes the new point. For each of
the two sets a model is trained while iteratively leaving out one
of the points. When there is an overall significant difference
between the two sets, it is considered to be a change point
and a cut is made.
ADWIN: Change detection based on raw data using adap-
tive windowing. ADWIN method works as follows: given a
sequence of signals it checks whether there are statistically
significant differences between the means of each possible split
of the sequence. If a statistically significant difference is found,
the oldest portion of the data backwards from the detected
point is dropped and the splitting procedure is repeated until
there are no significant differences in any possible split of
the sequence. More formally, given the GSR data stream,
suppose a1and a2are the means of the two subsequences
as a result of a split. Then the criterion for a change detection
is |a1−a2|>
cut, where
cut =1
2alog 4k
δ,a=1
1
k1+1
k2
,(1)
here kis total size of the sequence, while k1and k2are sizes
of the subsequences respectively.
TABLE I
DATA SET SUMMARY.
Number of users 5
Number of time series 72
Time series per user (mean) 14
Mean length (samples) 98721
Number of change points overall 368
Mean change points per series 6.5
III. EXPERIMENTAL STUDY
In this section we present the results from the conducted
experimental study on real GSR data collected during the
recent pilot field study. First, we give a concise description of
the constructed dataset and experiment setup, and then provide
a summary of the quantitative evaluation and some highlights
of the qualitative analysis of interesting cases.
Dataset description. Table I summarizes the main charac-
teristics of the data set. The data consists of the GSR data
measured on five persons in the course of the four weeks.
The data was collected from a watch-like device worn by
the persons during working hours. Since the sampling rate
is 4 Hz and the typical working day is roughly 8 hours, the
average length of the raw time series is 98721. All together
the data set contained 72 time series. 26 time series were
excluded from the experiments for either of the two reasons:
the GSR level showed very low variation or the contact of
the sensors was not sufficient to yield a usable signal (these
were detected automatically by a filter and then verified by
the visual inspection).
For each of the remaining 56 time series we annotated the
change points based on the visual inspection. Overall the set
of time series contains 368 change points with an average of
around 6.5 change points per time series.
The users participated in the study were instructed to anno-
tate any meeting in their agenda (MS Outlook Calendar) with
information about their feeling towards the meeting (“nice”,
“exciting”, “neutral”, “annoying”, or “tense”). Although this
information was available, it was not used in this investigation.
The reason for this is that the primary objective in this work
is to detect GSR peaks; however, a lot of the peaks do not
correspond to any meeting recorded in the agenda. Moreover,
the actual stress related to a meeting does not necessarily
shows up at the time of the meeting. It might precede the event
(see Figure 7) or continue to influence the person afterwards
(see Figure 8). In the ideal case, these labels reflect the state
transitions as shown in Figure 4, but in reality it is hard to
discern the separate state changes.
Therefore, instead of using the working agenda annotations
provided by the users, we used manually added labels based
on the visual inspection of the GSR time series. In the
experimental study presented in this paper, we labeled only the
change points, i.e. from the problem formulation perspective,
each point is labeled to be either a change point or not – that
our arousal detection approach will try to detect based on the
already observed GSR values.
TABLE II
TP AND FP RATES OF DETECTING THE CHANGE POINTS.THE MEAN μ
VALUES ARE PERCENTAGES WITH RESPECT TO PERFECT DETECTION.
μ(TP
P)σ(TP
P)μ(FP
TP+FP )σ(FP
TP+FP )
Fit 0.66 0.16 1.66 0.16
ADWIN 0.08 0.01 1.01 0.1
TABLE III
THE DISTANCE BETWEEN THE TIME OF THE ACTUAL CHANGE (ta)AND
THE TIME OF THE DETECTION (td).
μ(|ta−td|)σ(|ta−td|)
Fit 2.8 0.54
ADWIN 2.5 1.2
Experiment setup and evaluation. On each of 56 time series
we perform three steps: preprocess the data as discussed in the
previous section (see Figures 11 and 12), apply each of the
change detection methods, compare the labels to the changes
signalled by the method.
The techniques are applies on each time series in a progres-
sive way. That means that we assume that the data arrives as
a stream (one point at the time). Historic data is kept until a
change point is suspected. After that a new window is created
from the change point onwards.
The change points are evaluated by measuring the distance
between the point identified by a detection algorithm as a
change point and the closest actual change point within a
preset boundary threshold. The reason for doing this is that
there is no strict requirement that a change point should be
detected at exactly the point where it occurs. We should allow
for some leniency with respect to the actual time where it is
detected. Therefore, we measure the True Positive rate within
a window of 5 minutes around the actual change point. Instead
of the False Positive rate, the False Discovery rate is reported,
since the amount of True Negatives is very large with respect
to the True Positives.
Results. The results of the experiments are shown in Tables II
and III. As can be seen from Table II none of the methods
was able to catch all of the change points. The fit method
detected more change points than ADWIN, but at a cost of
more False Positives. The positioning of the change points is
better handled by ADWIN.
In Figure 14 and 15 there are a lot of False Positives in
the beginning of the time series. This is probably due to the
online encoding. In the beginning, if the signal is flat, small
fluctuations are blown-up by the discretization step. This might
lead to more False Positives. Yet the fit method shows this
behavior along the whole length of the time series.
There are two reasons why the True Positive rate is low for
ADWIN. The first is that it does not detect small peaks. The
second is that it also does not detect the change in cases where
the signal is slowly rising or falling (like in Figure 17).
Although we did not study thoroughly the effect of the
preprocessing techniques on the performance of the change
detection methods, some examples indicate that when time-
050 100 150 200 250 300 350 400
0
0.5
1
1.5
2
2.5
3
3.5
Time, hours
GSR
GSR
label
fit
ADWIN
Fig. 14. A flat signal followed by a high peak. On the down-curve of the
high peak there are many smaller peaks that are more difficult to detect.
050 100 150 200 250 300 350 400 450 500
0
2
4
6
8
10
Time, minutes
GSR
GSR
label
fit
ADWIN
Fig. 15. One of the stress time series and the change points. Green triangles
depict the ground truth, red diamonds depict the detection of the fit-method,
and the blue circles depict the detection of ADWIN.
series is filtered and then aggregated, but not discretised with
SAX, change detection may become less accurate (e.g. in
Figure 18 ADWIN missed two change points; cf. ADWIN
in Figure 16).
Discussion. The main difficulty of the stress detection task
is that arousal comes in many different forms. Since the
experiments were done in uncontrolled settings, it is difficult
160 170 180 190 200 210 220 230
0
2
4
6
8
Time, minutes
GSR
GSR
label
fit
ADWIN
Fig. 16. Closeup of the time series in Figure 15. ADWIN clearly detects the
high peaks, whereas the fit method is more sensitive to small local changes.
050 100 150 200 250 300 350 400
0
0.2
0.4
0.6
0.8
1
1.2
Time, minutes
GSR
GSR
label
fit
Fig. 17. Steadily increasing signal is not detected by ADWIN, yet there are
a lot of False Positives from the fit method.
160 170 180 190 200 210 220 230
0
2
4
6
8
Time, minutes
GSR
GSR
label
fit
ADWIN
ADWIN
Fig. 18. Detection results of ADWIN and fit on the same time series as in
Figure 16, but without SAX discretisation in preprocessing.
to interpret the patterns in the data. The manual labels are
not arbitrary, but their interpretation in terms of real arousal
is difficult.
Many examples suggest us that interpretation of the GSR
data can be rather ambiguous and deciding whether a particular
observed pattern corresponds to stress or something else (like a
physical exercise) is a non-trivial task even for a human expert.
(We asked the domain expert to analyze GSR curves like the
ones presented in the paper and he had confirmed that they
were ambiguous and additional information was required to
make a confident judgement whether the peaks correspond to
genuine stress or they are results of other factors). Therefore,
even “ideal” noise-free GSR data may be insufficient for
accurate determining the level of stress. This suggests that
the reliable translation of physiological data gathered by using
sensor technology into the “stress level rates” is only possible
when additional sources of information are available. For
example, apart from the GSR measurements, we can also use
measurements of acceleration in three dimensions. Exploring
the potential of accelerometer data for detecting the activity
context (e.g. physical exercises, walking, active discussion etc)
is an interesting direction for further research.
Other sources of additional data may include subjective
user feedback collected via questionnaires, annotation of the
events/signal, etc., as well as various external data extracted
e.g. from the social media, e-mail correspondence or electronic
agendas. Having access to such additional data facilitates the
use of pattern mining for finding relations between the in-
creases and decreases in the stress level with the characteristics
of the events of daily lives (what, where, when, with whom,
etc.).
IV. CONCLUSIONS AND FUTURE WORK
The detection of stressful events is a challenging task.
The information coming from sensor measurements is highly
ambiguous and dependent on hidden contexts. The detection
of separate stress peaks in the GSR data is also challenging
due to the varieties of patterns in the data. Moreover, it is
not clear without additional information whether certain peaks
correspond to a significant physiological process and how to
categorize them if they do.
In the further work, we plan to mine different sources of
data for stress detection and categorization. This includes the
statistics from the calendar, e-mail correspondence and social
media [18].
An additional source of information is the similarities or
differences between persons. Each person will handle stress
in a different way, but some might share characteristics when
it comes to anticipation, relaxing, or the general impact of
stress on observable variables. Using these sources of data
collected under more controlled settings we hope to be able
to get more reliable and more fine-grained categorization of
stress patterns.
ACKNOWLEDGEMENT
This research has been partly supported by EIT ICT Labs,
Health & Wellbeing thematic line (http://eit.ictlabs.eu) and
NWO HaCDAIS Project.
REFERENCES
[1] J. Bakker, L. Holenderski, R. Kocielnik, M. Pechenizkiy, and
N. Sidorova. Stress@work: From measuring stress to its understanding,
prediction and handling with personalized coaching. In Proc. of the
2nd ACM SIGHIT International Health Informatics Symposium, IHI’12.
ACM Press, 2012.
[2] A. Bifet and R. Gavald`
a. Learning from time-changing data with
adaptive windowing. In Proc. of the 7th SIAM Int. Conference on Data
Mining, SDM’07, 2007.
[3] W. Boucsein. Electrodermal activity. New York and London: Plenum
Press, 1992.
[4] K. Glanz and M. Schwartz. Stress, coping, and health behavior. Health
behavior and health education: Theory, research, and practice, pages
211–236, 2008.
[5] H. S. Hayre and J. C. Holland. Cross-correlation of voice and heart rate
as stress measures. Applied Acoustics, 13(1):57 – 62, 1980.
[6] D. Kifer, S. Ben-David, and J. Gehrke. Detecting change in data streams.
In Proceedings of the International Conference on Very Large Data
Bases, pages 180–191, Toronto, Canada, 2004. Morgan Kaufmann.
[7] P. J. Lang, M. K. Greenwald, M. M. Bradley, and A. O. Hamm.
Looking at pictures: Affective, facial, visceral, and behavioral reactions.
Psychophysiology, 30(3):261–273, 1993.
[8] J. Lin, E. J. Keogh, L. Wei, and S. Lonardi. Experiencing SAX: a
novel symbolic representation of time series. Data Min. Knowl. Discov.,
15(2):107–144, 2007.
[9] H. B. Mann and D. R. Whitney. On a test of whether one of two
random variables is stochastically larger than the other. Annals of Math.
Statistics, 18:50–60, 1947.
[10] S. Michie. Causes and management of stress at work. Occupational
and Environmental Medicine, 59(1):67, 2002.
[11] E. S. Page. Continuous inspection schemes. Biometrika, 41(1/2):100–
115, 1954.
[12] P. Paoli, D. Merlli´
e, and F. per a la Millora. Third European survey on
working conditions 2000. European Foundation for the Improvement of
Living and Working Conditions, 2001.
[13] M. Pechenizkiy, J. Bakker, I. ˇ
Zliobait˙
e, A. Ivannikov, and T. K¨
arkk¨
ainen.
Online mass flow prediction in cfb boilers with explicit detection of
sudden concept drift. SIGKDD Explor. Newsl., 11:109–116, May 2010.
[14] W. K. Pratt. Digital Image Processing. John Wiley & Sons, 1978.
[15] N. D. Ramirez-Beltran and J. A. Montes. Neural networks for on-
line parameter change detections in time series models. Computers &
Industrial Engineering, 33(1-2):337 – 340, 1997. Proc. of the 21st Int.
Conference on Computers and Industrial Engineering.
[16] P. Sanches, K. H ¨
o¨
ok, E. K. Vaara, C. Weymann, M. Bylund, P. Ferreira,
N. Peira, and M. Sj¨
olinder. Mind the body!: Designing a mobile stress
management application encouraging personal reflection. In Conference
on Designing Interactive Systems, pages 47–56, 2010.
[17] M. Severo and J. Gama. Change detection with Kalman Filter and
CUSUM. In Ubiquitous Knowledge Discovery, LNCS 6202, pages 148–
162. Springer Berlin / Heidelberg, 2010.
[18] E. Tromp and M. Pechenizkiy. Senticorr: Multilingual sentiment analysis
of personal correspondence. In Proc. of IEEE ICDM 2011 Workshops.
IEEE Press, 2011.
