Content uploaded by Tobias Kowatsch
Author content
All content in this area was uploaded by Tobias Kowatsch on Oct 01, 2017
Content may be subject to copyright.
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017
DESIGN AND LAB EXPERIMENT OF A
STRESS DETECTION SERVICE BASED ON
MOUSE MOVEMENTS
Research full-length paper
Track 12
Kowatsch, Tobias, University of St.Gallen, St.Gallen, Switzerland, tobias.kowatsch@unisg.ch
Wahle, Fabian, ETH Zurich, Zurich, Switzerland, fwahle@ethz.ch
Filler, Andreas, University of St.Gallen, St.Gallen, Switzerland, andreas.filler@unisg.ch
Abstract
Workplace stress can negatively affect the health condition of employees and with it, the performance
of organizations. Although there exist approaches to measure work-related stress, two major limitations
are the low resolution of stress data and its obtrusive measurement. The current work applies design
science research with the goal to design, implement and evaluate a Stress Detection Service (SDS) that
senses the degree of work-related stress solely based on mouse movements of knowledge workers. Using
van Gemmert and van Galen’s stress theory and Bakker and Demerouti’s Job Demands-Resource model
as justificatory knowledge, we implemented a first SDS prototype that senses mouse movements and
perceived stress levels. Experimental results indicate that two feature sets of mouse movements, i.e.
average deviation from an optimal mouse trajectory and average mouse speed, can classify high versus
low stress with an overall accuracy of 78%. Future work regarding a second build-and-evaluate loop
of a SDS, then tailored to the field setting, is discussed.
Keywords: Health Information System, Human-Computer Interaction, Workplace Stress, Task Perfor-
mance.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 2
1 Introduction
Work has changed over the past decades, resulting in steadily increasing workloads (Eurofound, 2012).
Moreover, working environments are one of the dominant predictors of mental health disorders (WHO,
2013). By 2020, 50% of the top ten medical problems worldwide will be stress-related with work con-
stituting a primary source of stress in modern society (Cartwright and Cooper, 2014). For example, a
recent study by Health Promotion Switzerland indicates that 25.4% of employees perceive work-related
stress and exhaustion, potentially leading to a loss of productivity to the amount of approximately 5.6
billion US Dollars (Igic et al., 2016).
Stress, defined as the “psychological and physical state that results when the resources of the individual
are not sufficient to cope with the demands and pressures of the situation” (Michie, 2002, p. 67), nega-
tively affects individual employees and with them, the performance of their organizations. In particular,
it threatens the health condition, quality of life, work-related goal achievements, self-esteem, confidence
and personal development (Michie, 2002; Nyberg et al., 2014). If job demands cannot be balanced by
individual resources in the long run, diastolic blood pressure, heart rate, blood glucose, cholesterol con-
centration, escapist drinking, and smoking among other symptoms of stress show first evidence of seri-
ous health conditions such as cardiovascular diseases, Type 2 diabetes, or mental disorders (Boedeker
and Klindworth, 2007; Geurts, 2014; Michie, 2002; Nyberg et al., 2014).
The detection of stress at the workplace represents therefore a prerequisite not only to anticipate any
negative health effects in the long term but also to offer just-in-time (organizational) health promotion
interventions (Mattke et al., 2013; Nahum-Shani et al., 2015; Nahum-Shani et al., 2016). Beyond organ-
izational and psychological barriers to directly report stress to colleagues or supervisors (Demerouti et
al., 2009), there exist several self-report instruments for measuring individual stress levels (Cohen et al.,
1983; Demerouti et al., 2003; Kessler et al., 2003; Siegrist et al., 2009). However, applying these instru-
ments has two major limitations: First, stress polls are usually conducted only two times per health
intervention with several weeks or even months in between, if at all health promotion programs are
implemented by corresponding organizations (Tims et al., 2013a; Tims et al., 2013b). Thus, the resolu-
tion of stress data is too low, i.e. short-term episodes of stress with serious negative health outcomes
cannot be identified reliably. Second, data collection with self-reports, if collected in higher frequencies,
is time-consuming, obtrusive and costly due to data collection, data analysis and data interpretation
activities.
Recent Information Systems literature, however, indicates significant relationships between motor ac-
tivity measured by mouse movements and emotional states (Grimes et al., 2013; Hibbeln et al., 2017).
In Human-Computer Interaction literature, it has been even shown a relationship between mouse move-
ments and perceived stress (Sun et al., 2014). However, with 71% the classification accuracy is limited,
as is the underlying mass-spring-damper model regarding the set of potential mouse features from which
perceived stress can be derived (ibid.). Our research question is therefore:
Which features of mouse movements are significantly related to the degree of perceived stress
such that a Stress Detection Service (SDS) is able to classify high versus low stress accurately?
In order to address this question, we apply design science research (Astor et al., 2013; Gregor and
Hevner, 2013; Hevner et al., 2004; Peffers et al., 2007; vom Brocke et al., 2013) with the goal to design,
implement and evaluate a SDS that senses the degree of perceived stress in knowledge workers solely
based on mouse movements. Combining the Job Demands-Resource (JD-R) model (Bakker and
Demerouti, 2007) and van Gemmert and van Galen (1997)’s stress theory with a particular focus on its
neuromotor noise concept as justificatory knowledge, we built the first two SDS modules that sense
mouse movements and perceived stress. Mouse movement data was collected during a laboratory ex-
periment and used to derive mouse movement features that can classify high versus low levels of per-
ceived stress accurately. These features finally inform the design of a third SDS module, the stress clas-
sification module, which is part of our future work.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 3
The remainder of this paper is structured as follows. Next, we present the conceptual foundations and
hypotheses of the current work from which the design requirements for the SDS are derived. Then, the
current implementation of the SDS is described. With a focus on the planned classification module and
the identification of relevant mouse features, we describe a lab experiment that was conducted to test
our hypotheses. We finally present the results of the lab experiment, discuss the limitations of this work
and provide an outlook of a second build-and-evaluate loop of the envisioned SDS in the context of a
longitudinal field study.
Finally, it must be noted that details about the motivation and related work of the envisioned SDS are
presented elsewhere (Kowatsch et al., 2015). By contrast, the current paper is distinct from that protocol
in so far as it adopts a design science approach and thus presents a build-and-evaluate loop including
theoretically derived design requirements, a description of the design artefact, i.e. the SDS, and presents
and discusses the actual empirical results from a first lab experiment.
2 Conceptual Foundations and Hypotheses
In this section, we present the conceptual foundations and hypotheses of the current work with the over-
all goal to outline the theoretical underpinnings of the SDS. The conceptual foundations are informed
by the JD-R model (Bakker and Demerouti, 2007) and stress theory (van Gemmert and van Galen, 1997)
with workplace stress being the focal theoretical construct. The JD-R model is adopted because it pro-
vides a holistic view of predictors and outcomes of work-related stress, i.e. it describes the broader
context of this research. By contrast, van Gemmert and van Galen’s stress theory is considered appro-
priate because it explains the relationship between workplace stress, variations in the motor system
through the concept of neuromotor noise, and human performance. Thus, it represents the justificatory
knowledge for the relationship between the motor system of an individual, i.e. the proxy to mouse move-
ments, and the degree of workplace stress. An overview of the current work’s research framework is
depicted in Figure 1. The rationale of the theoretical constructs and their relationships are described in
the following two subsections.
Figure 1. Research framework. Note: arrows indicate positive (+) and negative (-) relationships
2.1 Job Demands-Resource (JD-R) Model
The JD-R model proposes that the degree of work-related stress is positively influenced by the degree
of mental, emotional and physical job demands and that job resources of employees, for example, their
skills, negatively moderate this relationship (Bakker and Demerouti, 2007). Job resources can compen-
sate the job demands in a way that workplace stress is reduced. Or, put in other words, an imbalance of
job demands and job resources in which the demands are higher than the resources leads to increased
levels of stress perceptions. Job resources do also positively impact the motivation of employees that,
in turn, can increase the outcomes. It must be noted that this motivational aspect is not shown in Figure
1 as the focus of this research lies on the measurement of workplace stress only. By contrast, workplace
stress is negatively associated with the outcome parameters on various levels such as the individual
level, e.g. impacting health, or the organizational level, e.g. work performance (Bakker and Demerouti,
2007; Ganster and Rosen, 2013; Kuoppala et al., 2008; Michie, 2002).
+
Work-related Stress
Job Resources
Job Demands
Perceived Stress
(Self-report)
Neuromotor Noise
(Mouse Movements)
Research
Question
?
Outcomes
Health Condition &
Work Performance
-
+
-
Workplace*stress*
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 4
2.2 Stress Theory and the Concept of Neuromotor Noise
In contrast to the JD-R literature, in which workplace stress is usually measured via self-report instru-
ments, van Gemmert and van Galen’s stress theory (van Gemmert and van Galen, 1997) puts a neuro-
physiological lens on workplace stress. The theory, which has been already adopted in human-computer
interaction research (e.g. Lin and Wu, 2011), suggests that an imbalance of high job demands and low
job resources is reflected by increased information processing demands. Moreover, these “increased
processing demands (e.g. in dual-task situations) lead to increased levels of neuromotor noise and, there-
fore, to decreased signal-to-noise ratios in the [motor, the author(s)] system.” (van Gemmert and van
Galen 1997, p. 1300) Here, neuromotor noise is generated by cognitive activities in the brain. Particu-
larly in high-demand work situations, neuromotor noise results from a competition of individuals’ in-
formation processing resources. The resulting decrease of the signal-to-noise ratio has direct effects on
the motor system, which can be measured by increased variations of human movements, i.e. micro-
movements (“shivering”). For example, mouse movements have been shown to be valid proxies of cog-
nitive and affective processing (Grimes et al., 2013; Maehr, 2008; Zimmermann et al., 2003) as they
provide “continuous streams of output that can reveal ongoing dynamics of processing, potentially cap-
turing the mind in motion with fine-grained temporal sensitivity.“ (Freeman et al., 2011, p. 1)
Against this background, if each mouse movement is guided by a target, for example, to hit the send
button of an email application or to drop a file in a folder, then neuromotor noise may probably add so
called micro-movements to the mouse trajectory that are not consciously processed by the individual.
That is, we would not only expect mouse movement deviations from an optimal trajectory as shown
Figure 2, but also an increase in mouse movement speed due to these micro-movements that increase
the distance to the mouse movement target.
Figure 2. Hypothesized mouse trajectories in low (top) versus high (bottom) stress situations.
The overall time, however, to reach the target might be still comparable to a low stress situation without
any micro-movements induced by neuromotor noise. We therefore formulate our hypotheses related to
our research question as follows:
H1: Higher (lower) deviations from an optimal mouse trajectory are related to high (low)
perceived stress.
H2: Higher (lower) speed of mouse movements are related to high (low) perceived stress.
3 Design Requirements for a Stress Detection Service (SDS)
The theoretical considerations and hypotheses allow us now to derive the requirements for the design of
our envisioned SDS. First and foremost, the target audience and work context of the SDS must be de-
fined to which the SDS should be tailored to. Consistent with related work (e.g. Koldijk et al., 2013;
Sappelli et al., 2014), we consider knowledge workers in the current research as the target population,
i.e. employees that are seen as “the most valuable asset of a 21st-century institution (whether business
or non-business)” (Drucker, 1999, p. 79). Knowledge workers interact significantly with their mouse
rendering these interactions a mirror of their work and relevant proxy of variations in the motor system
(Sappelli et al., 2014). Since measuring stress itself could increase workplace stress and thus, negatively
impact the work performance, the sensing must be designed in a way that is not obtrusive. The first
requirement is therefore defined as follows:
target
target
Neuromotor noise / High Stress
Low Stress
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 5
Requirement 1: The design artifact (SDS) hast to provide an unobtrusive interface to sense mouse move-
ments of knowledge workers that predominantly use their mouse to perform their work-
related activities.
Second, besides promising results with regard to the relationship between mouse movements and (a)
perceived emotions (e.g. Grimes et al., 2013; Hibbeln et al., 2017; Koldijk et al., 2013) and (b) perceived
stress (e.g. Sun et al., 2014), it is not yet conclusive whether neuromotor noise influences the motor
system in a way such that the degree of workplace stress perceptions can be measured accurately, i.e.
with accuracy levels close to 100%. Thus, the design artifact needs to measure the ground truth, i.e.
perceived stress, too. This is not only a requirement to identify relevant features of mouse movements,
as hypothesized above. It also makes the “static feature” SDS a rather self-learning SDS that improves
its stress detection accuracy with each additional labeled mouse movement episode (Barata et al., 2016;
Pejovic and Musolesi, 2015). However, bearing in mind the burden of intervention that increases with
each labelling activity, i.e. indicating the perceived level of stress by the knowledge worker, it is pro-
posed to conduct a lab experiment (see below) to initially derive relevant features of mouse movements
that are correlated to perceived stress levels, before the SDS is used in the field to improve its feature
set to increase classification accuracy in a training phase.
Requirement 2: The design artifact (SDS) hast to provide an interface to sense mouse movements of
knowledge workers that predominantly use their mouse to perform their work-related
activities.
Third, the SDS must able to classify the mouse movements into workplace stress perceptions (e.g. binary
high vs. low stress levels). Two potential features of mouse movements have been hypothesized above
and are evaluated in Section 5. In addition, to have an effect and to tackle any serious health condition
in an early stage, the classification results must be communicated through an appropriate interface to
various SDS endpoints. These endpoints could be quantified-self feedbacks tailored to the employee,
relaxation tips to decrease short term episodes of stress or an (organizational) health intervention service,
for example, offered by an organization’s physician or medical consultant (Cook et al., 2007; Mattke et
al., 2013). The final requirement is therefore formulated as follows:
Requirement 3: The design artifact (SDS) hast to classify the mouse movements into classes of work-
place stress and to communicate the result to health intervention services.
4 Implementation of the SDS
The SDS consists of two sensing modules, one for mouse movements (Requirements 1) and one for
perceived stress (Requirement 2). Moreover, it includes a stress classification module with an applica-
tion programming interface that provides a trigger to (external) health intervention services (Require-
ment 3). An overview of the modules and their relationships is depicted in Figure 3.
In the current SDS implementation, the sensing module for mouse movements is written as a Java appli-
cation that runs on several operating systems. It gathers a data stream of mouse coordinates and meta
information (e.g. single-click, double click, etc.), attaches timestamps to it, and appends these data tuples
to a text file. While previous work (Freeman and Ambady, 2010; Sun et al., 2014; Visser et al., 2004)
employed artificially high sampling rates, for example, 500Hz and even higher, and special mouse hard-
ware and software drivers to increase the resolution of mouse movements, we decided to utilize a non-
artificial but standard sampling rate of today’s operating systems and of-the-shelf computer mouse hard-
ware, i.e. about 125Hz on average. This does not only add external validity to the potential findings of
the current work – e.g. artificially down-sampling high resolution mouse movement data as done by Sun
et al. (2014) generates a potential systematic empirical bias – but may allow organizations to use stand-
ard mouse hardware they have already in place.
The sensing module for perceived stress relies on LimeSurvey, an open source platform for self-report
data collection purposes (LimeSurvey.org). We have written a script for LimeSurvey that stores the
timestamp of perceived stress self-report data with a pre-defined label. This allows us to synchronize
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 6
the self-report data on perceived stress at a particular point in time with the data stream of mouse move-
ments for later analyses. This sensing module provides the ground truth to identify the relevant features
of mouse movements. It is therefore in particular relevant for the design and training phase of the SDS.
However, when the feature set results in accurate stress classifications it is no longer needed. The burden
of measuring stress would therefore disappear completely over time.
Figure 3. Schematic overview of the Stress Detection Service architecture. Note: The dotted
sensing module is required only for the identification of features of mouse movements
that are able to accurately classify work-related stress. Thereafter, this module can be
used optionally to increase the classification accuracy with regard to individual em-
ployees in a training phase, too.
Finally, the stress classification module with an application programming interface is currently imple-
mented in the form of a MatLab (R2016a) script.1 With regard to the implementation described above it
becomes obvious that the SDS implementation is currently tailored to controlled evaluation environ-
ments in the lab. However, if the evaluation indicates accurate classification results based on mouse
movements, we plan to implement an SDS that better fits to a field and (organizational) setting, for
example, an SDS that uses a database for storing the mouse coordinates and timestamps, and provides
an appropriate user interface that does not distract from work activities.
5 Evaluation of Relevant Features of Mouse Movements
In this section, we describe a laboratory experiment that has the objective to identify features of mouse
movements that are significantly related to perceived stress levels. For this purpose, we use the two
hypothesized features from Section 2.2, i.e. the average deviation from an optimal mouse trajectory (H1)
and the average speed of mouse movements (H2). Results will inform the design of SDS’s stress clas-
sification module. We first outline the experiment and measures, before the results are reported.
5.1 Experimental Procedure and Measures
An overview of the lab experiment is shown in Figure 4 (left). The experiment was approved by the
ethical committee of the authors’ institution. After subjects gave their informed consent, they were ran-
domly assigned to an experimental group (high stress condition) and a control group (low stress condi-
tion). They were seated in front of a computer screen and were given a mouse to work with until the end
of the experimental session. In addition, physiological arousal was measured throughout the experi-
mental session with the help of a skin conductance (SC) sensor on the non-dominant hand to control for
any physiological differences in the subjects that might impact the neuromotor noise effect. For that
purpose, a medical device (MindMedia’s NeXus-10) was used and the total sum of standardized SC
reactions (SCR) for each part of the experiment that involved mouse movements was calculated (see
1 The authors are happy to provide full access to the source code and the complied version of the sensing module of the current
SDS implementation, the LimeSurvey and MatLab scripts.
Stress Detection Service (SDS)
Knowledge
Worker
Mouse
Sensing of Mouse
Movements
Sensing of
Perceived Stress
Self-Report
Interface
Individual
Feedback
Stress
Reduction
Tips
…
Stress
Classification
Module
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 7
Boucsein, 2012, p. 181). To derive the SCR values, we decomposed the electrodermal activity data by
applying nonnegative deconvolution as described by Benedek and Kaernbach (2010).
Then, subjects read a cover story, which introduced a fictive objective of the study, i.e. to measure
subjects’ cognitive performance. All mouse movements were recorded from this point in time by the
corresponding sensing module of the SDS.
After subjects have read the cover story, they were asked to follow the instructions related to a square
task (Square Task 1). This square task was designed to test the hypotheses, i.e. to provide subjects with
pre-defined mouse targets and trajectories. This first square task involved listening to calm music for
one minute while tracing the 550px edges of a square on the screen using the mouse. Subjects were
given the instruction to trace the edges as accurate as possible while maintaining a swift pace. They were
also asked to perform a double click on each corner they pass, which induces a color change of the
corner, a visual feedback for interaction success. Figure 4 (right) shows the visual instructions of this
square task. Due to the characteristics of this square task, we could test the hypotheses regarding hori-
zontal and vertical mouse movements separately in addition to their composite, i.e. the average of the
hypothesized features in both dimensions. An overview of the derived features of mouse movements is
provided in Table 1. Finally, the square task allowed us also to derive an objective performance measure
from the mouse coordinates, i.e. the number of edges an individual could trace within the pre-defined
time. This control variable is used to test the negative impact of stress on task performance as proposed
by the JD-R model, and thus, to add external validity to the experimental procedure and findings.
Figure 4. Study procedure (left) and visual instructions of the square task (right).
Consistent with prior work (Sun et al., 2014), participants were then asked to rate their perceived stress
level on an 11-point Likert scale ranging from 0 (not stressed at all) to 10 (extremely stressed) with the
corresponding sensing module of our SDS (Perceived Stress 1).
After this baseline measurement, stress was induced in the subjects of the experimental group (Stress
Induction), whereas a filler task was given to the subjects of the control group (Filler Task). For the
stress induction, we used techniques related to the Trier Social Stress Test (Kirschbaum et al., 1993).
Accordingly, subjects of the experimental group were asked to prepare a five-minute presentation about
a self-assessment of their cognitive capabilities, which, according to the instructions, would be critically
examined by an academic expert on cognitive performance awaiting their presentation in the room next
door. Furthermore, a video camera was placed beside the computer screen, the recording was started
and subjects were told that the academic expert is observing them from now on.
Participants
Experimental Group Control Group
Randomization
Square Task 1
Mouse Movements
Square Task 1
Mouse Movements
Filler TaskStress Induction
Perceived Stress 1
Self-report
Perceived Stress 1
Self-report
Square Task 2
Mouse Movements
Square Task 2
Mouse Movements
Perceived Stress 2
Self-report
Perceived Stress 2
Self-report
1 min.
-
2 min.
1 min.
-
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 8
After this instruction, subjects had two minutes to prepare the presentation with a pen and paper, thus
no biasing mouse interactions were required. By contrast, subjects of the control group were given the
filler task to write down with a pen calm moments of their last holidays. They were also given two
minutes for this task.
Thereafter, subjects of both groups were asked again to conduct the one-minute square task (Square
Task 2). In addition to the instructions of Square Task 1 both groups were now asked to memorize the
different colors of the corners of the square after each double click together with the number of edges
they successfully traced during the task to increase cognitive load and to counterbalance, to some degree,
any learning effects from Square Task 1. Moreover, instead of relaxation music, both groups were ex-
posed to acoustic office background noise (e.g. talking, phone ringing or and printing sounds from stylus
printers) to create a more realistic acoustic workplace scenario. After Square Task 2 subjects were again
asked to provide their perceived stress level (Perceived Stress 2).
Finally, all subjects were debriefed and the true objective of the study was revealed. Subjects of the
experimental group were additionally asked to perform a relaxation exercise to calm down.
Feature
Description
Mouse ADx/y
Averaged deviation from the horizontal (ADx)
and vertical (ADy) edges of the square in pixels
Mouse ADtotal
Average of ADx and ADy
Mouse ASx/y
Average speed on the horizontal (ASx) and vertical (ASy) mouse
trajectory during the square task in pixels per millisecond
Mouse AStotal
Average of ASx and ASy
Table 1. Derived feature set from mouse movements according to H1 and H2
5.2 Results
Overall, 19 students from a business university participated in the experiment. We had to drop one
subject from further analyses due to incomplete physiological data. The descriptive statistics of the re-
maining 18 subjects are listed in Table 2. Boxplots of stress perceptions after the two square tasks for
each of both groups are shown in Figure 5 (left). In order to test the success of the stress induction, we
conducted a robust mixed ANOVA on the perceived stress data (see Field et al., 2012, p. 648; 10%
trimming, tsplit()). As a prerequisite, Levene’s test revealed that the variances of the perceived stress
(PS) variables are similar for the control and the experimental group (FPS1(1, 17) = 0.60, ns; FPS2(1,17)
= 0.01, ns). The assumption of homogeneity of variance is thus not violated. ANOVA results show that
the group allocation (main effect) is not significant (Q = 0.21, ns), but both the time (main effect of PS1
and PS2) (Q = 6.69, p = .01) and the interaction term of group allocation and time (Q = 6.71, p = .01).
Post-hoc tests revealed that there are no significant time differences in the control group (t(8) = -1.79,
ns) but only in the experimental group (t(8) = -3.49, p < .01). Overall, this indicates that the induction
of stress was successful.
Furthermore, boxplots of physiological arousal measured by the total sum of standardized skin conduct-
ance reactions during the two square tasks for each of both groups are shown in Figure 5 (right). Here,
we found no significant differences in physiological arousal between both groups (Q=.04, p = .85) and
also no interaction effect between group and square task factors (Q=.04, p = .85) by applying a robust
mixed ANOVA with R’s WRS2 package (bwtrim function). That is, physiological arousal did not bias
any neuromotor noise effect between the groups. However, we found a significant difference in physi-
ological arousal between the two square tasks (Q=17.9, p = .002) which can be explained by the increase
of cognitive load in both groups by memorizing the number of traced edges and colors of the square
corners.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 9
Finally, we evaluated the relative performance of both groups based on the number of edges that each
subject traced during the square tasks. Results indicate, though statistically not significant (p > .05), that
the increase in performance from the first to the second square task can be attributed to learning effects.
However, the increase in performance was lower for the experimental group (12.1%) compared to the
control group (22.4%). This finding is consistent with the negative effects of workplace stress as outlined
in the JD-R model and thus, adds external validity to the results of this lab experiment.
Variable
Experimental Group
Control Group
N
9
9
Gender
8 male, 1 female
7 male, 2 female
Age
23.78 (3.31)
26.22 (1.72)
Square Task 1
Performance (Number of Edges)
19.06 (5.88)
22.52 (7.51)
Perceived Stress 1
2.67 (2.18)
2.78 (2.59)
Physiological Arousal (SCR)
994 (327)
977 (237)
Mouse ADx
6.24 (2.26)
6.89 (1.43)
Mouse ADy
5.01 (1.47)
5.60 (1.77)
Mouse ADtotal
5.62 (1.79)
6.24 (1.44)
Mouse ASx
0.17 (0.05)
0.22 (0.06)
Mouse ASy
0.19 (0.05)
0.23 (0.06)
Mouse AStotal
0.18 (0.05)
0.23 (0.06)
Square Task 2
Performance (Number of Edges)
23.33 (6.60)
27.56 (4.93)
Perceived Stress 2
6.11 (2.03)
3.67 (2.06)
Physiological Arousal
1514 (358)
1493 (294)
Mouse ADx
6.58 (3.00)
8.01 (2.90)
Mouse ADy
4.46 (1.33)
5.25 (1.81)
Mouse ADtotal
5.52 (2.13)
6.63 (2.18)
Mouse ASx
0.21 (0.07)
0.26 (0.06)
Mouse ASy
0.22 (0.06)
0.27 (0.04)
Mouse AStotal
0.21 (0.06)
0.27 (0.05)
Table 2. Descriptive Statistics. Note: Means (Std. Dev.) are only provided for quantitative
variables
Against this background, we now focus on the mouse movement features of the experimental group.
Regarding the mouse trajectories and to provide an intuition about potential differences in low vs. highly
stressed participants, Figure 6 visualizes mouse movement patterns of participants with a perceived
stress from the low and the high end of the scale. Here, the differences in the deviation of the mouse
trajectories to the optimal square edges is consistent with the hypothesized effects of van Gemmert and
van Galen’s stress theory about the concept of neuromotor noise and can be visually clearly observed.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 10
Figure 5. Boxplots for perceived stress (left) and physiological arousal (right). Note: “Con”
stands for Control group and “Exp” stands for experimental group
Figure 6. Six mouse movement patterns of subjects with low (top, baseline) and high (bottom,
after stress induction) perceived stress levels.
To test our hypotheses, we first calculated Pearson’s product-moment correlations (Cohen et al., 2003)
of perceived stress and the six mouse movement features for both, the experimental group and the con-
trol group. Table 3 summarizes the correlation coefficients for all six features. In the experimental group,
we found five out of six (83%) positive and significant correlations with large and close to large effect
sizes. By contrast, we found no significant correlations at the .05 level for the same relationships in the
control group. We therefore conclude that these results support our hypotheses.
0
2
4
6
8
After Square Task 1 After Square Task 2
0 (no stress at all) - 10 (extremely stressed)
Group
Con
Exp
Perceived Stress
500
1000
1500
2000
During Square Task 1 During Square Task 2
Total Sum of Standardized Skin Conductance Reactions (SCRs)
Group
Con
Exp
Physiological Arousal
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 11
Pearson Correlation
Experimental Group
Control Group
Mouse ADx x perceived stress
.49* (p = .04)
-.05 (p = .85)
Mouse ADy x perceived stress
.35*(p = .15)
-.10 (p = .68)
Mouse ADtotal x perceived stress
.46* (p < .05)
-.08 (p = .75)
Mouse ASx x perceived stress
.54* (p = .02)
-.06 (p = .80)
Mouse ASy x perceived stress
.58* (p = .01)
-.07 (p = .77)
Mouse AStotal x perceived stress
.57* (p = .01)
-.07 (p = .79)
Table 3. Pearson correlation coefficients of the six mouse movement features and perceived
stress levels. Note: * indicates p < .05
To cross-validate these findings with respect to differences between both square tasks, we further con-
ducted Wilcoxon signed-rank tests for the experimental and control group to compare the medians of
the six mouse movement features. The results are shown in Table 4. They indicate that there are signif-
icant differences in all six features’ medians between Square Task 1 vs. Square Task 2 for the subjects
of the experimental group at the .95% confidence interval. By contrast, we could not identify any sig-
nificant differences for subjects of the control group. Again, these results support H1 and H2.
Wilcoxon Signed-Rank Test
Experimental Group
Control Group
Difference of Mouse ADx between both square tasks
p = .01*
p = .96
Difference of Mouse ADy between both square tasks
p = .05*
p = .89
Difference of Mouse ADtotal between both square tasks
p = .03*
p = 1.0
Difference of Mouse ASx between both square tasks
p = .02*
p = .68
Difference of Mouse ASy between both square tasks
p = .01*
p = .55
Difference of Mouse AStotal between both square tasks
p = .007**
p = .68
Table 4. Difference of the six mouse movement features between low vs. high perceived stress
by applying Wilcoxon signed-rank tests. Note: * / ** indicates p < .05 / .01
To finally test our hypotheses with respect to SDS’s classification module, i.e. to assess the applicability
of the proposed mouse movement features to predict high versus low stress levels, we finally defined
this task as a binary classification problem. Using receiver operating characteristic (ROC) analysis, we
assessed the univariate classification performance of each feature for separating the two square tasks in
both groups. Figure 7 shows the corresponding ROC curves for each of the six mouse movement fea-
tures. Among the area under the curve (AUC), Table 5 lists accuracy, sensitivity and specificity scores
of the optimal operating point on the ROC curve for the experimental group only, because the AUCs for
the control group are close to .50 (see Table 5, column two in brackets) indicating a random and thus,
inadequate classification.
While the AUCs for optimal line deviation features are consistently lower than features derived from
mouse movement speed, performance at the optimal operating point is equal for all proposed features.
The ADy feature and the AStotal feature have the highest sensitivity score with 0.89, trading off specificity
at 0.67. With AUC values starting at .75 and clearly above the random value of .50, and an average
classification accuracy of .78 for both feature sets, i.e. average deviation and average speed, the empir-
ical data supports the proposed hypotheses. It can be therefore concluded that both (a) average deviation
from an optimal mouse trajectory and (b) average speed of mouse movements are potentially relevant
mouse movement features of SDS’s classification module.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 12
Figure 7. Receiver operator curves for the six mouse movement features for experimental group
(Exp) and control group (Con).
Feature
Area Under the Curve
Accuracy
Sensitivity
Specificity
Mouse ADx
0.75 (Control Group 0.49)
0.78
0.78
0.78
Mouse ADy
0.78 (Control Group 0.47)
0.78
0.89
0.67
Mouse ADtotal
0.75 (Control Group 0.50)
0.78
0.78
0.78
Mouse ASx
0.81 (Control Group 0.43)
0.78
0.78
0.78
Mouse ASy
0.84 (Control Group 0.40)
0.78
0.78
0.78
Mouse AStotal
0.86 (Control Group 0.43)
0.78
0.89
0.67
Table 5. Analysis of the receiver operating characteristic. Note: Accuracy, sensitivity and spec-
ificity values are provided for the optimal operating point; AUC: Area under the curve
6 Discussion
This research presents a first steps towards a novel Stress Detection Service (SDS), which has the ob-
jective to unobtrusively measure the degree of work-related stress in knowledge workers solely based
on mouse movements. Relying on the conceptual foundations of van Gemmert and van Galen’s stress
theory and the JD-R model (Bakker and Demerouti, 2007), experimental results indicate that two feature
sets of mouse movements, i.e. average deviation from an optimal mouse trajectory and average mouse
movement speed, can classify high versus low stress with an overall accuracy of 78%. With 78%, the
classification accuracy of our SDS lies also slightly above the accuracy of prior work (Sun et al., 2014,
71%) but requires only one feature to reach that accuracy score. That is, the feature set is more parsimo-
nious compared to the work of Sun et al. as they have adopted the mass-spring-damper model with five
features as justificatory knowledge for their classification algorithm development.
0
0.2
0.4
0.6
0.8
1
False positive rate (1-Specificity)
0
0.2
0.4
0.6
0.8
1
True positive rate (Sensitivity)
Mouse ASx Exp
Random
Mouse ASx Con
0
0.2
0.4
0.6
0.8
1
False positive rate (1-Specificity)
0
0.2
0.4
0.6
0.8
1
True positive rate (Sensitivity)
Mouse ASy Exp
Random
Mouse ASy Con
0
0.2
0.4
0.6
0.8
1
False positive rate (1-Specificity)
0
0.2
0.4
0.6
0.8
1
True positive rate (Sensitivity)
Mouse AStotal Exp
Random
Mouse AStotal Con
0
0.2
0.4
0.6
0.8
1
False positive rate (1-Specificity)
0
0.2
0.4
0.6
0.8
1
True positive rate (Sensitivity)
Mouse ADx Exp
Random
Mouse ADx Con
0
0.2
0.4
0.6
0.8
1
False positive rate (1-Specificity)
0
0.2
0.4
0.6
0.8
1
True positive rate (Sensitivity)
Mouse ADy Exp
Random
Mouse ADy Con
0
0.2
0.4
0.6
0.8
1
False positive rate (1-Specificity)
0
0.2
0.4
0.6
0.8
1
True positive rate (Sensitivity)
Mouse ADtotal Exp
Random
Mouse ADtotal Con
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 13
Moreover, the current work presents an experimental design with a control group compared to Sun et
al. who chose a within-subjects design only. That is, we could not only test differences of repeated
measures but also among the experimental and control groups by our within- and between-subjects de-
sign. The current work complements therefore the work of Sun et al. (2014) and shows for the very first
time that mouse movements are indeed related to stress perceptions but are not necessarily related to
another physiological indicator of arousal and stress, i.e. skin conductance responses, a common view
of previous research (e.g. Chittaro and Sioni, 2014a; Liao et al., 2006; Moody and Galletta, 2015; Riedl,
2013; Riedl et al., 2013; Schnädelbach et al., 2012; Sun et al., 2014). That is, physiological arousal
increased significantly in both groups from Square Task 1 to Square Task 2 but there was only a signif-
icant difference of the mouse movement features in the experimental group (see Table 4). Consistent
with prior work (e.g. Grimes et al., 2013), this indicates that mouse movements might have the potential
to differentiate between positive vs. negative arousal (and stress), a quality physiological arousal meas-
ured by skin conductance reactions alone is not capable of (e.g. Sheng and Joginapelly, 2012).
Another aspect must be pointed out that seems to be relevant for measures in general and thus, also for
physiological mouse movements. There seems to be a discrepancy in the descriptive statistics of Table
2 (e.g. Mouse ADtotal = 5.62 during Task 1 vs. 5.52 during Task 2) and our hypotheses. One rationale of
this observation, among the fact that stress detection via mouse movements does not work for every-
body, may lie in intra-individual differences of subjects (e.g. Kehr and Kowatsch, 2015; Pitariu and
Ployhart, 2010; Ployhart and Vandenberg, 2010; Tams et al., 2014; vom Brocke and Liang, 2014). That
is, individuals’ physiological reactions to externally induced stress may differ systematically, for exam-
ple, with regard to their personality traits (Bibbeya et al., 2013) or the degree they are chronically
stressed (Jones et al., 2016). It is therefore recommended to control for any additional physiological and
psychological factors in future studies and to apply self-learning techniques (e.g. reinforcement learn-
ing) that adapt the feature set and their classification model to each individual knowledge worker over
time (Kaelbling et al., 1996; Szita and Szepesvari, 2010).
Even though this work reports promising results regarding a SDS, two major limitations of the current
work are the artificial square task and the limited study sample from which general findings cannot be
drawn, particularly, with respect to field settings. Moreover, the binary classification approach adopted
in this work yielded accurate results, but this demonstrates only a first step towards more complex mod-
els. That is, a combination of features can act as a base for decision models trained by machine learning
algorithms to go beyond the binary classification case and to further increase performance levels.
7 Future Work
In our future work, we use the findings of the current work to adapt the current SDS to a field setting
and asses its technical, organizational and legal feasibility. In line with additional design requirements
derived from knowledge workers of one of our industry partners, we will combine all current SDS mod-
ules into one integrated desktop application and compare detection accuracies with prior work.
Moreover, we will assess the accuracy of SDS’s stress detection module with respect to multi-class
classification problems (e.g. low, medium and high stress). We will also add a quantified-self module
that allows self-reflection of perceived stress data in the form of a visual diary. This quantified-self
module aims not only at improving self-determined stress reflection and regulation capabilities of em-
ployees but it has also the goal to motivate knowledge workers to continuously train their individual
classification model such that the obtrusive and time-consuming self-report activities can be replaced
by mouse movements in the long term.
We finally plan to integrate and evaluate existing interventions for the management of stress at the
workplace (Ryan et al., 2017). Examples are relaxation exercises (Chittaro and Sioni, 2014a), breathing
trainings (Chittaro and Sioni, 2014b), health literacy interventions (Jacobs et al., 2016) or job crafting
interventions (Kooij et al., 2017) that are triggered just-in-time by our envisioned SDS with the overall
goal to prevent chronic stress and serious mental health problems such as major depression.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 14
References
Astor, P. J., M. T. P. Adam, P. Jercic, K. Schaaff and C. Weinhardt (2013) Integrating Biosignals into
Information Systems: A NeuroIS Tool for Improving Emotion Regulation, Journal of Management
Information Systems 30 (3), 247-277.
Bakker, A. B. and E. Demerouti (2007) The Job Demands-Resources Model: State of the Art, Journal
of Managerial Psychology 22 (3), 309-328.
Barata, F., T. Kowatsch, P. Tinschert and A. Filler (2016) Personal MobileCoach: Tailoring Behavioral
Interventions to the Needs of Individual Participants, in: Proceedings of the UBICOMP 2016
Workshop Designing, Developing, and Evaluating The Internet of Personal Health (IoPH),
Heidelberg, Germany.
Benedek, M. and C. Kaernbach (2010) Decomposition of skin conductance data by means of
nonnegative deconvolution, Psychophysiology 47, 647-658.
Bibbeya, A., D. Carrolla, T. J. Roseboomb, A. C. Phillipsa and S. R. de Rooijb (2013) Personality and
physiological reactions to acute psychological stress, International Journal of Psychophysiology 90
(1), 28-36.
Boedeker, W. and H. Klindworth (2007) Hearts and Minds at Work in Europe: A European Work-
related Public Health Report on Cardiovascular Diseases and Mental Ill Health, Essen: BKK
Bundesverband.
Boucsein, W. (2012) Electrodermal Activity, (2nd ed.) Heidelberg, Germany: Springer.
Cartwright, S. and C. L. Cooper (2014), Towards Organizational Health: Stress, Positive Organizational
Behavior, and Employee Well-being, in: Bridging Occupational, Organizational and Public Health:
A Transdisciplinary Approach, G. F. Bauer and O. Hämmig (eds.), Springer, Dordrecht, 29-42.
Chittaro, L. and R. Sioni (2014a) Affective computing vs. affective placebo: Study of a biofeedback-
controlled game for relaxation training, International Journal of Human-Computer Studies 72 (8–9),
663-673.
Chittaro, L. and R. Sioni (2014b) Evaluating mobile apps for breathing training: The effectiveness of
visualization, Computers in Human Behavior 40, 56-63.
Cohen, J., P. Cohen, S. G. West and L. S. Aiken (2003) Applied multiple regression/correlation analysis
for the behavioral sciences, Mahwah, NJ, USA: Lawrence Erlbaum Associates Publishers.
Cohen, S., T. Kamarck and R. Mermelstein (1983) A Global Measure of Perceived Stress, Journal of
Health and Social Behavior 24 (4), 385-396.
Cook, R. F., D. W. Billings, R. K. Hersch, A. S. Back and A. Hendrickson (2007) A Field Test of a
Web-based Workplace Health Promotion Program to Improve Dietary Practices, Reduce Stress, and
Increase Physical Activity: Randomized Controlled Trial, Journal of Medical Internet Research 9
(2:e17).
Demerouti, E., A. B. Bakker, I. Vardakou and A. Kantas (2003) The convergent validity of two burnout
instruments: A multitrait- multimethod analysis, European Journal of Psychological Assessment 19
(1), 12-23.
Demerouti, E., P. M. Le Blanc, A. B. Bakker, W. B. Schaufeli and J. Hox (2009) Present but sick: a
three-wave study on job demands, presenteeism and burnout, Career Development International 14
(1), 50-68.
Drucker, P. F. (1999) Knowledge-Worker Productivity: The Biggest Challange, California
Managament Review 41 (2), 79-94.
Eurofound (2012) Fifth European Working Conditions Survey, Publications Office of the European
Union, Luxembourg.
Field, A., J. Miles and Z. Field (2012) Discovering Statistics Using R, (1st ed.) London: Sage.
Freeman, J. B. and N. Ambady (2010) MouseTracker: Software for studying real-time mental processing
using a computer mouse-tracking method, Behavior Research Methods 42 (1), 226-241.
Freeman, J. B., R. Dale and T. A. Farmer (2011) Hand in motion reveals mind in motion, Frontiers in
Psychology 2 (59), 1-6.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 15
Ganster, D. C. and C. C. Rosen (2013) Work Stress and Employee Health: A Multidisciplinary Review,
Journal of Management 39 (5), 1085-1122.
Geurts, S. (2014), Recovery from Work during Off-job Time, in: Bridging Occupational,
Organizational and Public Health: A Transdisciplinary Approach, G. F. Bauer and O. Hämmig
(eds.), Springer, Dordrecht, 193-208.
Gregor, S. and A. Hevner (2013) Positioning and Presenting Design Science Research for Maximum
Impact, MIS Quarterly 37 (2), 337-355.
Grimes, G. M., J. L. Jenkins and J. S. Valacich (2013) Exploring the Effect of Arousal and Valence on
Mouse Interaction, in: Proceedings of the International Conference on Information Systems (ICIS),
Milan, Italy.
Hevner, A. R., S. T. March, J. Park and S. Ram (2004) Design Science in Information Systems Research,
MIS Quarterly 28 (1), 75-105.
Hibbeln, M., J. L. Jenkins, C. Schneider, J. S. Valacich and M. Weinmann (2017) How Is Your User
Feeling? Inferring Emotion Through Human-Computer interaction Devices, MIS Quarterly 41 (1),
1-21.
Igic, I., B. Brunner, A. Elfering and S. Wieser (2016) Job-Stress-Index 2016, Health Promotion
Switzerland, Bern, Switzerland.
Jacobs, R. J., J. Q. Lou, R. L. Ownby and J. Caballero (2016) A systematic review of eHealth
interventions to improve health literacy, Health Informatics Journal 22 (2), 81-98.
Jones, A., J. C. Pruessner, M. R. McMillan, R. W. Jones, G. T. Kowalik, J. A. Steeden, B. Williams, A.
M. Taylor and V. Muthurangu (2016) Physiological adaptations to chronic stress in healthy humans
– why might the sexes have evolved different energy utilisation strategies?, The Journal of
Physiology 594 (15), 4297-4307.
Kaelbling, L. P., M. L. Littman and A. W. Moore (1996) Reinforcement Learning: A Survey, Journal
of Artificial Intelligence Research 4, 237-285.
Kehr, F. and T. Kowatsch (2015) Quantitative Longitudinal Research: A Review of IS Literature, and a
Set of Methodological Guidelines, in: Proceedings of the 23rd European Conference on Information
Systems (ECIS), Münster, Germany, 2015, Münster, Germany.
Kessler, R. C., C. Barber, A. L. Beck, P. A. Berglund, P. D. Cleary, D. McKenas, N. P. Pronk, G. E.
Simon, P. E. Stang, T. B. Üstün, et al. (2003) The World Health Organization Health and Work
Performance Questionnaire (HPQ), Journal of Occupational and Environmental Medicine 45 (2),
156-174.
Kirschbaum, C., K. M. Pirke and D. H. Hellhammer (1993) The 'Trier Social Stress Test'--a tool for
investigating psychobiological stress responses in a laboratory setting, Neuropsychobiology 28 (1-
2), 76-81.
Koldijk, S., M. Sappelli, M. Neerincx and W. Kraaij (2013), Unobtrusive Monitoring of Knowledge
Workers for Stress Self-regulation, in: Proc. of the 21th International Conference on User Modeling,
Adaptation, and Personalization (UMAP '13), S. Carberry, S. Weibelzahl, A. Micarelli and G.
Semeraro (eds.), Springer, Berlin, Germany, 335-337.
Kooij, D. T. A. M., M. van Woerkom, J. Wilkenloh, L. Dorenbosch and J. J. A. Denissen (2017) Job
crafting towards strengths and interests: The effects of a job crafting intervention on person-job fit
and the role of age, The Journal of Applied Psychology 102 (6), 971-981.
Kowatsch, T., F. Wahle, A. Filler, F. Kehr, D. Volland, S. Haug, G. J. Jenny, G. Bauer and E. Fleisch
(2015) Towards Short-Term Detection of Job Strain in Knowledge Workers with a Minimal-invasive
Information System Service: Theoretical Foundation and Experimental Design, in: Proceedings of
the 23rd European Conference on Information Systems (ECIS), Münster, Germany.
Kuoppala, J., A. Lamminpää and P. Husman (2008) Work Health Promotion, Job Well-being, and
Sickness Absences - A Systematic Review and Meta-Analysis, Journal of Occcupational
Environmental Medicine 50 (11), 1216-1227.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 16
Liao, W., W. Zhang, Z. Zhu, Q. Ji and W. D. Gray (2006) Toward a decision-theoretic framework for
affect recognition and user assistance, International Journal of Human-Computer Studies 64 (9),
847-873.
Lin, C.-J. and C. Wu (2011) Factors affecting numerical typing performance of young adults in a hear-
and-type task, Ergonomics 54 (12), 1159-1174.
Maehr, W. (2008) eMotion: Estimation of User's Emotional State by Mouse Motions, Saarbrücken,
Germany: VDM.
Mattke, S., H. Liu, J. P. Caloyeras, C. Y. Huang, K. R. Van Busum, D. Khodyakov and V. Shier (2013)
Workplace Wellness Programs Study, Santa Monica, CA, USA: RAND Corporation.
Michie, S. (2002) Causes and Management of Stress at Work, Occupational and Environmental
Medicine 59 (1), 67-72.
Moody, G. D. and D. F. Galletta (2015) Lost in Cyberspace: The Impact of Information Scent and Time
Constraints on Stress, Performance, and Attitudes Online, Journal of Management Information
Systems 32 (1), 192-224.
Nahum-Shani, I., E. B. Hekler and D. Spruijt-Metz (2015) Building Health Behavior Models to Guide
the Development of Just-in-Time Adaptive Interventions: A Pragmatic Framework, Health
Psychology 34 (Supplement), 1209-1219.
Nahum-Shani, I., S. N. Smith, B. J. Spring, L. M. Collins, K. Witkiewitz, A. Tewari and S. A. Murphy
(2016) Just-in-Time Adaptive Interventions (JITAIs) in Mobile Health: Key Components and Design
Principles for Ongoing Health Behavior Support, Annals of Behavioral Medicine ePrint ahead
doi:10.1007/s12160-016-9830-8.
Nyberg, S. T., E. I. Fransson, K. Heikkilä, K. Ahola, L. Alfredsson, J. B. Bjorner, M. Borritz, H. Burr,
N. Dragano, M. Goldberg, et al. (2014) Job Strain as a Risk Factor for Type 2 Diabetes: A Pooled
Analysis of 124,808 Men and Women, Diabetes Care 37 (8), 2268-2275.
Peffers, K., T. Tuunanen, M. Rothenberger and S. Chatterjee (2007) A Design Science Research
Methodology for Information Systems Research, Journal of Management Information Systems 24
(3), 45-77.
Pejovic, V. and M. Musolesi (2015) Anticipatory Mobile Computing: A Survey of the State of the Art
and Research Challenges, ACM Computing Surveys 47 (3), Article 47.
Pitariu, A. H. and R. E. Ployhart (2010) Explaining Change: Theorizing and Testing Dynamic Mediated
Longitudinal Relationships, Journal of Management 36 (2), 405-429.
Ployhart, R. E. and R. J. Vandenberg (2010) Longitudinal Research: The Theory, Design, and Analysis
of Change, Journal of Management 36 (1), 94-120.
Riedl, R. (2013) On the biology of technostress: literature review and research agenda, The DATA BASE
for Advances in Information Systems 44 (1), 18-55.
Riedl, R., H. Kindermann, A. Auinger and A. Javor (2013) Computer breakdown as a stress factor during
task completion under time pressure: Identifying gender differences based on skin conductance,
Advances in Human-Computer Interaction Article ID 420169, 1-8.
Ryan, C., M. Bergin, T. Chalder and J. S. Wells (2017) Web-based interventions for the management of
stress in the workplace: Focus, form, and efficacy., Journal of Occupational health 59 (3), 215-236.
Sappelli, M., S. Verberne, S. Koldijk and W. Kraaij (2014) Collecting a dataset of information behaviour
in context, in: Proceedings of the 4th Workshop on Context-Awareness in Retrieval and
Recommendation, Amsterdam, The Netherlands, 26-29.
Schnädelbach, H., A. Irune, D. Kirk, K. Glover and P. Brundell (2012) ExoBuilding: Physiologically
Driven Adaptive Architecture, ACM Trans. Comput.-Hum. Interact. 19 (4), 1-22.
Sheng, H. and T. Joginapelly (2012) Effects of Web Atmospheric Cues on Users’ Emotional Responses
in E-Commerce, AIS Transactions on Human-Computer Interaction 4 (1), 1-24.
Siegrist, J., N. Wege, F. Pühlhofer and M. Wahrendorf (2009) A short generic measure of work stress
in the era of globalization: effort-reward imbalance, International Archives of Occupational or
Environmental Health 82 (8), 1005-1013.
Kowatsch et al. / Mouse-based Stress Detection Service
The 11th Mediterranean Conference on Information Systems (MCIS), Genoa, Italy, 2017 17
Sun, D., P. Paredes and J. Canny (2014) MouStress: detecting stress from mouse motion, in:
Proceedings of the CHI '14 Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems, Toronto, Canada, 61-70.
Szita, I. and C. Szepesvari (2010) Model-based reinforcement learning with nearly tight exploration
complexity bounds, in: Proceedings of the Proceedings of the 27th International Conference on
Machine Learning, Haifa, Israel, 1-8.
Tams, S., K. Hill, A. O. de Guinea, J. Thatcher and V. Grover (2014) NeuroIS - Alternative or
Complement to Existing Methods? Illustrating the Holistic Effects of Neuroscience and Self-
Reported Data in the Context of Technostress Research, Journal of the Association for Information
Systems 15 (10), Article 1.
Tims, M., A. B. Bakker and D. Derks (2013a) The Impact of Job Crafting on Job Demands, Job
Resources, and Well-Being, Journal of Occupational Health Psychology 18 (2), 230-240.
Tims, M., A. B. Bakker, D. Derks and W. van Rhenen (2013b) Job Crafting at the Team and Individual
Level: Implications for Work Engagement and Performance, Group & Organization Management
38 (4), 427-454.
van Gemmert, A. W. A. and G. P. van Galen (1997) Stress, neuromotor noise, and human performance:
A theoretical perspective, Journal of Experimental Psychology: Human Perception and Performance
23 (5), 1299-1313.
Visser, B., M. P. D. Looze, M. P. D. Graaff and J. H. V. Dieën (2004) Effects of precision demands and
mental pressure on muscle activation and hand forces in computer mouse tasks, Ergonomics 47 (2),
202-217.
vom Brocke, J. and T.-P. Liang (2014) Guidelines for Neuroscience Studies in Information Systems
Research, Journal of Management Information Systems 30 (4), 211-234.
vom Brocke, J., R. Riedl and P.-M. Léger (2013) Application Strategies for Neuroscience in Information
Systems Design Science Research, Journal of Computer Information Systems 53 (3), 1-13.
WHO (2013) Mental health action plan 2013-2020, World Health Organization, Geneva, Switzerland.
Zimmermann, P., S. Guttormsen, B. Danuser and P. Gomez (2003) Affective computing - a rationale
for measuring mood with mouse and keyboard, International Journal of Occupational Safety and
Ergonomics 9 (4), 539-551.
