Machine Learning Findings on Geospatial Data of Users from the
TrackYourStress mHealth Crowdsensing Platform
R¨
udiger Pryss1, Dennis John2, Manfred Reichert1, Burkhard Hoppenstedt1, Lukas Schmid1, Winfried Schlee3,
Myra Spiliopoulou4, Johannes Schobel1, Robin Kraft1, Marc Schickler1, Berthold Langguth3, and Thomas Probst5
Abstract— Mobile apps are increasingly utilized to gather
data for various healthcare aspects. Furthermore, mobile apps
are used to administer interventions (e.g., breathing exercises)
to individuals. In this context, mobile crowdsensing constitutes
a technology, which is used to gather valuable medical data
based on the power of the crowd and the offered computational
capabilities of mobile devices. Notably, collecting data with
mobile crowdsensing solutions has several advantages compared
to traditional assessment methods when gathering data over
time. For example, data is gathered with high ecological validity,
since smartphones can be unobtrusively used in everyday life.
Existing approaches have shown that based on these advantages
new medical insights, for example, for the tinnitus disease, can
be revealed. In the work at hand, data of a developed mHealth
crowdsensing platform that assesses the stress level and fluctu-
ations of the platform users in daily life was investigated. More
specifically, data of 1797 daily measurements on GPS and stress-
related data in 77 users were analyzed. Using this data source,
machine learning algorithms have been applied with the goal
to predict stress-related parameters based on the GPS data
of the platform users. Results show that predictions become
possible that (1) enable meaningful interpretations as well as
(2) indicate the directions for further investigations. In essence,
the findings revealed first insights into the stress situation of
individuals over time in order to improve their quality of life.
Altogether, the work at hand shows that mobile crowdsensing
can be valuably utilized in the context of stress on one hand.
On the other, machine learning algorithms are able to utilize
geospatial data of stress measurements that was gathered by a
crowdsensing platform with the goal to improve the quality of
life of its participating crowd users.
Index Terms— mHealth, crowdsensing, stress, geospatial
data, machine learning
I. INTRODUCTION
The present life is increasingly characterized by psychoso-
cial stress, which is a risk factor for mental and somatic
diseases [1]. As a result, interdisciplinary teams try to find
technical solutions that may empower patients to better man-
age their stress-level over time [2], [3]. Mobile crowdsensing
that incorporates Ecological Momentary Assessments (EMA)
1Faculty of Computer Science, Engineering and Psychology, Ulm Uni-
versity, Germany
{ruediger.pryss,burkhard.hoppenstedt,lukas.schmid,
manfred.reichert,johannes.schobel,
robin.kraft,marc.schickler}@uni-ulm.de
2Lutheran University of Applied Sciences Nuremberg, Germany
3University Hospital Regensburg, Germany
4University of Magdeburg, Germany [email protected]
5Department for Psychotherapy and Biopsychosocial Health, Danube
is such a technology. EMA measurements, in turn, are
repeatedly performed in daily life to capture the experience
of a user in the best natural way to maximize the ecological
validity of the gathered data. The concept is well-known
in psychological research and has an already considerable
history [4], [5]. Following this, mobile crowdsensing has sev-
eral advantages compared to traditional assessment methods
when gathering data over time.
First, data is gathered with high ecological validity, since
smartphones can be unobtrusively used in everyday life.
Second, gathered data can be directly compared to data of
other crowd users in a meaningful way. Third, individuals can
collect data more easily compared to a traditional collection
procedure (e.g., paper and pencil questionnaires). Fourth, the
bias of experimenters is mitigated in crowdsensing collection
procedures compared to data collection in the lab as experi-
menters are not needed.
Technically, several sensing paradigms have been pro-
posed to relate crowd users to sensing tasks [6]. Furthermore,
it has been shown that sophisticated technical crowdsensing
platform must be developed to enable meaningful mea-
surements by crowd users [7]. In the context of mobile
Health (mHealth) solutions, four technical developments are
particularly important for the provision of a crowdsensing
platform that incorporates EMA measurements:
1) A proper data model must be conceived [8].
2) A flexible Application Programming Interface, which han-
dles data transfers based on the data model, must be imple-
mented [8].
3) An architecture must be conceived that enables sophisticated
crowdsensing collection procedures [9].
4) Mobile native apps must be implemented that enable (1) an
user experience that is welcomed by the crowd users [10],
[11] as well as (2) enables reliable sensor measurements.
Along these issues, including the use of results from
the implementation of four other crowdsensing platforms
[12], the TrackYourStress mHealth crowdsensing platform
(TYS) was implemented. The goal of TYS is to improve the
quality of life of its users by tracking their stress level and
related variables (e.g., stress reactivity, coping skills) over
time and providing personalized feedback. The daily tracking
procedure of TYS, in turn, comprises specific questions to
assess stress perception and stress-related parameters during
the daily routine of a user. Additionally, the smart mobile
device of a user can record the environmental sound level
and the GPS position while the user fills in the assessment
questionnaires, if he/she agrees in the app that these addi-
tional variables are recorded. The platform is now running for
roughly one year and has already gathered valuable stress-
related data of its registered users.
In the work at hand, the TYS data source was evaluated
with the goal to get insights on the gathered GPS data and
its relation to the collected daily stress values. As machine
learning promises to be useful in this context [13], respective
algorithms are applied to gathered data of TYS to find new
insights. To be more precise, it was investigated whether
geospatial data (i.e., the collected GPS positions) of TYS
users can be used as a predictor for the individual stress
situation of the participating users. The results that have been
revealed show their value and that they can be a starting
point for several other investigations. Furthermore, the results
show that a combination of EMA, mobile crowdsensing, and
machine learning is a powerful setting for collecting and
evaluating ecological valid stress data of crowd users over
time.
The remainder of this paper is organized as follows. In
Section II, related work, which is relevant in the context of
this paper, will be reviewed. Section III gives the background
information on the TrackYourStress platform. The used ma-
terials and methods for the data analysis are described in
Section IV, while Section V presents the obtained results.
The latter are then discussed in Section V, whereas Section
VII concludes the paper with a summary and an outlook.
II. RELATED WORK
Four categories of related work are relevant in the context
of this work. (1) Approaches that deal with mobile crowd-
sensing and EMA in the healthcare domain, (2) approaches
on stress measurement based on mobile technology in gen-
eral, (3) approaches that link GPS data with health aspects,
and (4) approaches that apply machine learning algorithms
to GPS-based data.
Regarding the first category, some recent works exist that
deal with generic crowdsensing approaches to enable human-
subject studies [9]. However, the use of crowdsensing in
the context of stress and related disorders is still rare. One
example for a crowdsensing solution dealing with a chronic
disorder that can be influenced by stress is the TrackY-
ourTinnitus mHealth crowdsensing platform [14]–[17]. In
turn, technical solutions that enable Ecological Momentary
Assessments without using mobile crowdsensing technology
have been already presented with valuable healthcare results
[18], [19].
Regarding the second category, approaches can be found
that utilize mobile technology in the context of stress mea-
surements. A recent review on smartphone-based self-reports
on stress provides an excellent overview [2]. Besides self-
reports on stress levels, objective data was collected in some
of these studies as well. Thereby, associations between stress-
levels and objective data were investigated. For example:
Stress-levels were predicted with accuracies of 60% and 71%
by accelerometer data in the work shown by [20].
Regarding the third category, GPS data is another category
of objective data that was considered and studies found
that stress-levels can be generally predicted by GPS data
[21], [22]. Advantages of GPS measurements to study health
aspects in general have been highlighted by [23], and a recent
review summarizes the studies on smartphone-based passive
sensing (including GPS) in the health context [24].
Regarding the fourth category, there exist approaches
that consider machine learning algorithms in the context
of GPS-based data. One important aspect is to identify the
features for the input parameters of the algorithms [25].
Based on valuable features, many works exist that investigate
mobility or travel behavior [26], [27]. Other approaches try
to figure out anomalies in GPS trajectories [28]. One study,
for example, used machine learning to link GPS-data with
the behavior of cows [29]. For machine learning in the
context of personal sensing and mental health in humans,
this review provides a good overview: [30]. Stress predictions
in the light of smartphone gathered data are also subject of
existing works [31], [32]. However, these approaches do not
particularly focus on stress predictions based on GPS-driven
data that was gathered by a mHealth crowdsensing platform.
Altogether, the presented approaches show that mobile
technology, Ecological Momentary Assessments, mobile
crowdsensing, machine learning, and GPS data are a solid
basis to address healthcare scenarios, but this discipline is
still in its infancy.
III. TRACKYOURSTRESS PLATFORM
TrackYourStress (TYS) is a mHealth crowdsensing plat-
form, which is built on four technical components. First, it
offers a website for user registration and other user-related
features (e.g., account management). Second, it offers an
Android and iOS application. Third, a MariaDB database is
used as the central repository for the data collected. Fourth, a
RESTful API is provided that enables the communication be-
tween the mobile applications, the website, and the database
[8].
In general, TYS was developed to track the individual
stress of its registered users. The tracking is based on a set
of questionnaires (registration, daily, weekly, monthly). In
addition, the environmental sound level and the GPS position
can be measured. Note that the users must allow the sensor
measurement when registering to TYS for the first time. This
way, we consider the privacy of the users [33]. In general,
TYS users accomplish three fundamental phases. First, they
have to register through the website or the mobile apps.
Second, they have to fill in a so-called registration ques-
tionnaire once. The latter captures the current stress situation,
demographic data (e.g., date of birth), and other stress-related
parameters. The completion of this registration questionnaire
is a fundamental prerequisite for users who want to use the
features of the continuous mobile crowdsensing procedure,
i.e., the daily, weekly, and monthly measurements. Also,
during the second phase, users have to accept or adjust a
notification schema. The notification schema determines how
often and in what way (i.e., fixed or random points in time)
the daily, weekly, and monthly assessment questionnaires are
applied to them. The number of daily assessments, in turn,
is restricted to a maximum of 20 times per day.
Third, after the registration questionnaire has been ac-
complished and the notification schemas are determined,
users can start with the daily, weekly, and monthly stress
assessments. For the application of the daily assessment
questionnaire, as well as the weekly and monthly ques-
tionnaire, notification features for both Android and iOS,
as well as a notification algorithm, were realized. After
a notification appears, the user may click on it. In the
latter case, the mobile application is started (if not already
running) and the respective assessment questionnaire (daily,
weekly, or monthly) is directly displayed to the user. Then,
he or she can fill in the questionnaire, and finally saves
the entered data. It is also possible that users can fill in
the questionnaires without notifications. While filling in the
questionnaire, either with or without using a notification,
the GPS position and the environmental sound level are
measured (if the app is allowed to measure them). The result
is then transferred to the database through the RESTful API
if the mobile app is online; otherwise the result is locally
stored until the device gets an online connection. A more
detailed technical description of the presented features can
be found in [8], [17].
IV. MATERIAL AND METHODS
This section provides relevant materials and methods,
which were the basis for this work. First of all, so far,
TrackYourStress (TYS) is only available in German, while
an English version is under development. Furthermore, the
developed website is the only component that is currently
public. The developed mobile apps are only available through
private invitations, i.e., distributing the app without using the
official App stores. For Android, we distribute APK files to
the users, while for iOS, we use Testflight. However, we
plan to release the mobile apps within this year to the Google
Playstore and the Apple App Store. The users that have been
invited and registered are recruited by students of the FOM
University of Applied Sciences in Munich and Augsburg,
Germany.
At the time of this analysis (January 2019), TrackY-
ourStress (TYS) had 119 registered users with 2256 filled out
daily assessment questionnaires. For the analysis of the work
at hand, we only used the daily assessment questionnaire
results. Its 9 questions are shown in Table I. It is noteworthy
that we use sliders as visual analogue scales (VAS), cat-
egories for Question 5 (C), and self-assessment manikins
(SAM; [34]). As the first data preparation step, we filtered
out all test users and all users without any GPS measurement.
104 users with 1879 daily measurements (including results to
all of the 9 questions and a measured GPS position, which is
related to the filled-out questions) remained after this phase.
Furthermore, we analyzed only those users, who had at least
>10 daily measurements (including the GPS positions).
After that, we had 77 users with 1797 daily measurements
and their related GPS positions.
As a next and very important step, we derived geospatial
features as the input for the machine learning algorithms.
The identification of such features in general is challenging
when applying machine learning algorithms. As for GPS-
driven data less works exist that have presented such features
[25], especially in the context of the work at hand, the
derivation of respective features is a challenging task. In
total, we derived 4 geospatial features for our analysis. These
features were then used as the predictors of eight selected
questions of the daily assessment questionnaire (see Table
I). Note that Question 5 (What stresses you at the moment?)
is not included in the analysis. Since it is the only question,
which has no numeric result values, it was omitted in this
analysis, but will be considered in future investigations.
For a better understanding of the used geospatial features,
we provide the considerations behind them and then we
present them more technically. The major idea is to identify
movement patterns of TYS users. Apparently, location clus-
ters may indicate striking whereabouts. Therefore, a cluster
feature is used to reveal the locations, in which the TYS users
are frequently remaining. Such clusters may indicate insights
to stress-related parameters while being there. If valuable
insights can be found, then they can be possibly predicted
and users might be empowered to better cope with the
stressful situations. Such clusters denote our first geospatial
feature. We derive these clusters based on a threshold for
the amount of GPS measurements within a specified location
radius.
The remaining three geospatial features are related to
movement patterns based on daily habits. The second feature
shall capture if TYS users reveal large and different distances
between many GPS measurement points. In this case, a TYS
user might have less location clusters compared to other
TYS users, and, in general, travels a lot. The third feature
shall reveal whether TYS users have always a large distance
between GPS measurements and, in addition, whether these
distances reveal to be in strikingly similar orders of magni-
tude. In other words, this feature expresses that TYS users
have large movement distances between GPS measurements,
but usually with a comparable distance between these GPS
measurement points. The last feature was used to detect
clusters, which cannot be considered as a normal location
cluster. For example, if a user travels to another country and
a cluster is identified there, this cluster shall be distinguished
to those that have been identified on daily habits.
In the following, the four geospatial features will be
technically described: First, we established the so-called
location cluster for each user in order to be able to check
whether a GPS measurement of this user is within this cluster
{0,1}. To create a respective location cluster feature, we
applied DBSCAN [35] to the measured GPS positions of
a user, with a resolution of 0.1, which denotes 0.1◦in
the haversine distance (i.e., roughly 11.12km). Regarding
DBSCAN, this leads to parameter values eps = 0.1and
min samples = 3. That means, a maximal distance of 0.1
to cluster neighbors and minimally 3 nodes are needed to
form a cluster. This feature would require extra calculations
if a GPS measurement would be too close to the poles, but in
our case, this cannot be happen. Following this, we receive a
feature with a binary decision whether a GPS measurement
is located within the cluster or not. The location cluster shall
reflect a region, in which users often stay to detect outliers.
Second, a feature denoted with absolute distance was
derived. For this feature, for each GPS measurement of a
user, the sum of all distances to all other GPS measurements
is calculated. This value is then divided through the amount
of all GPS measurements of a user. This feature shall reflect
large distances between all measured GPS positions of a
user. If this value is thus high, the GPS measurements
among each other reveal large distances, meaning that the
user has less location clusters. More precisely, for a GPS
measurement piof a user with its latitude and longitude
(xi, yi), this feature diis calculated as follows:
di:=
n
P
j=1
||pi−pj||2
n−1=
n
P
j=1
√(xi−xj)2+(yi−yj)2
n−1
Third, a feature denoted with relative distance was
derived. Here, again, for each GPS measurement of a user,
the sum of all distances to all other GPS measurements is
calculated. Opposed to the latter feature, here, the calculated
value is divided through the value with the largest distance.
This shall reflect the situation that a GPS measurement is at
the boundaries of a users’ movement radius (i.e., the relative
distance is high). More precisely, for a GPS measurement
piof a user with its latitude and longitude (xi, yi), this
feature riuses diand is calculated as follows:
ri:= di
max
i
di.
Fourth, a feature denoted with distance to central point
was derived. For this feature, the central point of a user
is calculated through the mean value of all GPS positions.
After that, for each GPS measurement, its distance to the
central point is considered. This feature shall reflect outliers
in the following sense: You do a business trip to another
country, which is also identified as a location cluster, but
the latter constitutes not a daily location cluster of the user.
Such alien clusters can then be considered as an outlier
location cluster. More precisely, for a GPS measurement pi
of a user with its latitude and longitude (xi, yi), this feature
ciis calculated as follows:
ci:= ||pi−c||2=q(xi−c1)2+ (yi−c2)2,
whereby c= (c1, c2)denotes the central point of the
respective user.
Using machine learning algorithms, these four features
were investigated as predictors for the answers of 8 TYS
questions of the daily assessment questionnaire. Then, with
respect to the amount of data and a better insight into the data
set, the value ranges of the 8 TYS questions were divided
into 4 equally distributed quartiles, i.e., very low (0 −25%)
Question Scale
1
How high is your momentary stress-level? VAS
2
How well can you control your momentary stress-level? VAS
3
How strongly are you experiencing your momentary
stress-level as negative/impairing?
VAS
4
How strongly are you experiencing your momentary
stress-level as positive/beneficial?
VAS
5
What stresses you at the moment? C
6
How is your mood right now? SAM
7
How is your arousal right now? SAM
8
How important is the current situation for you person-
ally?
VAS
9
How would you assess your ability to cope with the
currently experienced situation?
VAS
VAS=Visual Analogue Scale, C=Categories
SAM=Self-Assessment Manikins
TABLE I: TrackYourStress Daily Assessment Questions
(Quartile 0), low (25−50%) (Quartile 1), higher (50−75%)
(Quartile 2), and very high (75 −100%) (Quartile 3). Note
that very low means the lowest stress level, while the others
mean an increased stress level up to very high. Thus, we
predict for each of the 8 TYS questions these 4 quartiles,
resulting in (8x4 = 32) prediction values.
For the quartiles, in turn, the respective predictions are
based on the given assessments, while the user is reflected
through the 4 geospatial input features. Following this set-
ting, we can better compare different stress-levels and their
obtained prediction accuracies within one question. After
creating the features and the 4 quartiles for each TYS
question, we used the input for a Decision Tree (DT) and
a Support Vector Machine (SVM). For both methods, the
distribution between training and validation set was defined
as follows: For each of the four quartiles of each question,
100 random values were used for testing, while the rest of the
values was used for the validation. The two machine learning
algorithms have been chosen as they particularly address
high-dimensional datasets, which is the case for the given
dataset on TYS. Further note that the prediction models have
been validated based on a 10-fold cross-validation approach.
Finally note, by the best of our knowledge, no other works
could be found that have presented respective features in
the context of stress predictions. Our feature set was mainly
derived through interdisciplinary discussions with psychol-
ogists, medical experts, and computer scientists. However,
they are only a first step, and therefore be used as the basis
for further investigations.
V. RESULTS
Of the 77 participants, 46 (59.7%) were female. On
average, the participants were M= 33.95 years old (SD =
11.99). Table II shows the prediction accuracies for the
quartiles of each of the 8 analyzed daily TYS questions.
Furthermore, Table II shows in the row denoted with QT*,
how many GPS measurements could be evaluated in each
quartile. We exemplarily describe the meaning of one se-
lected accuracy results for a better understanding. Consider
therefore the value of Quartile 1 and Question 1 computed
by the SVM. The achieved accuracy of 72.8% means that
if a user has given an answer to Question 1 that would put
him or her in the stress level Quartile 1, then we are able to
predict this quartile with an accuracy of 72.8%, just based
on his or her given geospatial data. The results of Question
1 for the SVM are also shown in Fig. 1
Overall, it can be seen that there is a large variance in the
prediction accuracies among the four quartiles, as well as
the two machine learning algorithms. We discuss only three
selected and important results in the following. Although the
overall prediction accuracies are not high (highest achieved
accuracy is 78.5% for the Question6-Quartile3 combination;
using the SVM), valuable indicators can be nevertheless
obtained: (1) Some questions, for example, Questions 8
(How important is the current situation for you personally?),
have revealed a higher accuracy when using the DT over all
four quartiles than other questions (e.g., when comparing it
to Question 1).
This indicates that some questions seem to be more
accurately predicted based on the movement behavior of
TYS users. (2) As can be also obtained, some quartiles have
an accuracy of 0%. This can have several reasons: First,
the gathered GPS data was not enough for a prediction
or, second, the users have not provided GPS data that
is valuable with respect to the used feature set. Another
explanation is that the GPS data does not correlate with the
stress parameters in these quartiles. However, when having
a look at the row denoted with QT*, it is striking that
prediction accuracies of 0% are accompanied by less GPS
measurements. Therefore, the first explanation seems to be
most likely. (3) Finally, for this study, it can be observed that
the DT outperformed the SVM. Hence, further investigations
must be accomplished to compare the suitability of the
different algorithms. Note that in other works, also a DT
has outperformed other approaches on GPS-based data [25].
Whether the reasons of our work and [25] can be compared,
must be further on investigated. It can be possibly assumed
that if GPS trajectories, for example, like presented in [25],
show sudden transport changes (e.g., when changing from
walking to driving), a DT copes better with such changes.
VI. DISCUSSION
In general, the results indicate that using movement behav-
ior to predict stress parameters of the TYS users based on
machine learning is promising. Yet, this study has several
limitations, as explained below. First, the TYS data set is
rather small, and the training sample is even smaller. While
the choice of a small sample for training decreases the
likelihood of overfitting, it also decreases accuracy. Other
forms of sampling could be considered in the future. Second,
the TYS mobile apps are not officially released to the App
Stores. Following this, the invited TYS users might pose
characteristics that bias the generalizability of the presented
results. This includes the distribution of the sample data
across the quartiles. Third, stress-levels have been designed
manually and the features are also handcrafted. So, the
relation between features and stress-level is not semantically
QMLA Quartile
0
Quartile
1
Quartile
2
Quartile
3
WA
1
DT 0.429 0.318 0.340 0.000 0.384
1
SVM 0.106 0.728 0.075 0.000 0.269
1
QT* 826 476 356 148
2
DT 0.509 0.331 0.349 0.493 0.437
2
SVM 0.174 0.166 0.716 0.105 0.264
2
QT* 203 310 443 850
3
DT 0.458 0.368 0.402 0.500 0.432
3
SVM 0.135 0.600 0.253 0.100 0.248
3
QT* 956 406 292 152
4
DT 0.535 0.374 0.363 0.543 0.442
4
SVM 0.218 0.156 0.667 0.114 0.316
4
QT* 505 485 497 319
6
DT 0.000 0.000 0.396 0.510 0.488
6
SVM 0.000 0.000 0.275 0.785 0.683
6
QT* 34 97 424 1251
7
DT 0.597 0.400 0.392 0.450 0.456
7
SVM 0.168 0.187 0.481 0.290 0.241
7
QT* 494 746 386 180
8
DT 0.581 0.419 0.387 0.534 0.460
8
SVM 0.349 0.538 0.167 0.139 0.219
8
QT* 241 309 659 597
9
DT 0.000 0.362 0.414 0.591 0.539
9
SVM 0.000 0.224 0.534 0.094 0.203
9
QT* 130 213 433 1030
Q=Question, WA = Weighted Average
MLA = Machine Learning Algorithm
DT = Decision Tree, SVM = Support Vector Machine
QT* = Quantity, denoting the number of
evaluated GPS measurements for this quartile
TABLE II: Prediction Results
0,106
0,728
0,075
0
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0 1 2 3
Correctly Classified
Quartiles
SVM: training set = 400, validation set = 1431, folds = 10
Fig. 1: SVM Results for Question 1
grounded. Nonetheless, we belied that the results do not
justify further investigations on the associations between
geospatial and stress data that were gathered by a mHealth
crowdsensing platform. In particular, it should be investi-
gated next whether the differences in the obtained accuracies
among the quartiles are due to differences in the population
distribution, or that they can be explained differently. A
further step would be also to monitor stress-levels through
sensors (e.g., through the performance of mobile cortisol
assessments) rather than resorting to hand-crafted stress-
levels. Finally, differences between iOS and Android users in
the light of associations between geospatial and stress data
seem also to be worth of being investigated.
VII. SUMMARY AND OUTLOOK
This work has presented the TrackYourStress mHealth
crowdsensing platform (TYS). It was developed to assess
the stress levels and fluctuations of its users in daily life.
The goal of the platform is to learn more about stress and
related factors. We have also shown that mobile crowdsens-
ing combined with Ecological Momentary Assessments is
less considered for stress in particular and health questions
in general. As a particular question for TYS, we investigated
in the work at hand whether geospatial data (i.e., GPS
measurements) can predict stress-related data using machine
learning algorithms. Although the presented setting has nu-
merous limitations, the results have shown three valuable
insights. First, mobile crowdsensing and Ecological Momen-
tary Assessments are valuable instruments in the context
of stress assessment. Second, geospatial data allows for the
prediction of stress parameters with an accuracy that paves
the way for further investigations. Third, machine learning
algorithms seem to be worth being applied to EMA-driven
mobile crowdsensing data in the context of stress.
REFERENCES
[1] S. Cohen et al., “Psychological stress and disease,” Jama, vol. 298,
no. 14, pp. 1685–1687, 2007.
[2] H. ´
orarinsd´
ottir et al., “Smartphone-based self-assessment of stress
in healthy adult individuals: a systematic review,” Journal of medical
Internet research, vol. 19, no. 2, 2017.
[3] D. Mart´
ın et al., “Managing and controlling stress using mhealth:
Systematic search in app stores,” JMIR mHealth and uHealth, vol. 6,
no. 5, 2018.
[4] A. Stone, J. Schwartz, J. Neale, S. Shiffman, C. Marco, M. Hickcox,
J. Paty, L. Porter, and L. Cruise, “A comparison of coping assessed by
ecological momentary assessment and retrospective recall.” Journal of
personality and social psychology, vol. 74, no. 6, p. 1670, 1998.
[5] S. Shiffman, A. Stone, and M. Hufford, “Ecological Momentary
Assessment,” Annu. Rev. Clin. Psychol., vol. 4, pp. 1–32, 2008.
[6] H. Ma et al., “Opportunities in mobile crowd sensing,” IEEE Com-
munications Magazine, vol. 52, no. 8, pp. 29–35, 2014.
[7] R. Kraft et al., “Design and Implementation of a Scalable Crowd-
sensing Platform for Geospatial Data of Tinnitus Patients,” in 32th
IEEE Int’l Symposium on Computer-Based Medical Systems. IEEE
Computer Society Press, 2019.
[8] R. Pryss et al., “Requirements for a flexible and generic api enabling
mobile crowdsensing mhealth applications,” in 4th Int’l WS on Re-
quirements Engineering for Self-Adaptive, Collaborative, and Cyber
Physical Systems. IEEE, 2018, pp. 24–31.
[9] H. Xiong et al., “Sensus: a cross-platform, general-purpose system
for mobile crowdsensing in human-subject studies,” in Proc of the
Int’l Conf on Pervasive and Ubiquitous Computing. ACM, 2016, pp.
415–426.
[10] K. Agrawal et al., “Towards incentive management mechanisms in
the context of crowdsensing technologies based on trackyourtinnitus
insights,” Procedia computer science, vol. 134, pp. 145–152, 2018.
[11] F. Beierle et al., “What data are smartphone users willing to share
with researchers?” Journal of Ambient Intelligence and Humanized
Computing, pp. 1–13, 2019.
[12] R. Pryss et al., “Differences between android and ios users of the
trackyourtinnitus mobile crowdsensing mhealth platform,” in 31st
International Symp on Computer-Based Medical Systems. IEEE,
2018, pp. 411–416.
[13] M. L. P. et al., “Finding tinnitus patients with similar evolution of
their ecological momentary assessments,” in IEEE 31st Int’l Symp on
Computer-Based Medical Systems, 2018, pp. 112–117.
[14] T. Probst et al., “Emotional States as Mediators between Tinnitus
Loudness and Tinnitus Distress in Daily Life: Results from the
“TrackYourTinnitus” application,” Scientific Reports, vol. 6, 2016.
[15] T. Probst, R. Pryss, B. Langguth, and W. Schlee, “Emotion dynamics
and tinnitus: Daily life data from the “trackyourtinnitus” application,”
Scientific reports, vol. 6, 2016.
[16] W. Schlee et al., “Measuring the moment-to-moment variability of
tinnitus: the trackyourtinnitus smart phone app,” Frontiers in Aging
Neuroscience, vol. 8, 2016.
[17] R. Pryss et al., “Prospective crowdsensing versus retrospective ratings
of tinnitus variability and tinnitus–stress associations based on the
trackyourtinnitus mobile platform,” International Journal of Data
Science and Analytics, pp. 1–12, 2018.
[18] U. Ebner-Priemer and T. Kubiak, “Psychological and psychophysio-
logical ambulatory monitoring,” European Journal of Psychological
Assessment, vol. 23, no. 4, pp. 214–226, 2007.
[19] D. De Beurs et al., “The role of mobile phone technology in under-
standing and preventing suicidal behavior,” 2015.
[20] E. Garcia-Ceja et al., “Automatic stress detection in working environ-
ments from smartphones’ accelerometer data: A first step.” IEEE J.
Biomedical and Health Informatics, vol. 20, no. 4, pp. 1053–1060,
2016.
[21] S. Vhaduri et al., “Estimating Drivers’ Stress from GPS Traces,”
in Proc of the 6th Int’l Conf on Automotive User Interfaces and
Interactive Vehicular Applications. ACM, 2014, pp. 1–8.
[22] D. Ben-Zeev et al., “Next-generation psychiatric assessment: Using
smartphone sensors to monitor behavior and mental health.” Psychi-
atric rehabilitation journal, vol. 38, no. 3, p. 218, 2015.
[23] J. Kerr et al., “Using global positioning systems in health research:
a practical approach to data collection and processing,” American
journal of preventive medicine, vol. 41, no. 5, pp. 532–540, 2011.
[24] V. Cornet and R. Holden, “Systematic review of smartphone-based
passive sensing for health and wellbeing,” Journal of biomedical
informatics, vol. 77, pp. 120–132, 2018.
[25] Y. Zheng et al., “Understanding mobility based on gps data,” in Proc
of the 10th Int’l Conf on Ubiquitous Computing. ACM, 2008, pp.
312–321.
[26] K. Ellis et al., “Identifying active travel behaviors in challenging envi-
ronments using gps, accelerometers, and machine learning algorithms,”
Frontiers in public health, vol. 2, p. 36, 2014.
[27] A. Jahangiri and H. Rakha, “Applying machine learning techniques
to transportation mode recognition using mobile phone sensor data,”
IEEE transactions on intelligent transportation systems, vol. 16, no. 5,
pp. 2406–2417, 2015.
[28] Z. Liao et al., “Anomaly detection in gps data based on visual
analytics,” in Symposium on Visual Analytics Science and Technology.
IEEE, 2010, pp. 51–58.
[29] M. Williams et al., “A novel behavioral model of the pasture-based
dairy cow from gps data using data mining and machine learning
techniques,” Journal of dairy science, vol. 99, no. 3, pp. 2063–2075,
2016.
[30] D. Mohr et al., “Personal sensing: understanding mental health using
ubiquitous sensors and machine learning,” Annual review of clinical
psychology, vol. 13, pp. 23–47, 2017.
[31] M. Gjoreski et al., “Automatic detection of perceived stress in campus
students using smartphones,” in Int’l Conf on Intelligent Environments.
IEEE, 2015, pp. 132–135.
[32] H. Sarker et al., “Finding significant stress episodes in a discontinuous
time series of rapidly varying mobile sensor data,” in Proc of the Conf
on Human Factors in Computing Systems. ACM, 2016, pp. 4489–
4501.
[33] F. Beierle et al., “Context data categories and privacy model for mobile
data collection apps,” Procedia computer science, vol. 134, pp. 18–25,
2018.
[34] M. Bradley and P. Lang, “Measuring Emotion: The Self-Assessment
Manikin and the Semantic Differential,” Journal of Behavior Therapy
and Experimental Psychiatry, vol. 25, no. 1, pp. 49–59, 1994.
[35] M. Ester et al., “A density-based algorithm for discovering clusters
a density-based algorithm for discovering clusters in large spatial
databases with noise,” in Proc of the Second Int’l Conf on Knowledge
Discovery and Data Mining, ser. KDD’96. AAAI Press, 1996, pp.
226–231.
