Content uploaded by Filipe Barata
Author content
All content in this area was uploaded by Filipe Barata on Oct 26, 2018
Content may be subject to copyright.
Personal MobileCoach: Tailoring Be-
havioral Interventions to the Needs of
Individual Participants
Abstract
MobileCoach, an open source behavioral intervention platform,
has been developed to provide health professionals with an
authoring tool to design evidence-based, scalable and
low-cost digital health interventions (DHI). Its potential
meets the lack in resources and capacity of health care
systems to provide DHI for the treatment of noncom-
municable diseases. In the current work, we introduce
the first personalization approach for MobileCoach with
the purpose of identifying the needs of participants,
tailoring the treatment and, as a consequence, enhanc-
ing the capability of MobileCoach-based DHIs. The per-
sonalization approach is then exemplified by a very first
prototype of a DHI for people with asthma that is able
to detect coughing by just using a smartphone’s micro-
phone. First empirical results with five healthy subjects
and 80 coughs indicate its technical feasibility as the
detection accuracy yielded 83.3%. Future work will
focus on the integration of personalized sensing and
supporting applications for MobileCoach.
Author Keywords
Open Source Platform; Behavioral Intervention; Public
Health; Personalization; Machine Learning;
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. Copyrights
for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to
post on servers or to redistribute to lists, requires prior specific permis-
sion and/or a fee. Request permissions from Permissions@acm.org.
Ubicomp/ISWC'16 Adjunct , September 12-16, 2016, Heidelberg, Germa-
ny
© 2016 ACM. ISBN 978-1-4503-4462-3/16/09…$15.00
DOI: http://dx.doi.org/10.1145/2968219.2972713
Filipe Barata
ETH Zürich
8092 Zürich, CH
fbarata@ethz.ch
Tobias Kowatsch
University St.Gallen & ETH
Zürich
9000 St.Gallen, CH
tobias.kowatsch@unisg.ch
Peter Tinschert
University St.Gallen
9000 St.Gallen, CH
peter.tinschert@unisg.ch
Andreas Filler
University of Bamberg & ETH
Zürich
96047 Bamberg, DE
afiller@ethz.ch
1089
UBICOMP/ISWC ’16 ADJUNCT, SEPTEMBER 12-16, 2016, HEIDELBERG, GERMANY
ACM Keywords
J.3 Life and Medical Science; I.2 Artificial Intelligence;
K.3.2 Learning; J.4 Social and Behavioral Sciences;
F.1.1 Models of Computation
Introduction
To support an individual to adopt a health-enhancing
behavior has been the motivation for the design of
effective and efficient health interventions. Based on
behavioral health models [1, 9, 13, 17, 24] and tech-
niques [1, 22], behavior change interventions have
been the basis for several applications in the health
domain and investigated thoroughly in the last dec-
ades. The existing bottleneck at resources and the ca-
pacity to provide behavioral health interventions for the
treatment of noncommunicable diseases (NCDs) pose
one of the greatest challenges for health care systems
worldwide [27]. Furthermore, NCDs represented the
leading cause of death in 2012 responsible for more
than two thirds of the global deaths [5].
With respect to these shortcomings, Internet of Person-
al Health systems, in particular digital health interven-
tions (DHIs), have the potential to play a crucial role in
reducing the burden for an individual patient as well as
for the whole health care system. Not only could they
create the possibility of improving the effectiveness and
efficiency of health interventions, but also reduce their
costs [2]. In contrast to their potential, the design and
implementation of evidence-based, scalable and low-
cost DHI is sparsely empirically supported [23].
In this context, MobileCoach, an open source platform
for DHIs, was introduced with the main objective “to
give scientists, public and behavioral health experts a
software platform that allows them to design evidence-
based, scalable and low-cost health interventions.” [8,
p. 1] With a modular design, which enables developers
of DHIs to extend and expand MobileCoach to their
specific needs, it is the rule-based engine which moni-
tors health states and triggers state transitions that lies
at its heart. Moreover, MobileCoach bears the potential
for the much-needed scalable and low-cost supply of
DHIs to address the global healthcare insufficiencies,
especially with regard to the previously mentioned
NCDs.
Asthma, a chronic disease involving the airways and
the lungs, ranks among the most prevalent NCDs. 30
million children and adults (less than 45 years old)
suffer from it in Europe [10] and 3,630 people die from
it each year in the USA [7]. According to clinical guide-
lines, asthma control, a term used to describe the
course of the disease, can be assessed by symptoms
such as wheezing, breathlessness, chest tightness and
coughing. Coughing, particularly the amount of coughs
per day or per night, is reported to provide an objective
assessment which correlates with the standard meas-
ure of asthma control [21].
In addition to being used as a measure of asthma con-
trol, coughing, a common symptom for many respirato-
ry diseases [15], is a well-recognized indicator for the
improvement of diagnostics. As a consequence, many
efforts have been made towards the development of
objective audio cough monitoring systems, which can
be traced all the way back to the 1950s [3]. Recent
advances have been accomplished by employing ma-
chine learning to automatically detect coughs into a
semi-automated [4] or fully automated procedure [19].
The latter is a smartphone-based solution that merely
makes use of a smartphone’s built-in microphone to
record coughing sounds and to subsequently detect and
count personal coughs.
In conclusion, a smartphone not only enables recording
and monitoring of cough sounds in near real-time but
also provides a low-cost and scalable scope for a sens-
ing application which is able to trigger health interven-
tions for people with asthma based on an objective
assessment of asthma control. In this current work we
1090
UBICOMP/ISWC ’16 ADJUNCT, SEPTEMBER 12-16, 2016, HEIDELBERG, GERMANY
elaborate on the development and design of the afore-
said application for the purpose of integrating it into the
MobileCoach platform as a trigger for future behavioral
health interventions. Furthermore, we want to enhance
the capability of such a sensing application by using
machine learning techniques for personalization.
Personalization is known from the web customization
process in which organizations determine the needs of
the users and provide them advertising services, shop-
ping services and filtering without the users having to
ask for it explicitly [28]. In the context of MobileCoach,
personalization may first help identifying the needs of
the participants in the intervention more precisely and
therefore develop a more sensitive trigger for DHIs,
second it may help to tailor DHIs themselves, hence
enhance their capability. With regard to the sensing
application for cough monitoring this could mean that
the system recognizes when it may lack the capability
to monitor the participant’s coughing rates precisely
and therefore start on its own behalf to adapt the sys-
tem’s model by learning from the participant’ individual
coughs.
The remainder of this current work is organized as
follows. We briefly outline a novel personalization ap-
proach for MobileCoach. We then describe coughing
detection by means of a smartphone, present first em-
pirical results from a technical feasibility study and
describe how it can be related to a personalized Mo-
bileCoach DHI. We conclude with a summary and an
outlook on future work.
The personal MobileCoach
In order to add personalization to the current Mo-
bileCoach, we illustrate its conceptual design first. The
MobileCoach system as it has been introduced [8] fol-
lows the sequential logic of a state machine, where the
state transitions are determined by intervention rules.
The state corresponds to an aggregation of all signifi-
cant variables, which are relevant to the intervention
progress of the participant. Whenever a state transition
is triggered by an intervention rule, significant partici-
pant variables are changed and therefore a change in
the state machine occurs.
In particular, a participant registers himself / herself to
the system by filling out an online baseline assessment
(e.g. nickname, age, mobile number, intentions and the
like), the consequence of which is the assignment of an
initial state. By this action a fully automated dialog
between participant and system commences. This dia-
log is executed over a text message service and con-
sists of questions with pre-defined answer schemes.
Depending on the participant’s answer, state transitions
are triggered and states are changed, based on the
intervention flow formalized as rules, to lastly tailor the
follow-up communication between participant and sys-
tem.
It is just this exchange of information between partici-
pant and system over a longer period of time that we
want to utilize to personalize MobileCoach. The more
communication occurs and information is exchanged,
the more evidence we have to possibly derive a com-
prehensive picture of the participant’s needs. With re-
gard to the conceptual design described above, the
implications are that state transitions can no longer be
triggered by taking into account only the last answer of
the participant and the system’s current state. In fact,
in addition to the participant’s answer, the current state
and all the previous states have to be considered be-
fore a participant-specific behavioral health intervention
can be triggered. With this in mind, a personalized
MobileCoach would not only be able to recognize the
interventions that had the biggest impact on the inter-
vention progress of the participant, but also omit the
ones that had less and lastly determine the treatment
that is most efficient for the participant. Finally, differ-
ent forms of personalization have been used in different
fields. Among the most prominent, web search person-
Figure 1: In this picture we show
the coughing detection applica-
tion, which makes use of the
described algorithm. This applica-
tion also enables a recording
functionality, which was used in
the test.
1091
SESSION: IOPH
alization, where the goal is to tailor search queries to a
particular individual, based on her interests and prefer-
ences [16]. With regard to mobile application, person-
alized training periods have been exploited to infer the
mood state of an user by analyzing communication
history and application usage patterns of a smartphone
[20].
The personalized cough detection module
for MobileCoach
As described in the previous section, personalization
should occur over time. For machine learning based
cough detection, this could look as follows: After each
night, the application detects and counts coughs and
possibly triggers DHI based on the amount of coughs.
Additionally, the application may choose a number of
recorded sound sequences, preferably sequences which
have been on the verge of being detected as cough
sounds or being ignored, respectively. These chosen
sequences are then queried to the participant to be
labeled as coughs or non-coughs respectively. Alterna-
tively, this data could also be judged and labeled by a
third party such as cough experts or on-demand work-
ers (e.g. [12]). Finally, by continuously integrating this
new labeled data in the model learning process, per-
sonalization is achieved. This special case of supervised
learning is known in literature as active learning and is
characterized by its queries [25].
Results
The subject of the current work has been the develop-
ment of a coughing detection module for MobileCoach
to provide personalization as described above. The
developed module is based on classification of spectral
features from the audio signal of a smartphone’s mi-
crophone. The system pursues a similar implementation
as in [6]. However, support vector machine classifica-
tion was used instead of simple decision-tree classifica-
tion, which proved to generate better results in terms
of sensitivity, specificity and accuracy with respect to
our settings. A first test included a population of 5
healthy subjects (2 female, 3 male, mean age: 27, SD:
2.55), recorded by means of a bespoke app running on
Android 6.0 OS on a HTC M8 smartphone. In this first
test, 16 intentional coughs were recorded instead of
natural coughs. The participants were instructed to
intentionally cough while being recorded by the afore-
said smartphone. Analogously, the participant was
asked to read for a limited amount of time. These re-
cordings build the coughing data and non-coughing
data, from which a predicting model was learned. The
evaluation of the test yielded 86.7% sensitivity and
81.0% specificity, with an accuracy of 83.3%.
Future Work
We emphasized in this work the expendability of
MobileCoach towards personalization and a new sensing
module for smartphones. In the matter of asthma,
further sensing modules for MobileCoach can be
thought of as for instance, smartphone-based
spirometry. Spirometry is the mainstay for measuring
lung function and reports suggest that its functionality,
namely the measurement of instantaneous flow and
cumulative volume of exhaled air, may be reproducable
by means of a mobile phone [11, 18]. Even beyond
sensing, we can consider supportive DHIs. For
example, wind instruments have been used in music
therapy to improve the skill of asthmatic patients to
attack asthma [14]. This finding can potentially be
exploited by a smartphone-based wind instrument
application [26] and be embedded in a behavioral DHI.
Against this background, we want to extend
MobileCoach with various sensing and supporting DHIs
in our future work: not only to provide a broader tool-
set to intervention experts, but also to design efficient
1092
UBICOMP/ISWC ’16 ADJUNCT, SEPTEMBER 12-16, 2016, HEIDELBERG, GERMANY
and effective interventions, which are tailored to the
needs of the participants and provide objective
(physiological) data on therapy adherence.
Acknowledgments
We would like to thank Gabriella Chiesa and Niklas
Elser of the CSS Health Lab for their valuable feedback.
This work is part-funded by CSS Insurance, Switzer-
land.
References
1. Abraham, C. and Michie, S. A Taxonomy of
Behavior Change Techniques Used in Interventions.
Health Psychology, 27, (3), 379-387.
2. Agarwal, R., Gao, G., DesRoches, C. and Jha, A. K.
The Digital Transformation of Healthcare: Current
Status and the Road Ahead. Information Systems
Research, 21, (4), 796-809.
3. Bickerman, H. A. and Itkin, S. E. The effect of a
new bronchodilator aerosol on the air flow
dynamics of the maximal voluntary cough of
patients with bronchial asthma and pulmonary
emphysema. Journal of chronic diseases, 8, (5),
629-636.
4. Birring, S. S., Fleming, T., Matos, S., Raj, A. A.,
Evans, D. H. and Pavord, I. D. The Leicester Cough
Monitor: preliminary validation of an automated
cough detection system in chronic cough. European
Respiratory Journal, 31, (5), 1013-1018.
5. Bloom, D. E., Cafiero, E. T., Jané-Llopis, E.,
Abrahams-Gessel, S., Bloom, L. R., Fathima, S.,
Feigl, A. B., Gaziano, T., Mowafi, M., Pandya, A.,
Prettner, K., Rosenberg, L., Seligman, B., Stein, A.
Z. and Weinstein, C. The Global Economic Burden
of Non-communicable Diseases. World Economic
Forum and the Harvard School of Public Health,
Geneva, Switzerland, 2011.
6. Casaseca-de-la-Higuera, P., Lesso, P., McKinstry,
B., Pinnock, H., Rabinovich, R., McCloughan, L. and
Monge-Alvarez, J. Effect of downsampling and
compressive sensing on audio-based continuous
cough monitoring. IEEE, 2015.
7. Centers for Disease Control Prevention Asthma
facts—CDC’s national asthma control program
grantees. 2013.
8. Filler, A., Kowatsch, T., Haug, S., Wahle, F.,
Staake, T. and Fleisch, E. MobileCoach: A Novel
Open Source Platform for the Design of Evidence-
based, Scalable and Low-Cost Behavioral Health
Interventions - Overview and Preliminary
Evaluation in the Public Health Context. In
Proceedings of the 14th annual Wireless
Telecommunications Symposium (WTS 2015)USA,
New York, 2015. IEEE,
9. Fishbein, M. and Yzer, M. C. Using Theory to Design
Effective Health Behavior Interventions.
Communication Theory, 13, (2), 164-183.
10. Gibson, G. J., Loddenkemper, R., Sibille, Y. and
Lundback, B. The European Lung White Book:
Respiratory Health and Disease in Europe.
European Respiratory Society, 2013.
11. Goel, M., Saba, E., Stiber, M., Whitmire, E.,
Fromm, J., Larson, E. C., Borriello, G. and Patel, S.
N. SpiroCall: Measuring Lung Function over a
Phone Call. ACM, 2016.
12. Goodman, J. K., Cryder, C. E. and Cheema, A. Data
collection in a flat world: The strengths and
weaknesses of Mechanical Turk samples. Journal of
Behavioral Decision Making, 26, (3), 213-224.
13. Green, L. W. and Kreuter, M. W. Health program
planning: An educational and ecological approach.
1093
SESSION: IOPH
14. Hänninen, S. Breathing woodwinds: music therapy
for asthma and COPD rehabilitation.
15. Irwin, R. S., Curley, F. J. and French, C. L. Chronic
cough: the spectrum and frequency of causes, key
components of the diagnostic evaluation, and
outcome of specific therapy. American Review of
Respiratory Disease, 141, (3), 640-647.
16. Jay, P., Shah, P., Makvana, K. and Shah, P. Review
on web search personalization through semantic
data. IEEE, 2015.
17. Kok, G., Schaalma, H., Ruiter, R. A. C., Van
Empelen, P. and Brug, J. Intervention mapping:
protocol for applying health psychology theory to
prevention programmes. Journal of health
psychology, 9, (1), 85-98.
18. Larson, E. C., Goel, M., Boriello, G., Heltshe, S.,
Rosenfeld, M. and Patel, S. N. SpiroSmart: using a
microphone to measure lung function on a mobile
phone. ACM, 2012.
19. Larson, E. C., Lee, T., Liu, S., Rosenfeld, M. and
Patel, S. N. Accurate and privacy preserving cough
sensing using a low-cost microphone. ACM, 2011.
20. LiKamWa, R., Liu, Y., Lane, N. D. and Zhong, L.
Moodscope: Building a mood sensor from
smartphone usage patterns. ACM, 2013.
21. Marsden, P. A., Satia, I., Ibrahim, B., Woodcock,
A., Yates, L., Donnelly, I., Jolly, L., Thomson, N. C.,
Fowler, S. J. and Smith, J. A. Objective Cough
Frequency, Airway Inflammation and Disease
Control in Asthma. Chest.
22. Michie, S., Richardson, M., Johnston, M., Abraham,
C., Francis, J., Hardeman, W., Eccles, M. P., Cane,
J. and Wood, C. E. The Behavior Change Technique
Taxonomy (v1) of 93 Hierarchically Clustered
Techniques: Building an International Consensus
for the Reporting of Behavior Change
Interventions. Annals of Behavioral Medicine, 46,
(1), 81-95.
23. Moja, L., Kwag, K. H., Lytras, T., Bertizzolo, L.,
Brandt, L., Pecoraro, V., Rigon, G., Vaona, A.,
Ruggiero, F., Mangia, M., Iorio, A., Kunnamo, I.
and Bonovas, S. Effectiveness of computerized
decision support systems linked to electronic health
records: a systematic review and meta-analysis.
Amercian Journal of Public Health, 104, (12), 12-
22.
24. Schwarzer, R. Health Behavior Change: How to
Predict and Modify the Adoption and Maintenance
of Health Behaviors. Applied Psychology, 57, (1),
1-29.
25. Settles, B. Active learning literature survey.
University of Wisconsin, Madison, 52, (55-66), 11.
26. Wang, G. Designing smule’s iphone ocarina. 2009.
27. WHO Scaling up action against NCDs: How much
will it cost? World Health Organization, Genéve,
Switzerland, 2011.
28. Zafar, M. S. and Ahmad, F. A Study on
Personalization and Customization Mechanisms of
Vehicular Cloud Platform. IEEE, London, 2014.
1094
UBICOMP/ISWC ’16 ADJUNCT, SEPTEMBER 12-16, 2016, HEIDELBERG, GERMANY
