Predicting Feedback Compliance in a Teletreatment
Application
Harm op den Akker
Roessingh Research and Development
Enschede, the Netherlands
Email: h.opdenakker@rrd.nl
Val Jones
Telemedicine Group, University of Twente
Enschede, the Netherlands
Email: v.m.jones@ewi.utwente.nl
Hermie Hermens
Roessingh Research and Development
Telemedicine Group, University of Twente
Enschede, the Netherlands
Email: h.hermens@rrd.nl
Abstract—Health care provision is facing resourcing challenges
which will further increase in the 21st century. Health care
mediated by technology is widely seen as one important element
in the struggle to maintain existing standards of care. Personal
health monitoring and treatment systems with a high degree of
autonomic operation will be required to support self-care. Such
systems must provide many services and in most cases must
incorporate feedback to patients to advise them how to manage
the daily details of their treatment and lifestyle changes. As in
many other areas of healthcare, patient compliance is however an
issue. In this experiment we apply machine learning techniques to
three corpora containing data from trials of body worn systems
for activity monitoring and feedback. The overall objective is
to investigate how to improve feedback compliance in patients
using personal monitoring and treatment systems, by taking into
account various contextual features associated with the feedback
instances. In this article we describe our first machine learning
experiments. The goal of the experiments is twofold: to determine
a suitable classification algorithm and to find an optimal set of
contextual features to improve the performance of the classifier.
The optimal feature set was constructed using genetic algorithms.
We report initial results which demonstrate the viability of this
approach.
Index Terms—Mobile healthcare, activity monitoring, feedback
compliance, machine learning, genetic algorithms.
I. INTRODUCTION
An ambulant system has been developed designed to guide
the patient to reach a healthy distribution of activity over the
day. The system consists of a 3D-accelerometer to assess the
patient’s daily activity pattern in counts per minute, combined
with a PDA for providing feedback. By comparing his activity
to some predetermined reference activity pattern the patient is
provided with feedback messages at regular intervals advising
them to be more or less active or that they are performing
well. This system was used in three different patient groups:
chronic low back pain (CLBP) patients [1], chronic fatigue
syndrome (CFS) patients [2], [3] and people suffering from
obesity (BMI > 30). In the case of CLBP and CFS patient
populations, the goal of the feedback is to spread activity over
the day, while for obesity patients, the goal is to encourage
them to increase activity over all.
In this research we are looking at the responses to the
individual feedback message with a view to developing a
method of generating messages in a smarter, more efficient
and personalized manner. Related work in the field is reported
in [4], [5] where the aim is to cluster diabetic patients based on
the type of messages that seem to provoke a positive response
in the patients. They report preliminary, but promising results
in a dynamic clustering system that learns the preferences
of users over time. Our overall objective is to investigate
how to improve feedback compliance in patients by taking
into account various contextual features associated with the
feedback instances.
II. DATASETS
For this research we used retrospective data consisting
of three datasets (or corpora): the CLBP corpus, the CFS
corpus and the Obesity corpus. The patients from our three
populations had carried the monitoring system on average 13.6
days ( = 11:7). In all three studies, the protocol included an
experimental group who received feedback on some days, and
on other (control) days did not; as well as a control group
who never received feedback. A total of 45 patients received
feedback from the system. For these patients, feedback was
given on average on 13 days ( = 8:8). Patients were asked
to wear the system during waking hours (approximately from
8 am, until 10 pm). In all three studies the measurement
system consisted of a PDA connected wirelessly to a 3D-
accelerometer measuring the patient’s physical activity levels
throughout the day. Based on measured values, a variety of
feedback messages are given to the patient via the PDA screen.
The system logs acceleration as an integrated value, summed
up over the three axis of movement per 60 second interval
[6] as well as the timing and content of the given feedback
messages.
Table I gives an overview of some relevant statistics on these
three corpora. Due to gaps in the sensor data, or erroneous
timing of the feedback messages (sometimes two messages
would be generated too close to each other, rendering one
of them useless) we could not use all the data for our data
analysis. The last row in Table I shows the total number of data
instances that were used for our machine learning experiments
following exclusion of the problem data.
III. METHOD
To see how people respond to the feedback messages we
use a compliance measure. In this article we report on a
compliance measure which compares the amount of activity
performed in the 30 minute interval before the feedback