920 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 58, NO. 4, APRIL 2011
Adaptive Sleep–Wake Discrimination for
Wearable Devices
Walter Karlen*, Member, IEEE, and Dario Floreano, Senior Member, IEEE
Abstract—Sleep/wake classification systems that rely on phys-
iological signals suffer from intersubject differences that make
accurate classification with a single, subject-independent model
difficult. To overcome the limitations of intersubject variability,
we suggest a novel online adaptation technique that updates the
sleep/wake classifier in real time. The objective of the present study
was to evaluate the performance of a newly developed adaptive clas-
sification algorithm that was embedded on a wearable sleep/wake
classification system called SleePic. The algorithm processed ECG
and respiratory effort signals for the classification task and ap-
plied behavioral measurements (obtained from accelerometer and
press-button data) for the automatic adaptation task. When trained
as a subject-independent classifier algorithm, the SleePic device
was only able to correctly classify 74.94 ± 6.76% of the human-
rated sleep/wake data. By using the suggested automatic adaptation
method, the mean classification accuracy could be significantly im-
proved to 92.98 ± 3.19%. A subject-independent classifier based
on activity data only showed a comparable accuracy of 90.44 ±
3.57%. We demonstrated that subject-independent models used
for online sleep–wake classification can successfully be adapted
to previously unseen subjects without the intervention of human
experts or off-line calibration.
Index Terms—Adaptation, context awareness, personal health,
physiological signal classification, point-of-care, wearable.
I. INTRODUCTION
MONITORING sleep–wake behavior of subjects at homeallows the early detection of sleep disorders and is re-
ducing health care costs [1]. Ambulatory health applications
require comfortable devices that embed wearable sensors, elec-
tronics, and intelligent signal processing. The design of wearable
sleep/wake discrimination systems is particularly challenging.
The most common physiological signal used for sleep discrim-
ination in clinical settings is the recording of brain activity with
an EEG [2]. Unfortunately, EEG cannot be easily recorded with
a wearable system and is subject to an increased level of noise.
An alternative method is needed. It has also been shown that
Manuscript received July 11, 2010; accepted November 16, 2010. Date of
publication December 17, 2010; date of current version March 18, 2011. This
work was supported by the Solar Impulse project grant of Ecole Polytechnique
Fe´de´rale de Lausanne (EPFL). The work of D. Floreano was supported in part by
the CURVACE Project sponsored by the Future and Emerging Technology Di-
vision within the seventh Framework Programme for Research of the European
Commission under FET-Open Grant 237940. Asterisk indicates corresponding
author.
∗W. Karlen is with the Electrical and Computer Engineering in Medicine
Group, the University of British Columbia, Vancouver, BC V6T 1Z4, Canada
(e-mail: walterk@ece.ubc.ca).
D. Floreano is with the Laboratory of Intelligent Systems, Institute of Micro-
engineering, Ecole Polytechnique Fe´de´rale de Lausanne, CH-1015 Lausanne,
Switzerland (e-mail: dario.floreano@epfl.ch).
Digital Object Identifier 10.1109/TBME.2010.2097261
during sleep, intersubject differences in EEG [3] and cardio-
respiratory signals [4], [5] are more pronounced than intrasub-
ject variations. Consequently, any signal processing and clas-
sification algorithm tuned to a model user is bound to produce
highly variable results in different persons. This suggests that on
a mobile device an efficient user adaptation strategy is required.
A. Background
Sleep–wake behavior is normally monitored using
polysomnographic analysis that includes the recording of EEG
[6]. Polysomnography is usually conducted in sleep centers
which requires the patient to stay overnight. More recently,
portable recorders were used for ambulatory sleep recordings
that allow the patient to go home overnight. The portable sys-
tems are modular, supporting a multitude of sensors required
for polysomnographic analysis. Recent attempts to integrate
sensors and electrodes into textiles made the recorders more
wearable. Despite these advances, the devices often remain
bulky. Furthermore, the portable systems were only used for
recording and not for signal processing or classification. Instead
of polysomnographic recordings, the less accurate actigraphy
method is often used for long-term sleep studies [7], [8]. Actig-
raphy is a passive measure of sleep/wake behavior. Miniature
accelerometers in a watch-like device are used to record the
movement patterns of the subject. These wristbands are small,
light weight, and low power and therefore easy to wear over mul-
tiple days. Several classification algorithms have been suggested
for actigraphy analysis [9]–[13]. However, they do not provide
real-time detection of sleep and wake. Furthermore, they often
incorrectly classify low activity tasks (e.g., reading or watching
television) as sleep because the measured behavioral quiescence
is not unique to sleep [8], [11]. Furthermore, actigraphy is not
a good tool for detecting wakefulness in subjects with irregular
or fragmented sleep schedules [14].
We have previously demonstrated online sleep/wake classifi-
cation based on power spectral density estimates of ECG, respi-
ration effort (RSP) [4], and optionally accelerometer (ACC) sig-
nals [15]. We showed that if an artificial neural network (ANN) is
trained and tested later on the same user, a mean correct sleep–
wake classification of 94.23 ± 1.65% can be achieved [15].
However, when the ANN classifier was tested on data from
users who did not contribute to the training of the classifier, the
accuracy dropped significantly to 88.59± 6.66%. This indicated
that at least some of the signals do not generalize well for other
users and a single model cannot be used for accurate classifica-
tion in a larger population. In our previous work, we remarked
that an ANN could be trained for each user individually [4].
However, obtaining the necessary training dataset with accurate
0018-9294/$26.00 © 2011 IEEE