IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 2, APRIL 2009 71
Sleep and Wake Classification With
ECG and Respiratory Effort Signals
Walter Karlen, Student Member, IEEE, Claudio Mattiussi, and Dario Floreano, Senior Member, IEEE
Abstract—We describe a method for the online classification of
sleep/wake states based on cardiorespiratory signals produced by
wearable sensors. The method was conceived in view of its applica-
bility to a wearable sleepiness monitoring device. The method uses
a fast Fourier transform as the main feature extraction tool and
a feedforward artificial neural network as a classifier. We show
that when the method is applied to data collected from a single
young male adult, the system can correctly classify, on average,
95.4% of unseen data from the same user. When the method is ap-
plied to classify data from multiple users with the same age and
gender, its accuracy is reduced to 85.3%. However, receiver oper-
ating characteristic analysis shows that compared to actigraphy,
the proposed method produces a more balanced correct classifica-
tion of sleep and wake periods. Additionally, by adjusting the clas-
sification threshold of the neural classifier, 86.7% of correct classi-
fication is obtained.
Index Terms—Biomedical signal analysis, electrocardiography,
neural classifier, respiratory effort, sleep and wake classification,
wearable computing.
I. INTRODUCTION
I NCREASED sleepiness during daytime has been identifiedas an important cause of accidents in transportation and fac-
tory plants [1]. It is therefore a major health interest to contin-
uously monitor and report the sleepiness level of high-risk per-
sons, such as pilots, truck drivers, or shift workers. Continuously
updated information about the people’s “need for sleep” could
help these people better schedule their breaks and sleep times.
Currently, transport industries focus mainly on emergency sit-
uation prevention by means of vehicle-centered systems that
alert the user either by monitoring the vehicle performances
(e.g., lane deviation) or operators’ behavioral responses (e.g.,
eye blinks). Other fatigue detection techniques include fitness-
for-duty tests and mathematical alertness models [2], [3]. Our
approach consists in using mathematical models in combina-
tion with physiological measurements to establish a continuous
sleepiness profile of the subject and give warnings even before a
certain task begins or an emergency situation related to fatigue
can occur.
Manuscript received March 28, 2008; revised June 13, 2008. Current
version published March 25, 2009. This work was supported by the Solar
Impulse project grant of Ecole Polytechnique Fédérale de Lausanne (EPFL).
Preliminary results of this paper were first presented at the IEEE International
Biomedical Circuits and Systems Conference BIOCAS07, Montreal, QC,
Canada, November 2007. This paper was recommended by Associate Editor
Ralph Etienne-Cummings.
The authors are with the Laboratory of Intelligent Systems, Institute of
Micro-engineering, Ecole Polytechnique Fédérale de Lausanne (EPFL),
Lausanne CH-1015, Switzerland (e-mail: walter.karlen@epfl.ch; claudio.mat-
tiussi@epfl.ch; dario.floreano@epfl.ch).
Digital Object Identifier 10.1109/TBCAS.2008.2008817
Different mathematical models to estimate sleepiness have
been suggested [4]. These models are mainly based on the pre-
vious sleep and wake durations (homeostatic process) and daily
alertness rhythms (circadian process). In this paper, we describe
a method for the estimation of the homeostatic component with
a wearable device. In order to be wearable, the sleep/wake de-
tection device should be energetically autonomous. Since the
person is expected to wear it for several days, the device should
also be lightweight and comfortable. This puts tight restrictions
not only on the choice of signals for the classification task, but
also on the signal recording, processing, and on the computa-
tional requirements of the classifier. In addition, such a device
is intended for the public at large. Therefore, it should be easy to
use and should not depend on complicated calibration methods.
The gold standard for assessing sleep in humans is the
analysis of brain-wave patterns (EEG) first described by
Rechtschaffen and Kales [5]. The most common sleep analysis
method is called polysomnography (PSG), which combines
EEG recordings with different physiological signals, such as
electromyography (EMG), electrooculography (EOG), respira-
tory effort, blood oxygen saturation, electrocardiograms (ECG),
and video analysis. In PSG, 30-s epochs of the signals are used
for decision making. The method is normally carried out in a
controlled hospital environment and needs medical assistance
for setting up sensors, monitoring, and analysis. Although the
analysis is typically computer assisted [6], it still requires a
sleep expert and is therefore expensive and time consuming. It
is difficult to integrate PSG sensors into a wearable device, as
they are rather bulky, power consuming, and highly susceptible
to noise. Furthermore, EEG recordings require many electrodes
to be glued to the scalp, which makes it very cumbersome and
uncomfortable for the user.
In home environments, where PSG is typically not available,
physicians rely on actigraphy for sleep monitoring [7]. In this
method, the acceleration of the extremities (typically wrist)
are recorded over several days with a watch-like device using
miniature accelerometers and a storage medium. Periods of
low activity are later classified as sleep by offline computer
processing. Many different classification algorithms have been
suggested for actigraphy [8], [9], but often they cannot cope
with the problem of misclassifying low activity tasks, such as
reading and watching television or the case where the sensor
band is not worn [7], [9]. Recently, alarm clocks using ac-
celerometers have been commercialized [10], [11]. The activity
is used to detect the best sleep phase for easy wakeup in a given
time window (10 to 30 min). However, the accelerometers
are only active at night and the clocks do not calculate sleep
duration.
1932-4545/$25.00 © 2009 IEEE