Art 1

Biomedical Signal Processing and Control 10 (2014) 2133
Contents lists available at ScienceDirect
Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc
Review
Signal processing techniques applied to human sleep EEG signalsA

review
Shayan Motamedi-Fakhr a, , Mohamed Moshre-Torbati a , Martyn Hill a ,
Catherine M. Hill b , Paul R. White a,b
a
b
Faculty of Engineering and the Environment, University of Southampton, UK

Faculty of Medicine, University of Southampton, UK
a r t i c l e
i n f o
Article history:
Received 11 December 2012
Received in revised form
26 September 2013
Accepted 13 December 2013
Available online 15 January 2014
Keywords:
Sleep
EEG
Signal processing
Quantitative analysis
Artefact reduction
Feature selection/classication
a b s t r a c t
A bewildering variety of methods for analysing sleep EEG signals can be found in the literature. This article
provides an overview of these methods and offers guidelines for choosing appropriate signal processing
techniques. The review considers the three key stages required for the analysis of sleep EEGs namely,
pre-processing, feature extraction, and feature classication. The pre-processing section describes the
most frequently used signal processing techniques that deal with preparation of the sleep EEG signal
prior to further analysis. The feature extraction and classication sections are also dedicated to highlight the most commonly used signal analysis methods used for characterising and classifying the sleep
EEGs. Performance criteria of the addressed techniques are given where appropriate. The online supplementary materials accompanying this article comprise an extended taxonomy table for each section,
which contains the relevant signal processing techniques, their brief descriptions (including their pros
and cons where possible) and their specic applications in the eld of sleep EEG analysis. In order to
further increase the readability of the article, signal processing techniques are also categorised in tabular
format based on their application in intensively researched sleep areas such as sleep staging, transient
pattern detection and sleep disordered breathing diagnosis.
2013 Elsevier Ltd. All rights reserved.
Contents
1.
2.
3.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.
Artefact processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1.
Artefact detection and rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2.
Methods for suppressing artefacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.
Sleep EEG segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.
Temporal features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1.
Instantaneous statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.2.
Zero-crossing and period-amplitude analysis (PAA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.3.
Hjorth parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.4.
Detrended uctuation analysis (DFA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.
Spectral features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1.
Non-parametric spectral estimation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2.
Coherence analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.3.
Parametric spectral estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.4.
Subspace methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.5.
Higher-order spectral analysis (HOSA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Corresponding author at: University of Southampton, Room 3026, Building 13, Higheld, Southampton SO17 1BJ, UK. Tel.: +44 2380592873.
E-mail addresses: Smf1g08@Soton.ac.uk (S. Motamedi-Fakhr), M.M.Torbati@soton.ac.uk (M. Moshre-Torbati), M.Hill@soton.ac.uk (M. Hill),
C.M.Hill@soton.ac.uk (C.M. Hill), Pwr@isvr.soton.ac.uk (P.R. White).
1746-8094/$ see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.bspc.2013.12.003
22
22
23
23
23
24
25
25
25
25
25
25
25
25
25
26
26
26
22
S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133
3.3.
4.
5.
Timefrequency features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1.
Short time Fourier transform (STFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2.
The wavelet transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.3.
Matching pursuits (MP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.4.
Empirical mode decomposition (EMD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.
Nonlinear features/complexity measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.1.
Fractal dimension (FD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.2.
Correlation dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.3.
Entropy measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.4.
Lyapunov exponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Feature classication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.
Neural network (NN) classication (supervised learning) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1.
Multilayer perceptron (MLP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.
Clustering (unsupervised learning) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1.
Self-organising maps (SOM) or Kohonen maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.
Statistical classication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1.
Linear discriminant analysis (LDA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2.
Support vector machines (SVM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.3.
Hidden Markov model (HMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.
Fuzzy classication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.
Combined classiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary and conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conicts of interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix A.
Supplementary data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1. Introduction
Sleep is a crucial part of everyday life. It directly affects our cognitive performance, learning capabilities, and general physical and
emotional well-being. Sleep is the primary activity of the brain
in infancy and is thought to be a factor in neural plasticity [1,2].
Sleep problems in early life may result in lasting neurocognitive
decits. Krueger et al. [3] point out that during sleep one gives up
the opportunities to reproduce, eat, drink or socialise and one is
subject to predation. Sleep could only have evolved despite these
high evolutionary costs if it serves a crucial, primordial function.
Understanding and measuring brain activity in sleep is an exciting
frontier of neuroscience, and polysomnography (PSG) provides a
data-rich source for understanding sleep in both health and disease. PSG combines multiple signals in sleep typically including
neurophysiological signals:
EEG (usually 48 channels), EOG and EMG (submentalis and/or
tibialis muscle)combined with cardiorespiratory signals such as:
ECG
Oxyhaemoglobin saturation
Oral-nasal air ow
Abdominal and thoracic excursions
Visual inspection of these neurophysiological signals forms the
basis for standard sleep staging [4]. Signal processing allows the
extraction of detailed information from such signals. Applications
of these methods in relation to sleep EEG range from simple time
and frequency domain analysis to implementation of sophisticated nonlinear pattern recognition and classication algorithms.
Kubicki et al. [5] emphasise that going beyond the well-known and
commonly used Rechtschaffen and Kales scoring criteria [4] will
not be possible without the use of signal processing techniques
and computer aided analyses to reveal further information on the
microstructure of sleep. The body of literature developed for the
analysis of sleep EEG is vast and therefore this review paper provides a synthesis of a selection of this literature to generate an
overview of signal processing techniques applied to human sleep
EEG analysis and their relative merits.
26
26
26
26
27
27
27
27
27
28
28
28
29
29
29
29
29
29
29
30
30
30
30
30
30
The characteristic of the PSG signals to be analysed is rather

challenging. The underlying signals are inherently non-stationary
and the relationships between the different measurement channels
maybe time-varying. Many of the physical processes giving rise to
the observed signals are nonlinear in nature with the result that
the measured signals exhibit non-Gaussian statistics. To further
confound the problem many signal components of interest may
be observed in the presence of contaminating noise at a comparatively poor signal noise ratio. The standard principled approach to
the development of signal processing algorithms is generally based
on modelling the underlying processes and using that model to
then develop optimal algorithms. As is common in many biomedical applications, this approach ounders due to a lack of realistic
signal models and so a more heuristic approach is ourished leading
to the plethora of techniques observed in the eld.
Signal processing techniques will be considered in three sections: pre-processing, feature extraction and feature classication,
which constitute the basic underlying tasks in the automated
analysis of sleep signals. Each section describes its most frequently reported signal analysis methods. An extended Taxonomy
Table which summarises the sleep EEG related applications of
the surveyed techniques, containing a signicantly greater selection of references has also been collated for each section and
may be found in the accompanying supplemental materials. A
nal supplementary Taxonomy Table re-categorises the addressed
signal processing techniques based on their applications in wellestablished areas of sleep research such as obstructive sleep apnoea
(OSA) diagnosis and automatic sleep staging. Studies included in
this survey are limited to surface EEG signals in sleeping humans
(including the paediatric population). In short, this review aims to
synthesise the complex eld of sleep EEG analysis to inform both
the signal processing and the sleep research communities.
2. Pre-processing
There are several objectives when pre-proposing of PSG signals
including: normalisation, calibration, detrending and equalisation,
but these are aspects which are common to many signal processing
applications and require only the most basic of processing strategies. The following focuses on two rather more specic aspects of
pre-processing which are common to many EEG analysis problems

and are more widely encountered in other biomedical applications,
specically artefact removal and segmentation.
Note that we do not aim to give mathematical details of artefact
handling techniques here, rather the goal is to provide an overview
of applications of signal processing techniques in the analysis of
contaminated sleep EEGs. An extended list of the techniques and
their applications in sleep EEG artefact processing are provided in
Table 1 (supplementary materials).1
2.1. Artefact processing
EEG recordings, and in particular sleep EEG recordings, typically suffer from various types of artefacts. In biomedical signal
processing, artefacts are unwanted patterns not caused by the
underlying physiological event of interest. Thus, depending on the
purpose of the analysis, judgement must be made as to what is,
and what is not, an artefact. For reliable analysis special attention
should be given to identifying artefacts which if ignored, may significantly inuence the results and the consequent conclusions. Sleep
EEGs are often recorded in conjunction with other physiological
signals which may interfere with the measured sleep EEGs. Such
artefacts include, but are not limited to:
ocular artefacts that can occur due to eye movements (slow or
fast) which affect the electrical eld of the cornealretinal dipole
muscle artefacts (EMG interference) which occupy a broad frequency range and can appear in the form of spikes or continuous
interference
electrical eld changes induced by the cardiac muscle depolarisation which interfere with EEG signals (ECG interference)
Other types of artefact include:
mains (power-line) interference at 50 or 60 Hz depending on local
standards
head, body and chest movements
changes in electrolyte concentration at electrodes (e.g. sweat
artefacts)
Methods for processing artefacts in sleep EEG have been
reviewed in a general context in [6] whereas the specic application to infants and newborns (which are subject to similar forms
of artefacts) are considered in [7]. Fig. 1 illustrates sections of sleep
EEG contaminated by typical examples of artefacts.
Once a portion of data corrupted by an artefact has been successfully identied then there are various strategies which can be
adopted, depending on the form of that artefact. In extreme cases
the entire epoch which contains the artefact may have to be discarded, more commonly an artefact detector is employed and the
data which is contaminated by the artefact is identied and discarded. Alternatively, in some instances, it is possible to estimate
the underlying EEG using suitable signal processing techniques by
suppressing the artefact.
2.1.1. Artefact detection and rejection
The simplest approach to removing artefacts is to detect periods
when they are present and to reject that data. One approach is
1
Note that Tables 13 in the supplementary materials are designed to be
self-contained; each table includes the names of the relevant signal processing techniques (sorted based on frequency of use in the literature), a brief description of each
technique and an accompanying key reference, and instances where each technique
has been employed in sleep EEG analysis followed by the corresponding references.
We recommend the reader to use the supplemental materials provided alongside
this article.
23
to regard an artefact as any signicant deviation from the norm

(e.g. a sudden change in sleep EEG amplitude which lasts for
2 s), allowing detection to be performed by seeking changes (nonstationarities) in the measured signal. Care needs to be exercised
with this approach since the underlying EEG is itself non-stationary,
so algorithm parameters need to be selected so that only artefact
related non-stationarities are detected. Energy operators can be
useful markers of sudden changes (e.g. spikes) as they are sensitive
to instantaneous uctuations [8] but can lack sensitivity to more
subtle changes in the signal spectrum. Alternatively one can use an
autoregressive (AR) model of the signal within a Kalman lter setting to predict future samples of the time series and examine data
for signicant deviations from these predictions [9].
2.1.2. Methods for suppressing artefacts
Frequency selective lters (low-pass, high-pass, band-pass,
band-stop) have been almost universally used in artefact
processing, e.g. for muscle artefact removal [10]. Their success
requires that the artefact and EEG occupy distinct frequency bands,
a requirement which is rarely satised in practice. For example,
conventional frequency selective ltering cannot eliminate EMG
artefacts due to their broad spectrum (10200 Hz [11]).
One approach to artefact reduction can be adopted if a reference
measurement of the contaminating artefact signal can be obtained,
yielding a dual channel rejection scheme. By processing the EEG and
the reference signal one can attempt to remove the artefact. This
can be achieved in a variety of ways including using time domain
regression, Wiener lters, frequency domain regression or adaptive ltering (note that frequency domain regression is equivalent
to Wiener ltering if the signals are zero mean, for example if a
high pass lter has been applied [12]). The performance of such
approaches is limited by the quality of the reference measurement
and cross-contamination by the EEG signal of interest provides an
absolute limit on performance [13]. Regression analysis and Wiener
ltering have been used for removing EMG artefacts in high resolution studies [14,15] and adaptive ltering used to remove EOG
from general EEG signals boundaries [16]. Frequency selective ltering is applied on each channel independently and the above
dual channel methods exploit a dedicated reference channel. An
alternative approach is offered by ICA (independent component
analysis) which exploits the multichannel character of most EEG
signals to decompose the data into a set of random variables which
are maximally independent [17]. This decomposition is regarded as
representing the source signals which underlie the measured data
set. The electromagnetic nature of EEG signals means that these signals full the assumptions of the simplest form of ICA algorithm,
i.e. ones based on an instantaneous mixture model. Consequently
EEG analysis in general has been a fruitful application area for ICA
[18] and this has included application to sleep EEG. The methods
rely upon the idea that the artefact and EEG can be regarded as
arising from different (statistically independent) sources. Thus by
estimating those source components one can, in principle, identify which components relate to the artefact and which relate to
the EEG. This allows the reconstruction of the sensor signals using
only the EEG related sources and hence eliminating the artefact.
A key problem with ICA is the robust and automatic identication/separation of components relating to EEG and artefact. Further,
if the signal is temporally segmented then the ICA decomposition
lacks consistency between segments and care needs to be exercised
when recombining the data across segment boundaries [16].
The primary use of ICA is in the removal of artefacts of an
electromagnetic character, e.g. ECG, EMG, EOG and mains interference. This is because successful ICA decomposition requires that
the artefact, as well as the EEG, to conform to the instantaneous
mixing condition. ICA has been frequently, and successfully, used
for sleep EEG artefact reduction and is particularly well-suited
24
Fig. 1. Common EEG artefacts. (A) 50 Hz mains interference appears as a thickened signal caused by superposition of 50 Hz mains waves on the EEGs. (B) Movement causes
a sudden and signicant deviation from the background EEG and (C) severe movement can clip the EEG. (D) ECG interference appears as a pulsed EEG, it occurs when the
pulses on the ECG are superimposed on the EEG. (E) Sweat artefact is a slow drift of the baseline EEG.
to high resolution studies where the number of EEG electrodes

exceeds the usual 48 channels conventionally used in PSGs. In
such instances the large number of channels allows for good
source separation, although the number of sources identied
correspondingly increases exacerbating the problem of source
identication.
There are a very large number of ICA algorithms that have
been proposed for solving ICA under the instantaneous mixing
condition. The methods differ in a variety of aspects but one of
the main distinctions is in the metrics used to assess statistical independence. Crespo-Garcia et al. [19] have compared the
performance of four such algorithms for EMG interference reduction.
It has been demonstrated that a computerised system is capable of detecting most PSG artefacts including EMG, ECG and EOG
interference in addition to sweat and high frequency artefacts [21].
The programme requires tuning (manually by an expert or automatically by self-learning) to be more or less sensitive to different
artefact types depending on the state of the sleep EEG signal.
2.2. Sleep EEG segmentation
The non-stationary character of PSG is inconsistent with many
signal processing methods which assume the underlying signals
are stationary. Such problems are encountered in a great many
signal processing applications and a widely adopted solution is to
segment the signal in the time domain into small sub-sections such
that within each the signal can be regarded as being approximately
stationary (a so-called assumption of quasi-stationarity). Rather
more pragmatically the analysis of PSG signals requires one to
identify the approximate temporal location of events, so that
naturally any processing scheme uses data corresponding to a
nite duration window, which even if not explicitly regarded as a
segment, can be considered as such.
Segmentation can be performed uniformly or non-uniformly
(e.g. adaptively). Uniform segmentation is the most widely used
segmentation routine carried out in sleep EEG pre-processing.
Dividing the signal into regular xed duration segments with
typical durations of 30, 20 or even 1 s. The choice of a suitable
segment length is always a compromise. Longer segment lengths
risk including a signicant degree of non-stationarity, which may
cause processing schemes to loose sensitivity or to generate misleading results. The advantage of longer segments however is
that, assuming the signal remains well approximated as stationary
within the segment; then the reliability of estimates from wellconstructed processing algorithms will be enhanced. The use of
longer segments will also usually increase the overall computational burden of the algorithm, since in most instances the
computation load is related to the data length in a super-linear
fashion. The choice of segment length is usually also inuenced
by the algorithm being used, some methods are more data hungry than others and require the use of longer segment lengths to
obtain reliable results even in truly stationary environments. It
should be noted that this conict can force one to adopt segment
lengths which exceed those over which the signal can be reasonably regarded as stationary and the resulting method can suffer
signicantly.
Adaptive (non-uniform) segmentation is a more sophisticated
approach to segmentation, it aims to examine the underlying signals to determine a suitable segment length, allowing the use
of longer segments when the signal is approximately stationary and shorter segments when there are rapid changes [20,21].
An advantage of this approach is that feature extraction after
adaptive segmentation will be more reliable since features are
extracted from homogenous epochs rather than constant length
ones [22]. In order to perform such a segmentation one needs to
be able to detect changes in the underlying signal structure, this
can be done using a wide variety of metrics including statistical and spectral measures [2326], auto-regressive (AR) modelling
[20], time varying AR models [27] (which allow for limited temporal variation within a segment) and various energy measures
[28,29].
Whilst adaptive segmentation offers the potential for improved
performance it suffers from two distinct disadvantages. The rst
is that it clearly imposes an additional computational burden; the
degree to which this is important depends on the application and
method adopted. The more fundamental issue is the rather circular
nature of the principle, specically, in order to detect boundaries
to perform the segmentation one needs to, in some manner, analyse the data within the segment. The analysis used as the basis of
the segmentation should be as powerful as the method used postsegmentation, or the segment boundaries will only occur when
there is a gross non-stationarity and more subtle features in the
data will be obscured. This leads to a vision where segmentation is
inherent in the whole processing strategy, rather than a separate
process as is traditionally the case, such approaches whilst conceptually feasible are not currently described and so are not discussed
further.
3. Feature extraction
Analyses of time series are often carried out by extracting
features from the signal of interest. Features can be dened as
parameters which provide information about the underlying structure of a signal. There have been numerous techniques applied to
sleep EEG signals for the purpose of feature extraction. Names,
descriptions and applications of these techniques in sleep EEG
processing are provided in supplementary Table 2. Note that in
general, studies employ more than one feature in their analyses
and hence more often than not, features are complementary. The
rest of this section briey describes some of the most frequently
used features and feature extraction techniques in sleep EEG analysis. A review of quantitative analysis techniques applied to EEG
signal can be found in [30].
3.1. Temporal features
Temporal features are characteristics obtained from the signal in
the time domain. Some of the more widely used temporal features
and the associated processing techniques are described below. For
specic applications of the features/methods described below in
the analysis of sleep EEG signals see supplementary Table 2 which
also contains a wider range of temporal features.
3.1.1. Instantaneous statistics
These are among the simplest features which can be derived
from a time series and are the most frequently used temporal features in sleep EEG analysis. Such measures consider each sample
as if it were a univariate process and overlook any correlations
between samples. These statistics include measures derived from
moments of the waveform including the mean absolute amplitude, standard deviation/variance, skewness and kurtosis, as well as
measures relating to the probability density function of the waveform, such as mode, median or the entropy.
3.1.2. Zero-crossing and period-amplitude analysis (PAA)
Zero-crossings are the points at which the waveform crosses
the x-axis. They are simple to compute and zero crossing rate
encodes frequency information, because data which is dominated
by high frequencies has a rapid zero crossing rate, whereas low
zero-crossing rates are associated with low frequency processes
[31,32]. The limited information offered by zero crossings can be
supplemented by also characterising the waveform between the
zero-crossing points, leading to PAA [33]. This approach can be
adopted within frequency bands to mitigate the effects of noise and
to reduce the issues associated with signals comprised of multiple
components [33], but this is not always necessary and the results
can compare favourably with spectral based approaches [34].
3.1.3. Hjorth parameters
The parameters are based on variance of the derivatives of the
waveform and have been used for some time to characterise EEG
waveforms [35]. There are three Hjorth parameters which are commonly dened (although one can readily extend the principle) to
describe activity, mobility (shape) and complexity of EEG signals
[35,36]. Being based on derivatives, the Hjorth parameters are sensitive to noise and hence the signal of interest is commonly ltered
prior to calculation of these parameters [37].
3.1.4. Detrended uctuation analysis (DFA)
DFA is a method to characterise long range temporal correlations in a time series [38,39] and can be used as a measure of
self-similarity [40]. It is based on identify trends in the signals variance when analysed with different block length and is inherently
25
suitable for the analysis of non-stationary, noisy signals such as the

sleep EEGs [41,42].
3.2. Spectral features
The most commonly extracted features from sleep EEGs are
the spectral features. They are essentially parameters which characterise the signal in the frequency domain. Sleep EEGs are
traditionally divided into ve frequency bands namely delta
(04 Hz), theta (57 Hz), alpha (812 Hz), sigma (1315 Hz) and
beta (1630 Hz). Many of the visual scoring criteria for sleep
EEG signals are based on these frequency bands, which explain
in part the widespread use of spectral features in the analysis of these signals. Most spectral measures are only meaningful
when the underlying signal is stationary which emphasises the
importance of prior segmentation of the EEG signals. This section describes some of the most frequently used spectral features
and their associated signal processing techniques and highlights
the important properties of each technique where appropriate.
Table 2 in supplementary materials provides an extended list of
spectral features and their application in the human sleep EEG
analysis.
3.2.1. Non-parametric spectral estimation methods
Non-parametric methods for spectral estimation are the most
general form of spectral analysis, since they do not rely upon an
underlying model of the signal being analysed. They are based on
the Fourier transform, typically implemented via a Fast Fourier
Transform (FFT) algorithm. Welchs method (or segment averaging)
is, for instance, a well-known non-parametric power spectral density (PSD) estimator [43]. Non-parametric spectral estimators have
been by far the most frequently used techniques in sleep EEG analysis, largely because they are simple to implement and interpret.
These methods all inherently require one to compromise between
frequency resolution and variability in the estimated spectrum:
with the limited length of data available in a data segment achieving a reasonable compromise can be a challenging issue. Advice on
the use of the FFT for sleep EEG analysis is given by Campbell [44].
Once the spectrum has been estimated one may seek to reduce
the spectrum further to extract features from it, rather than using
the spectrum as a feature set. For instance the spectrum can be
summed in bands to yield estimates of energy in each of the
standard sleep EEG bands or one might look for spectral peaks.
Alternatively one might compute values like the spectral entropy
(SEN) [45] which characterises spectral shape, being a minimum
for a at (white) spectrum and a maximum for a single peak.
3.2.2. Coherence analysis
When multiple signals are measured cross-spectral analysis
becomes a powerful tool. The coherence is a normalised form of
the cross-spectrum and can be thought of as frequency domain
correlation [46,47]. It reects the degree of synchrony between
frequency components of two signals and can provide estimates
of functional connectivity in the brain [15,48]. A specic example
of its application in sleep EEG is as an indicator of sleep spindles,
which gives rise to a high coherence between two EEG channels
at about 13 Hz. Coherence analysis is most commonly used in high
resolution studies where a greater number of EEG channels are used
[49,50].
A related approach is the Directed Transfer Function (DTF)
method (also known as directed coherence) which provides information about causation (which coherence lacks) and so is suitable
for investigating functional connectivity in different brain regions
[51]. DTF is sensitive to phase shifts between signals but robust in
the presence of noise [48].
26
3.2.3. Parametric spectral estimation

The use of a model-based approach to spectral estimation is
attractive in that it has the potential to allow accurate spectral
estimates to be constructed with relatively short data lengths,
assuming the data conforms to the assumed model. The spectral
estimation problem then becomes one of estimating the model
parameters. The models employed are typically that of a digital lter excited by white noise. The different form of the assumed digital
lter gives rise to the different form of spectral estimator, in particular, methods based on auto-regressive (AR) modelling are popular
since the solution of the resulting parameter estimation problem is
a linear system of equations. One advantage of this approach is that
the model parameters (or parameters extracted from the model
such as reection coefcients or partial correlation coefcients) are
a natural set of features to represent the signal. A critical issue in
parametric modelling is the choice of the model order, i.e. the number of model parameters [52]. Whilst Akaikes information criterion
(AIC) [31,53] has been frequently used to determine the model
order many alternative methods exist [54]. A review of parametric
modelling techniques applied to EEG analysis can be found in [55].
3.2.4. Subspace methods
Subspace methods (also referred to as high resolution or super
resolution techniques) are a form of parametric spectral estimation, although strictly the output typically should not be regarded
as a spectrum. They are based on an assumption that the signal
consists of sinusoids in white noise and exploit the Eigen-structure
of the resulting correlation matrix [56]. Probably the most famous
method in this class is the Multiple Signal Classication (MUSIC)
algorithm [57]. In most EEG applications the underlying model for
these methods (sinusoids in noise) is not well matched to practical
EEG signals, consequently these approaches have received limited
attention in this eld, but subspace methods have been applied to
the analysis of sleep spindles [58].
3.2.5. Higher-order spectral analysis (HOSA)
The principles behind power spectral analysis have natural
extensions to higher orders [59]: power being a second order property of the signal in that it is the average of the signal squared, i.e.
the second moment. The power spectrum is a complete representation of a Gaussian process, whereas for non-Gaussian signals the
power spectrum is incomplete. The non-Gaussian nature of PSG
signals suggests that HOSA potentially provides new information
regarding the underlying processes. The use of HOSA allows one
to reveal phase coupling between different frequency components,
suggestive of non-linear generation mechanisms, something which
has been realised for sleep EEG [60]. Furthermore, measures based
on third order (bispectral) analysis have been used in estimating
the depth of anaesthesia (and sleep) from EEG signals [5658].
A signicant problem when applying HOSA is that they require
considerable quantities of data in order to obtain good estimates.
For example when estimating the bispectrum, which is the lowest
order HOSA method, it is recommended one uses N2 samples [61],
where N is the number of points in the FFT, so to compute a bispectrum with a 512 point FFT one should use approximately a quarter
of a million samples, or 1000 s of data at a 250 Hz sample rate.
3.3. Timefrequency features
Timefrequency analysis is a powerful tool which allows
decomposition of signals into both time and frequency [62]. It thus
provides a means for analysing signals which are non-stationary,
such as sleep EEGs. In the analysis of such signals one is often
interested in the evolution of the frequency content with time.
This is particularly important in the analysis of sleep EEGs where
many of the events (e.g. arousals, sleep spindles, alpha intrusions)
are manifested by sudden changes in amplitude and frequency

characteristics. Some of the more commonly used timefrequency
methods in the analysis of sleep EEGs are highlighted below, with
more details provided in supplementary Table 2.
3.3.1. Short time Fourier transform (STFT)
Is the simplest form of timefrequency analysis [62]. Often one
only considers the squared magnitude of the STFT and this squared
magnitude is termed the spectrogram. To compute the STFT the
signal of interest is uniformly segmented into many short duration
overlapping portions, the data in each portion are then windowed
and Fourier transformed. The result is a set of Fourier transforms
at different points in time revealing how these spectral properties change from one segment to another; that is the evolution of
frequencies with time. The timefrequency resolution of STFT is
directly determined by the segment size: the smaller the segment,
the higher the time resolution and, the lower the frequency resolution. By increasing the size of the segments one can increase
the frequency resolution at the cost of reduced time resolution.
Also note that longer segments may violate the quasi-stationarity
assumption required by the Fourier transform. Hence, one should
take into account issues about time and frequency resolutions in
addition to stationarity of the signal prior to the analysis. The STFT
has been widely used in the analysis of sleep EEG mainly due to its
simplicity and ease of implementation.
3.3.2. The wavelet transform
The wavelet transform is closely related to the STFT. Whilst STFT
can be regarded as representing a signal as a set of windowed
sinusoids of different frequencies, the wavelet transform represents a signal using a function which is scaled and shifted in time.
The scaling factor and time shifts can then be translated into frequency and time parameters respectively [63]. The character of the
wavelet transform implies that it uses variable size windows to
achieve timefrequency decomposition; short duration functions
representing high frequency components and long duration function representing low frequencies. This more closely matches the
character of many signals observed in the real world and means the
wavelet transform can offer signicant advantages over the STFT
[64]. On the other hand, an orthogonal discrete wavelet transform
[65] is generally not time shift invariant, that is, different time shifts
in the input do not result in time shifted version of the decomposition but a different decomposition, which may limit its use in
certain applications [66].
3.3.3. Matching pursuits (MP)
MP is a more recently developed timefrequency analysis
method. It is based on signal description via a collection of mathematical functions (commonly Gaussian modulated sinusoids)
called dictionaries. An advantage of MP is the large dictionary size
which is not limited to a certain form of function (as opposed to the
Fourier transform which uses only sinusoids or the wavelet transform which employs a mother wavelet function) [67]. MP achieves
timefrequency decomposition by nding the best matches that
t the structure of the signal from the dictionary. Parameterisation of the identied matches in time, frequency, amplitude and
energy results in a complete decomposition of the signal [68].
Analyses which employ MP benet from high timefrequency
resolution, accurate transient event description and appropriate
characterisation of non-stationarities. Furthermore, parameterisation of sleep EEGs using MP is compatible with the conventional
visual scoring criteria. An addressed issue of the MP algorithm in
analysing sleep EEGs is the statistical bias which is caused by the
structure of the employed dictionary. This can be alleviated by use
of stochastic dictionaries [69,70]. A possible shortcoming of the
method is its high computational cost which may limit its use in
real-time applications. In short, since its introduction, MP has been

widely used in the analysis of transient patterns in sleep EEGs; its
high timefrequency resolution makes MP an ideal candidate for
estimating the dominant frequency of sleep spindles [71].
3.3.4. Empirical mode decomposition (EMD)
EMD is a heuristic decomposition technique which provides a
signal representation which does not rely upon an assumption of
stationarity. In this method, the signal is broken down into basis
functions (called the intrinsic mode functions IMFs) which have
distinct oscillatory modes. The modes are constrained to have a
well-dened instantaneous frequency and these instantaneous frequencies are then computed utilising the Hilbert transform. The
calculated instantaneous frequencies can then be combined to produce an energytimefrequency representation of the signal. This
timefrequency plot is known as the Hilbert spectrum [72]. Use of
EMD in sleep EEG analysis has not been very wide, even though the
method offers the advantage of being well-suited to the analysis
of non-stationary signals. This might be a consequence of the fact
that EMD naturally represent signals as a combination of amplitude
and frequency modulated narrow-band components, the sources in
sleep EEG do not, in most instances, give rise to such components,
so the EMD representation does not allow ready interpretation in
terms of components which are readily associated with physiological sources.
To clarify the concept of timefrequency analysis further, Fig. 2
shows the timefrequency plots of the rst trend shown in Fig. 1
(50 Hz mains interference artefact). Short time Fourier transform
and Hibert spectrum of that signal are shown below.
3.4. Nonlinear features/complexity measures
It is traditionally assumed that EEG signals are generated from
stochastic processes and hence statistical methods can characterise
them. An alternative view suggests that EEG signals may be generated from a deterministic nonlinear process [73]. Since nerve cells
are highly nonlinear in nature, this view is intuitively appealing
[74]. In fact, Fell et al. [75] have shown that EEG signals are highly
unlikely to be fully characterised by linear stochastic models.
A nonlinear dynamical system is typically described by its states
(the system variables) which evolve according to a system of nonlinear equations governing the dynamics. The space which these
variables span is known as the phase space (or the state space) and
this contains all the possible states that the system can produce.
An important concept in nonlinear dynamical systems is the
attractor. An attractor is a subset of the phase space which, given
enough time, the system tends to evolve towards [76]. In systems
with multiple attractors, the initial conditions determine which
attractor the system eventually approaches. Attractors have properties which can be estimated and used as descriptive features. For a
comprehensive review on nonlinear dynamical analysis of EEG signals, see Stam [77]. Much work on nonlinear systems concentrates
on extracting information about the behaviour of the system given
knowledge of the governing equations. Nonlinear systems encompass a wide range of behaviours and consequently analytical tools
often seek to provide values which in some sense summarise the
behaviour of the system, usually in the vicinity of attractors (since
those attractors dene the asymptotic behaviour of the system).
In signal analysis problems the dynamics of the nonlinear system
is rarely known a priori and one usually seeks to approximate the
summary statistics, or features, directly from the data.
Nonlinear features have been widely applied to the analysis of
sleep EEGs. They can provide complementary information to characterise specic waveforms as well as different stages of sleep. An
important issue in nonlinear analysis of time series is the reliability
and interpretability of the potential results. Successful application
27
and interpretation of these techniques requires a good understating of the method and the application [74,78]. However, careful
analysis of nonlinearities can reveal useful information which is
otherwise hidden [79]. The rest of this section briey describes
some of the nonlinear features/techniques which have been more
frequently applied to the analysis of sleep EEGs.
3.4.1. Fractal dimension (FD)
The FD directly measures the complexity of the measured signal. The basic idea comes from quantication of dimensionalities
of fractals (i.e. geometries which are self-similar on different scales
[80]). A fractal is an object with a non-integer dimension. Familiar geometric objects have integer dimensions: curves have a
dimension of one, whilst planes are two dimensional. Fractals are
mathematical objects whose dimension is non-integer; for instance
an object of fractal dimension between one and two might be a line
of innite length which is contained within a nite area. The concept of FD has been expanded to the analysis of time series, the
principle being that a simple time-series would have a lower fractal dimension than a more complex one. For a single-channel sleep
EEG signal, FD can range from one to two, that is, its dimension is
at least one and cannot be greater than two [81]. There are several algorithms for the calculation of FD from a time series [82,83]
which have been used in the analysis of sleep EEGs to date. FDs are
suitable for detection of transients in EEG signals [84,85] since they
can be applied to short segments of data and are relatively stable
measures of complexity [86].
3.4.2. Correlation dimension
The correlation dimension provides a bound on the fractal
dimension of the attractor of the underlying dynamical system [87].
The dimension of the underlying attractor is fractal (non-integer) if
the system is chaotic. The correlation dimension is commonly considered for use in the analysis of sleep EEGs [88], this relatively wide
usage is, in part, a consequence of an efcient and straightforward
numerical algorithm for its estimation [87]. The correlation dimension has been applied to the analysis of both neonatal [89,90] and
adult [91,92] sleep EEGs and has, arguably, been most successful
used for sleep staging. Its use in identifying the different stages of
sleep is based on the observation that in deeper sleep the measured
complexity tends to be lower [45,84,89,90,93]. Accurate estimation
of correlation dimension requires a large sample size (large segments) which limits its applicability; it is therefore generally not
suitable for parameterisation of short transient events such as sleep
spindles or arousals. Extensions of the approach to non-stationary
signals do exist [94] but have not been widely adopted.
3.4.3. Entropy measures
Entropy is a statistical measure of complexity and does not rely
upon a nonlinear description of the data. There are a variety of denitions of entropy measures which have been proposed. Estimating
the entropy directly from the time-series requires computation of
the joint probability density function between time series, a process
which is data intensive, many entropy measures therefore attempt
to approximate this quantity in an efcient manner. Approximate
entropy (ApEn) [95] is one such widely considered method. ApEn
reects the conditional probability that two time series remain similar to each other for the next m samples, given that they have
previously been similar. If the signals have a high degree of regularity (i.e. low degree of complexity) it is more likely that they will
remain similar for subsequent samples and hence they produce
a low ApEn value [9698]. ApEn has several desirable properties
[108]: it is robust in the analysis of short data segments, resilient
to outliers and strong transients, is capable of dealing with noise
by appropriate estimation of its parameters and can be applied to
both stochastic and deterministically chaotic signals.
28
Fig. 2. Illustration of two timefrequency analysis methods. The EEG signal used here is the one from Fig. 1 highlighting 50 Hz main interference. The spectrogram clearly
shows the 50 Hz noise and its rst harmonic at 100 Hz, the low frequency activity around 17 s mark is also emphasised. Similarly, the Hilbert spectrum shows an activity
around 50 Hz localised in both time and frequency, the low frequency activity around the 17 s mark is also highlighted.
The precise length of data for reliable estimation of ApEn

depends on the systems complexity [96], with values of at least
1000 samples being recommended for some EEG applications.
However, ApEn is a biased estimate of complexity and the alternative, closely related idea of sample entropy (SampEn) [99] has been
suggested. It is claimed that SampEn is consistent, has reduced bias
and is largely independent of sample sizes [99]. Both methods have
been used in the analysis of sleep EEGs, for instance, SampEn has
been used to capture and reect the distinctive characteristics of
sleep in different stages [100].
3.4.4. Lyapunov exponents

The Lyapunov exponents represent the average convergence
or divergence rate of trajectories in phase space. These exponents can be positive, zero or negative and their interpretation
depends on their sign [77,101,102]. Positive Lyapunov exponents
are taken as an indication of chaos. There are as many Lyapunov
exponents as there are dimensions in the underlying dynamical
equations. Knowledge of the system equations allows one to compute Lyapunov exponents directly, but in practice these equations
are not available in EEG analysis problems and the exponents must
be estimated from time series data [103]. Commonly only the
Largest Lyapunov Exponent (LLE) is considered, since if it is positive then the underlying dynamics will be chaotic [104]. Whether
it is suitable to model sleep EEGs as chaotic processes is a matter of contention; positive Lyanpunov exponents can be found in
sleep EEGs [74], but such observations can be the consequence
of alternative interpretations [79,92]. Despite this, the Lyapunov
exponents provide information which characterises the attractor
in the dynamical system; a feature which may be useful in understanding the signal. This parameterisation has been used for sleep
staging [45] and as a diagnostic feature [105].
4. Feature classication
Features are measurable characteristics of a time series used
to reduce the signals dimension whilst maintaining information
vital to subsequent operations, e.g. classication. Once features are
extracted from a signal one can perform classication by grouping
the features; that is dividing the feature space into a discrete number of regions: one for each class to be classied. For instance, in
sleep staging we may have ve possible classes namely, wake, stage
1, stage 2, stage 3 sleep and REM, or in automatic spindle detection,

we may have two, spindle vs. non-spindle.
In classication, features within each category (or class) share
some similarity and maybe classied based on some measure
of that similarity. In sleep EEG analysis we commonly have to
deal with sleep micro and macro structural classication problems. Classication can be undertaken for a diagnostic purpose,
e.g. classifying patients with obstructive sleep apnoea vs. controls, or can aid automation (classication of transient events in
sleep or classifying sleep stages). Numerous techniques have been
proposed and used for sleep EEG signal classication. Table 3 in
supplementary materials is dedicated to these methods and their
specic applications in the analysis of sleep EEGs. As with the previous sections, we start with the most frequently used techniques
addressed in the literature. Information about the less frequently
used methods can still be found in Table 3 (supplementary
materials).
4.1. Neural network (NN) classication (supervised learning)

Neural networks or articial neural networks (ANN) were originally developed as the result of attempts to understand how the
human brain learns and solves problems. They are mathematical
models inspired by neuronal interactions in the brain and can be
used to model a wide range of complex systems. An ANN commonly consists of an input layer, an output layer and a number of
hidden layers in between. Layers are composed of articial neurons (or nodes). Neurons linearly combine their weighted inputs
and produce an output based on their nonlinear output function.
Using several neurons in conjunction, ANNs become capable of
modelling very complex nonlinear systems [114]. Parameters of
an ANN are generally the weights by which the inputs are scaled.
These weights are obtained through training (or learning), that is,
adaptively updating the weights in the ANN in response to inputs
to obtain the known correct output. For example, for classication
of sleep stages, one can use a previously scored polysomnogram
with known sleep stages to train an ANN after which the trained
ANN should classify new polysomnograms [106].
ANNs are adaptive and can deal with classication problems
which are not linearly separable. In other words, decision boundaries generated by ANNs are not constrained to be lines. ANNs have
been very frequently used in the analysis of sleep EEGs [107109].
ANNs are generally computationally efcient for classication of
new patterns (feedforward conguration). However, training them

with back-propagation can be computationally intensive.
There are many different types of ANNs with different architectures and training algorithms; some of the most commonly used
are briey described below. General information regarding ANNs
can be found in [110] and a review of ANNs in the context of EEG
analysis is provided in [111]. Table 3 in supplementary materials provides an extended list of ANNs and highlights their specic
applications in the analysis of sleep EEGs.
4.1.1. Multilayer perceptron (MLP)
An MLP is one of the earliest forms of ANN. It consists of an
input layer, one or more hidden layers and an output layer [112].
The nonlinear output (or activation) function of neurons in the
hidden layer(s) usually follows a sigmoid function (a smooth s
shaped graph which is differentiable everywhere) [110,113]. Given
sufcient neurons in the hidden layers, MLPs can be used to approximate any continuous function arbitrarily well [114]. Thus, MLPs
(and ANNs in general) benet from considerable exibility in classication. However, this exibility may also result in sensitivity
to overtting [115]. MLPs are widely used in the analysis of sleep
EEGs for a variety of purposes such as sleep staging, transient event
detection and obstructive sleep apnoea (OSA) diagnosis. Providing
that there is enough reliable data for training, the MLP remains an
effective classier.
4.2. Clustering (unsupervised learning)
Cluster analysis, or clustering, is the process of dividing a set
into natural homogenous subgroups where elements within each
subgroup are similar to each other and different from those within
other subgroups [116]. The simplest and the most widely employed
clustering algorithm is the k-means. This algorithm assumes that
the feature space consists of k clusters (this can be decided beforehand based on knowledge of the nature of the data) whose centres
are randomly distributed. It then iteratively adjusts those random
centre points to get closer to the centre of the actual clusters. This
adjustment is done by minimising a squared error function. The
algorithm is computationally efcient and simple to implement.
Its downside however, is that due to random initialisation of the
centroids, clustering may result in error; that is the algorithm may
fail to detect the true clusters. Hence, in practice, clustering is
done with a number of sets of random initial points rather than
a single set. For the exact description of the algorithm and more
details, see Jain [117]. Due to its simple implementation, use of kmeans algorithm in the analysis of sleep EEGs has been relatively
frequent; Table 3 (supplementary materials) shows its specic
applications.
4.2.1. Self-organising maps (SOM) or Kohonen maps
SOMs are neural network based clustering methods for the analysis and visualisation of high dimensional data [118]. When dealing
with a classication problem, SOMs do not need prior knowledge
about the number of classes and they can achieve classication
through maximal separation of input features [119122]. SOMs
can yield satisfactory results with a comparatively small training
set and can also be signicantly faster than conventional MLPs for
exploratory classication problems [118].
The advantages mentioned above make SOMs a competitive
candidate in classication problems [123]. Their use in sleep EEG
analysis has been moderate but diverse, ranging from automatic
sleep staging [124] and sleep stage characterisation [125] to classication of transient events such as K-complexes [126], emphasising
the exibility of this technique.
29
4.3. Statistical classication

Statistical classication is another popular approach for classifying feature spaces. In this article, the term statistical classier is
used in a broad fashion to accommodate a range of diverse methods. Statistical classiers usually separate the feature space into
different classes by calculating the probability that a certain feature belongs to a given class, in other words, they are based on
probabilistic models [127] (note that SVMs are an exception in that
they do not rely on an explicit statistical model for the feature
distribution, but can be derived from within a broader statistical
framework).
The rest of this section briey describes some of the more
frequently used statistical classiers in the analysis of sleep EEG signals, for additional techniques and more details on the applications
see Table 3 in supplementary materials.
4.3.1. Linear discriminant analysis (LDA)
This is a form of linear classier which divides the feature space
into several classes by hyper-planes (planes which exist in one
lower dimension than the feature space, for instance, a two dimensional feature space can be separated into two by a line and a three
dimensional feature space by a plane) [127,128]. LDA, also known as
Fishers Linear Discriminant analysis, is simple to implement (particularly when separating only two classes), easy to interpret and
generally works well in classifying linearly separable data; however, there exist problems for which the error probability of LDA is
close to unity even when the data are linearly separable [129]. Further, the fact that LDA is constrained to problems in which features
are linearly separable severely limits its utility. Fishers quadratic
discriminant analysis is an extension to LDA and can separate problems based on quadratic class boundaries. LDA and its variants have
been used relatively frequently in the analysis of sleep EEGs, see
Table 3 for further applications of LDA in sleep EEG analysis.
4.3.2. Support vector machines (SVM)
The principle behind SVM is to nd separating boundaries
(hyper-planes) which optimise the spaces between two classes,
in particular it maximises the margin around the hyper-plane
[110,130]. Most practical SVMs map the current input feature vector into a higher dimensional feature space through some nonlinear
mapping and then exploit the use of optimal hyper-planes for
separating these nonlinear functions of the features [130]. The consequence is that whilst SVMs are linear classiers in the space
associated with nonlinear functions of the features, in the original feature space the decision boundaries are nonlinear. They are
exible (i.e. can be generalised to suit different needs), robust and
not sensitive to overtraining; they may however, suffer from high
computational expense [131]. Use of SVMs in sleep EEG analysis
has been moderate but inclusive of a relatively wide range of applications including but not limited to automatic sleep staging [37],
arousal detection [132] and sleep spindle recognition [133]. SVMs
have been frequently successful in classication problems and can
work well even in situations where the training data set is small
[132]. For a thorough tutorial on SVMs see Burges [134].
4.3.3. Hidden Markov model (HMM)
An HMM extends the concept of a Markov model (or Markov
chain model). A Markov process (Markov chain) is a stochastic
model which attempts to characterise a system by its possible states
and transition probabilities between those states at any one time.
If the states of the system are deterministic and visible, i.e. each
transition leads to a certain outcome, we have a Markov model,
however, if the states are not directly visible (hidden) but are
reected in the outcome (observation), we have a Hidden Markov
model (HMM) [135,136]. HMMs can be applied to the analysis of
30
non-stationary signals by dividing the signal into segments and

computing feature sets for each segment. As an example for the
analysis of sleep EEGs, one might regard the states as the discrete
sleep stages and the methods developed for HMMs can be used
to estimate model parameters, such as the transition probabilities,
and the likelihood of observing a particular state sequence [137].
HMM is a powerful tool for classication of time series, for example
they are almost ubiquitous in the eld of automatic speech recognition. They achieve classication by identifying the most likely
state sequence associated with an observed sequence of feature
vectors. They are also capable of dealing with nonlinearly separated
data [131]. They have been moderately used in sleep EEG analysis
and their use has been relatively focused on sleep staging related
applications.
The above three methods conclude most of the statistical classiers used in the analysis of sleep EEGs. Bayesian classiers have
also been employed but their use has not been very wide, see Table
3 (supplementary materials) for the details.
4.4. Fuzzy classication
Fuzzy sets were introduced by Zadeh in 1965 as an attempt
to mathematically describe ambiguous classes, those are classes
which do not have a well dened boundary such as the class of
beautiful women or the class of tall men [138]. Ever since its
introduction, the concept of fuzziness has become a practical and
widely used tool in pattern recognition and classication. Fuzzy
classiers do not assume absolute membership of a single class
and hence they are more suitable to real data where boundaries
between subgroups might not be well dened [139]. Fuzzy classiers are versatile; they have very good generalisation capabilities
and can be effectively integrated with other classication systems
for better performance (e.g. neuro-fuzzy classiers, fuzzy decision
trees, fuzzy k-nearest neighbour algorithm) [116].
Use of fuzzy classiers and their simple variations have been
frequent and successful in the analysis of sleep EEGs. Table 3
(supplementary materials) shows the instances where fuzzy classication or fuzzy based reasoning has been employed to analyse
sleep EEGs.
4.5. Combined classiers
Last but not least, combined classiers are, as the name suggests, classiers which are based on two or more conventional
classication routines. Although these are generally bound to be
computationally less efcient, they tend to achieve more accurate
classication, hence they can be used for ofine analysis where
execution speed is not the main concern. A widely used combined
classier is the neuro-fuzzy classier (NFC). NFCs are particularly
good at approximating nonlinearities; however, their complexity
grows exponentially with additional inputs, this is referred to as
the curse of dimensionality [140].
For more detailed information on most of the classication procedures mentioned in this section, the authors refer the interested
readers to the review articles by Lotte et al. [131] and Jain et al.
[117]. Table 3 (supplementary materials) also provides a comprehensive list of the above (and additional) classication techniques
and their applications in sleep EEG analysis.
5. Summary and conclusion
This review article provides an overview of signal processing
techniques applied to the analysis of sleep EEG signals in both
paediatric and adult populations. The analysis has been broken
down into three main parts: pre-processing, feature extraction and
feature classication. Each of these is supported by supplementary

materials:
Table 1 provides further details of pre-processing techniques,
Table 2 serves as a guide to aid sleep researchers choose the
appropriate signal processing techniques for their application of
interest and provides brief descriptions of signal processing techniques in addition to their existing applications in the analysis of
sleep EEG signals,
Table 3 provides an overview of classication techniques as an
aid to researchers in selecting appropriate algorithms, and
Table 4 summarises the more frequently addressed topics of
sleep EEG analysis, splitting them into nine main categories
(sleep staging, sleep stage characterisation, OSA diagnosis, cyclic
alternating pattern detection, spindle detection/analysis, arousal
detection/analysis, other transient detection/analysis, cortical
interactions in sleep and other exploratory analyses) and providing details of signal processing techniques which have been
employed in their analysis. Utilising this table, one can conveniently navigate through the article and quickly identify the topic
of interest (e.g. paediatric sleep stage characterisation or sleep
spindle analysis).
Many relevant areas of sleep research continue to generate new
and interesting ndings utilising biosignals such as EEGs. High temporal resolution of EEGs has the potential to be an excellent tool in
understanding activity of the sleeping brain. The introduction of
functional magnetic resonance imaging (fMRI) in recent years and
the possibility of performing combined EEG/fMRI studies open new
possibilities in understanding functional dimensions of sleep EEG
[141].
Sleep research is increasingly reliant on advanced signal
processing, yet there seem to be few practically accepted standards
between the two communities. A bottleneck in the rapid growth
of sleep research is the subjective nature of manual sleep scoring
and the existing inter-scorer variability. There are now a number
of expert systems which can efciently mimic an expert human.
This variability could be reduced signicantly, if the denitions of
sleep EEG patterns and artefacts were universally accepted across
the disciplines. With more robust denitions and improved common protocols, computerised programmes can outperform human
scorers and potentially remove the subjectivity of sleep scoring.
Table 4 in supplementary materials not only helps in crossreferencing but also highlights potential gaps in sleep research. For
instance, there has been relatively little work done on cortical interactions in sleep, and more research in this area (i.e. neuroscientic
view of the sleeping brain) seems to be benecial. It can also been
seen that some sleep patterns (such as sleep spindles) are more
heavily researched than others which may not be clinically justied. In short, identifying and addressing these gaps will enable
sleep research to progress faster.
Conicts of interest
No conicts of interest.
Appendix A. Supplementary data
Supplementary material related to this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.bspc.2013.12.003.
References
[1] E. Touchette, D. Petit, J.R. Seguin, M. Boivin, R.E. Tremblay, J.Y. Montplaisir,
Associations between sleep duration patterns and behavioral/cognitive functioning at school entry, Sleep 30 (2007) 12131219.

[2] M.P. Walker, R. Stickgold, Sleep, memory, and plasticity, Annu. Rev. Psychol.
57 (2006) 139166.
[3] J.M. Krueger, D.M. Rector, S. Roy, H.P.A. Van Dongen, G. Belenky, J. Panksepp,
Sleep as a fundamental property of neuronal assemblies, Nat. Rev. Neurosci.
9 (2008) 910919.
[4] A. Rechtschaffen, A. Kales, A manual of standardized terminology, techniques
and scoring system for sleep stages of human subjects, U.S. National Institute
of Neurological Diseases and Blindness, Neurological Information Network,
1968.
[5] S. Kubicki, W.M. Herrmann, The future of computer-assisted investigation of
the polysomnogram: sleep microstructure, J. Clin. Neurophysiol. 13 (1996)
285294.
[6] P. Anderer, S. Roberts, A. Schlogl, G. Gruber, G. Klosch, W. Herrmann, P. Rappelsberger, O. Filz, M.J. Barbanoj, G. Dorffner, B. Saletu, Artifact processing
in computerized analysis of sleep EEG a review, Neuropsychobiology 40
(1999) 150157.
[7] E. Walls-Esquivel, M.F. Vecchierini, C. Heberle, F. Wallois, Electroencephalography (EEG) recording techniques and artefact detection in early premature
babies, Clin. Neurophysiol. 37 (2007) 299309.
[8] P. Hae-Jeong, J. Do-Un, P. Kwang-Suk, Automated detection and elimination
of periodic ECG artifacts in EEG using the energy interval histogram method,
IEEE Trans. Biomed. Eng. 49 (2002) 15261533.
[9] M. Rohalova, P. Sykacek, M. Koska, G. Dorffner, Detection of the EEG artifacts
by the means of the (extended) Kalman lter, in: I. Frollo, M. Tysler, A. Plackova (Eds.), Measurement Science Review 1 (1) (2001) 293296, Institute of
Measurement Science, Smolenice, Slovakia.
[10] D.P. Brunner, R.C. Vasko, C.S. Detka, J.P. Monahan, C.F. Reynolds, D.J. Kupfer,
Muscle artifacts in the sleep EEG: automated detection and effect on all-night
EEG power spectra, J. Sleep Res. 5 (1996) 155164.
[11] A. Boxtel, P. Goudswaard, L.R.B. Schomaker, Amplitude and bandwidth of the
frontalis surface EMG: effects of electrode parameters, Psychophysiology 21
(1984) 699707.
[12] J.C. Woestenburg, M.N. Verbaten, J.L. Slangen, The removal of the
eye-movement artifact from the EEG by regression-analysis in the frequencydomain, Biol. Psychol. 16 (1983) 127147.
[13] B. Widrow, J.R. Glover, J.M. McCool, J. Kaunitz, C.S. Williams, R.H. Hearn, J.R.
Zeidler, E. Dong, R.C. Goodlin, Adaptive noise cancelling principles and
applications, Proc. IEEE 63 (1975) 16921716.
[14] R.J. Davidson, EEG measures of cerebral asymmetry conceptual and methodological issues, Int. J. Neurosci. 39 (1988) 7189.
[15] R.T. Pivik, R.J. Broughton, R. Coppola, R.J. Davidson, N. Fox, M.R. Nuwer, Guidelines for the recording and quantitative-analysis of electroencephalographic
activity in research contexts, Psychophysiology 30 (1993) 547558.
[16] P. He, G. Wilson, C. Russell, M. Gerschutz, Removal of ocular artifacts from the
EEG: a comparison between time-domain regression method and adaptive
ltering method using simulated data, Med. Biol. Eng. Comput. 45 (2007)
495503.
[17] J.K.A. Hyvarinen, E. Oja, Independent Component Analysis, John Wiley & Sons,
2001.
[18] T.P. Jung, S. Makeiq, C. Humphries, T.W. Lee, M.J. McKeown, V. Iraqui, T.J.
Sejnowski, Removing electroencephalographic artifacts by blind source separation, Psychophysiology 37 (2000) 163178.
[19] M. Crespo-Garcia, M. Atienza, J.L. Cantero, Muscle artifact removal from
human sleep EEG by using independent component analysis, Ann. Biomed.
Eng. 36 (2008) 467475.
[20] G. Bodenstein, H.M. Praetorius, Feature extraction from the electroencephalogram by adaptive segmentation, Proc. IEEE 65 (1977) 642652.
[21] J.S. Barlow, Methods of analysis of nonstationary EEGs, with emphasis on segmentation techniques a comparative review, J. Clin. Neurophysiol. 2 (1985)
267304.
[22] K. Paul, V. Krajca, Z. Roth, J. Melichar, S. Petranek, Comparison of quantitative
EEG characteristics of quiet and active sleep in newborns, Sleep Med. 4 (2003)
543552.
[23] J.S. Barlow, Computer characterization of trace alternant and REM-sleep patterns in the neonatal EEG by adaptive segmentation an exploratory-study,
Electroencephalogr. Clin. Neurophysiol. 60 (1985) 163173.
[24] J.S. Barlow, O.D. Creutzfeldt, D. Michael, J. Houchin, H. Epelbaum, Automatic
adaptive segmentation of clinical EEGs, Electroencephalogr. Clin. Neurophysiol. 51 (1981) 512525.
[25] V. Krajca, S. Petranek, I. Patakova, A. Varri, Automatic identication of
signicant graphoelements in multichannel EEG recordings by adaptive
segmentation and fuzzy clustering, Int. J. Bio-Med. Comput. 28 (1991)
7189.
[26] M. Arnold, A. Doering, H. Witte, J. Dorschel, M. Eisel, Use of adaptive Hilbert
transformation for EEG segmentation and calculation of instantaneous respiration rate in neonates, J. Clin. Monit. 12 (1996) 4360.
[27] N. Amir, I. Gath, Segmentation of EEG during sleep using time-varying autoregressive modeling, Biol. Cybern. 61 (1989) 447455.
[28] R. Agarwal, J. Gotman, Adaptive segmentation of electroencephalographic
data using a nonlinear energy operator, in: ISCAS99: Proceedings of the
1999 IEEE International Symposium on Circuits and Systems, vol. 4, 1999,
pp. 199202.
[29] R. Agarwal, J. Gotman, Computer-assisted sleep staging, IEEE Trans. Biomed.
Eng. 48 (2001) 14121423.
[30] N.V. Thakor, S.B. Tong, Advances in quantitative electroencephalogram analysis methods, Annu. Rev. Biomed. Eng. 6 (2004) 453495.
31
[31] A. Isaksson, A. Wennberg, L.H. Zetterberg, Computer-analysis of EEG signals

with parametric models, Proc. IEEE 69 (1981) 451461.
[32] M. Carrozzi, A. Accardo, F. Bouquet, Analysis of sleep-stage characteristics in
full-term newborns by means of spectral and fractal parameters, Sleep 27
(2004) 13841393.
[33] B.A. Geering, P. Achermann, F. Eggimann, A.A. Borbely, Period-amplitude analysis and power spectral analysis: a comparison based on all-night sleep EEG
recordings, J. Sleep Res. 2 (1993) 121129.
[34] R. Armitage, R. Hoffmann, T. Fitch, C. Morel, R. Bonato, A comparison of period
amplitude and power spectral-analysis of sleep EEG in normal adults and
depressed outpatients, Psychiatry Res. 56 (1995) 245256.
[35] B. Hjorth, EEG analysis based on time domain properties, Electroencephalogr.
Clin. Neurophysiol. 29 (1970) 306310.
[36] A. Saastamoinen, E. Huupponen, A. Varri, J. Hasan, S.L. Himanen, Computer
program for automated sleep depth estimation, Comput. Methods Programs
Biomed. 82 (2006) 5866.
[37] S. Gudmundsson, T.P. Runarsson, S. Sigurdsson, Automatic sleep staging
using support vector machines with posterior probability estimates. International Conference on Computational Intelligence for Modelling, Control
and Automation, 2005 and International Conference on Intelligent Agents,
Web Technologies and Internet Commerce, in: M. Mohammadian (Ed.), IEEE,
Vienna, Austria, 2005, p. 6.
[38] C.K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E. Stanley, A.L. Goldberger, Mosaic organization of DNA nucleotides, Phys. Rev. E 49 (1994)
16851689.
[39] L. Jong-Min, K. Dae-Jin, K. In-Young, S.I. Kim, Analysis of scaling exponents
of waken and sleeping stage in EEG, in: J. Mira, A. Prieto (Eds.), Proceedings,
Part 1, 6th International Work-Conference on Articial and Natural Neural
Networks, IWANN 2001 Granada, Connectionist Models of Neurons, Learning Processes, and Articial Intelligence, Springer-Verlag, Granada, 2001, pp.
450456, Spain, June 1315.
[40] Y. Shen, E. Olbrich, P. Achermann, P.F. Meier, Dimensional complexity and
spectral properties of the human sleep EEG, Clin. Neurophysiol. 114 (2003)
199209.
[41] L. Jong-Min, K. Dae-Jin, K. In-Young, P. Kwang-Suk, S.I. Kim, Detrended uctuation analysis of EEG in sleep apnea using MIT/BIH polysomnography data,
Comput. Biol. Med. 32 (2002) 3747.
[42] S. Leistedt, M. Dumont, J.P. Lanquart, F. Jurysta, P. Linkowski, Characterization of the sleep EEG in acutely depressed men using detrended uctuation
analysis, Clin. Neurophysiol. 118 (2007) 940950.
[43] P. Welch, The use of fast Fourier transform for the estimation of power spectra:
a method based on time averaging over short, modied periodograms, IEEE
Trans. Audio Electroacoust. 15 (1967) 7073.
[44] I.G. Campbell, EEG recording and analysis for sleep research, Curr. Protoc.
Neurosci. (2009) (chapter 10, unit 10.12).
[45] J. Fell, J. Roschke, K. Mann, C. Schaffner, Discrimination of sleep stages: a comparison between spectral and nonlinear EEG measures, Electroencephalogr.
Clin. Neurophysiol. 98 (1996) 401410.
[46] G.M. Jenkins, D. Watts, Spectral Analysis and its Applications, 1968.
[47] R.B. Duckrow, H.P. Zaveri, Coherence of the electroencephalogram during the
rst sleep cycle, Clin. Neurophysiol. 116 (2005) 10881095.
[48] M. Kaminski, K. Blinowska, W. Szelenberger, Topographic analysis of coherence and propagation of EEG activity during sleep and wakefulness,
Electroencephalogr. Clin. Neurophysiol. 102 (1997) 216227.
[49] P. Achermann, A.A. Borbely, Coherence analysis of the human sleep electroencephalogram, Neuroscience 85 (1998) 11951208.
[50] P. Achermann, A.A. Borbely, Temporal evolution of coherence and power in
the human sleep electroencephalogram, J. Sleep Res. 7 (1998) 3641.
[51] M.J. Kaminski, K.J. Blinowska, A new method of the description of the
information-ow in the brain structures, Biol. Cybern. 65 (1991) 203210.
[52] R.E. Herrera, M.G. Sun, R.E. Dahl, N.D. Ryan, R.J. Sclabassi, Vector autoregressive model selection in multichannel EEG, in: Proceedings of the 19th Annual
International Conference of the IEEE Engineering in Medicine and Biology
Society, vol. 19, Pts 16 Magnicent Milestones and Emerging Opportunities
in Medical Engineering, IEEE, New York, 1997, pp. 12111214.
[53] H. Akaike, New look at statistical-model identication, IEEE Trans. Automatic
Control AC19 (1974) 716723.
[54] F. Gustafsson, H.k. Hjalmarsson, Twenty-one mL estimators for model selection, Automatica 31 (1995) 13771392.
[55] J. Pardey, S. Roberts, L. Tarassenko, A review of parametric modelling techniques for EEG analysis, Med. Eng. Phys. 18 (1996) 211.
[56] S.V. Vaseghi, Advanced Signal Processing and Digital Noise Reduction, WileyTeubner, 1996.
[57] A. Subasi, E. Ercelebi, A. Alkan, E. Koklukaya, Comparison of subspace-based
methods with AR parametric methods in epileptic seizure detection, Comput.
Biol. Med. 36 (2006) 195208.
[58] O. Caspary, P. Nus, Adaptive spectral analysis of sleep spindles based on subspace tracking, in: H. Boom, C. Robinson, W. Rutten, M. Neuman, H. Wijkstra
(Eds.), Proceedings of the 18th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, vol. 18, Pts 15, IEEE, New York,
1997, pp. 976977.
[59] C.L. Nikias, J.M. Mendel, Signal processing with higher-order spectra, IEEE
Signal Process. Mag. 10 (1993) 1037.
[60] K. Schwab, M. Eiselt, C. Schelenz, H. Witte, Time-variant parametric estimation of transient quadratic phase couplings during electroencephalographic
burst activity, Methods Inf. Med. 44 (2005) 374383.
32
[61] J.W.A. Fackrell, P.R. White, J.K. Hammond, R.J. Pinnington, A.T. Parsons, The
interpretation of the bispectra of vibration signals: I. Theory, Mech. Syst.
Signal Process. 9 (1995) 257266.
[62] L. Cohen, TimeFrequency Analysis, Prentice-Hall, New Jersey, 1995.
[63] C.K. Chui, An Introduction to Wavelets, Academic Press, Inc., 1992.
[64] M. Jobert, C. Tismer, E. Poiseau, H. Schulz, Wavelets a new tool in sleep
biosignal analysis, J. Sleep Res. 3 (1994) 223232.
[65] P.S. Addison, The Illustrated Wavelet Transform Handbook, Taylor & Francis,
2002.
[66] P.J. Durka, From wavelets to adaptive approximations: timefrequency
parametrization of EEG, Biomed. Eng. Online 2 (2003) 1.
[67] J. Zygierewicz, K.J. Blinowska, P.J. Durka, W. Szelenberger, S. Niemcewicz, W.
Androsiuk, High resolution study of sleep spindles, Clin. Neurophysiol. 110
(1999) 21362147.
[68] S.G. Mallat, Z.F. Zhang, Matching pursuits with timefrequency dictionaries,
IEEE Trans. Signal Process. 41 (1993) 33973415.
[69] P.J. Durka, D. Ircha, K.J. Blinowska, Stochastic timefrequency dictionaries for
matching pursuit, IEEE Trans. Signal Process. 49 (2001) 507510.
[70] K.J. Blinowska, P.J. Durka, Unbiased high resolution method of EEG analysis in timefrequency space, Acta Neurobiol. Exp. (Warsz.) 61 (2001)
157174.
[71] E. Huupponen, W. De Clercq, G. Gomez-Herrero, A. Saastamoinen, K. Egiazarian, A. Varri, B. Vanrumste, A. Vergult, S. Van Huffel, W. Van Paesschen, J.
Hasan, S.L. Himanen, Determination of dominant simulated spindle frequency
with different methods, J. Neurosci. Methods 156 (2006) 275283.
[72] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.-C. Yen, C.C.
Tung, H.H. Liu, The empirical mode decomposition and the Hilbert spectrum
for nonlinear and non-stationary time series analysis, Proc. Math. Phys. Eng.
Sci. 454 (1998) 903995.
[73] B.H. Jansen, Quantitative-analysis of electroencephalograms is there chaos
in the future, Int. J. Bio-Med. Comput. 27 (1991) 95123.
[74] D. Gallez, A. Babloyantz, Predictability of human EEG a dynamic approach,
Biol. Cybern. 64 (1991) 381391.
[75] J. Fell, J. Roschke, C. Schaffner, Surrogate data analysis of sleep electroencephalograms reveals evidence for nonlinearity, Biol. Cybern. 75 (1996)
8592.
[76] J.W. Milnor, Attractor, in: Scholarpedia, 2006.
[77] C.J. Stam, Nonlinear dynamical analysis of EEG and MEG: review of an emerging eld, Clin. Neurophysiol. 116 (2005) 22662301.
[78] B. Henry, N. Lovell, F. Camacho, Nonlinear dynamics time series analysis, in:
M. Akay (Ed.), Nonlinear Biomedical Signal Processing, vol. 2, 2001.
[79] M. Palus, Nonlinearity in normal human EEG: cycles, temporal asymmetry, nonstationarity and randomness, not chaos, Biol. Cybern. 75 (1996)
389396.
[80] P.S. Addison, Fractals and Chaos: An Illustrated Course, IOP publishing Ltd.,
1997.
[81] A. Accardo, M. Afnito, M. Carrozzi, F. Bouquet, Use of the fractal dimension for
the analysis of electroencephalographic time series, Biol. Cybern. 77 (1997)
339350.
[82] T. Higuchi, Approach to an irregular time series on the basis of the fractal
theory, Phys. D 31 (1988) 277283.
[83] M.J. Katz, Fractals and the analysis of waveforms, Comput. Biol. Med. 18 (1988)
145156.
[84] R. Acharya, O. Faust, N. Kannathal, T. Chua, S. Laxminarayan, Non-linear analysis of EEG signals at various sleep stages, Comput. Methods Programs Biomed.
80 (2005) 3745.
[85] J.E. Arle, R.H. Simon, An application of fractal dimension to the detection
of transients in the electroencephalogram, Electroencephalogr. Clin. Neurophysiol. 75 (1990) 296305.
[86] M.T.R. Peiris, R.D. Jones, P.R. Davidson, P.J. Bones, D.J. Myall, Fractal Dimension
of the EEG for Detection of Behavioural Microsleeps, IEEE, Shanghai, China,
2006, pp. 4.
[87] P. Grassberger, I. Procaccia, Characterization of strange attractors, Phys. Rev.
Lett. 50 (1983) 346.
[88] R. Ferri, L. Parrino, A. Smerieri, M.G. Terzano, M. Elia, S.A. Musumeci, S. Pettinato, C.J. Stam, Non-linear EEG measures during sleep: effects of the different
sleep stages and cyclic alternating pattern, Int. J. Psychophysiol. 43 (2002)
273286.
[89] S. Janjarasjitt, M.S. Scher, K.A. Loparo, Nonlinear dynamical analysis of the
neonatal EEG time series: the relationship between sleep state and complexity, Clin. Neurophysiol. 119 (2008) 18121823.
[90] M.S. Scher, H. Waisanen, K. Loparo, M.W. Johnson, Prediction of neonatal state
and maturational change using dimensional analysis, J. Clin. Neurophysiol. 22
(2005) 159165.
[91] J. Roeschke, J. Aldenhoff, The dimensionality of humans electroencephalogram during sleep, Biol. Cybern. 64 (1991) 307314.
[92] P. Achermann, R. Hartmann, A. Gunzinger, W. Guggenbuhl, A.A. Borbely,
All-night sleep EEG and articial stochastic-control signals have similar
correlation dimensions, Electroencephalogr. Clin. Neurophysiol. 90 (1994)
384387.
[93] T. Kobayashi, S. Madokoro, Y. Wada, K. Misaki, H. Nakagawa, Human sleep
EEG analysis using the correlation dimension, Clin. Electroencephalogr. 32
(2001) 112118.
[94] J. Skinner, M. Molnar, C. Tomberg, The point correlation dimension: performance with nonstationary surrogate data and noise, Integr. Psychol. Behav.
Sci. 29 (1994) 217234.
[95] S.M. Pincus, I.M. Gladstone, R.A. Ehrenkranz, A regularity statistic for medical
data-analysis, J. Clin. Monit. 7 (1991) 335345.
[96] S.M. Pincus, Approximate entropy as a measure of system-complexity, Proc.
Natl. Acad. Sci. U.S.A. 88 (1991) 22972301.
[97] S.M. Pincus, A.L. Goldberger, Physiological time-series analysis what does
regularity quantify, Am. J. Physiol. 266 (1994) H1643H1656.
[98] Y. Fusheng, H. Bo, T. Qingyu, Approximate entropy and its application in
biosignal analysis, in: M. Akay (Ed.), Nonlinear Biomedical Signal Processing,
IEEE Press, 2000.
[99] J.S. Richman, J.R. Moorman, Physiological time-series analysis using approximate entropy and sample entropy, Am. J. Physiol. Heart Circ. Physiol. 278
(2000) H2039H2049.
[100] J. Ge, P. Zhou, X. Zhao, M. Wang, Sample Entropy Analysis of Sleep EEG Under
Different Stages, IEEE, Beijing, China, 2007, pp. 15271530.
[101] H.G. Schuster, W. Just, Deterministic Chaos: An Introduction, Wiley-VCH,
2005.
[102] J. Roschke, J. Fell, P. Beckmann, The calculation of the 1st positive Lyapunov
exponent in sleep EEG data, Electroencephalogr. Clin. Neurophysiol. 86 (1993)
348352.
[103] A. Wolf, J.B. Swift, H.L. Swinney, J.A. Vastano, Determining Lyapunov exponents from a time-series, Phys. D 16 (1985) 285317.
[104] J. Fell, J. Roschke, P. Beckmann, Deterministic chaos and the 1st positive Lyapunov exponent a nonlinear-analysis of the human electroencephalogram
during sleep, Biol. Cybern. 69 (1993) 139146.
[105] X.Y. Wang, L. Chao, M. Juan, Nonlinear dynamic research on EEG signals in
HAI experiment, Appl. Math. Comput. 207 (2009) 6374.
[106] T. Shimada, T. Shiina, Y. Saito, CT sleep stage diagnosis system with neural
network analysis, in: H.K. Chang, Y.T. Zhang (Eds.), Engineering in Medicine
and Biology Society, 1998. Proceedings of the 20th Annual International Conference of the IEEE, vol. 2074, IEEE, Hong Kong, China, 1998, pp. 20742077.
[107] I.N. Bankman, V.G. Sigillito, R.A. Wise, P.L. Smith, Feature-based detection
of the K-complex wave in the human electroencephalogram using neural
networks, IEEE Trans. Biomed. Eng. 39 (1992) 13051310.
[108] D.R. Liu, Z.Y. Pang, S.R. Lloyd, A neural network method for detection of
obstructive sleep apnea and narcolepsy based on pupil size and EEG, IEEE
Trans. Neural Netw. 19 (2008) 308318.
[109] M.E. Tagluk, M. Akin, N. Sezgin, Classication of sleep apnea by using
wavelet transform and articial neural networks, Expert Syst. Appl. 37 (2009)
16001607.
[110] S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice Hall PTR,
1994.
[111] C. Robert, J.F. Gaudy, A. Limoge, Electroencephalogram processing using neural networks, Clin. Neurophysiol. 113 (2002) 694701.
[112] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University
Press, Inc., 1995.
[113] T.M. Mitchell, Machine learning and data mining, Commun. ACM 42 (1999)
3036.
[114] G. Cybenko, Approximation by superpositions of a sigmoidal function, Math.
Control Signals Syst. 2 (1989) 303314.
[115] D. Balakrishnan, S. Puthusserypady, Multilayer perceptrons for the classication of brain computer interface data, in: Bioengineering Conference, 2005.
Proceedings of the IEEE 31st Annual Northeast, 2005, pp. 118119.
[116] J.C. Bezdek, S.K. Pal, Fuzzy models for pattern recognition: background, significance, and key points, in: Fuzzy Models for Pattern Recognition, IEEE Press,
1992.
[117] A.K. Jain, M.N. Murty, P.J. Flynn, Data clustering: a review, ACM Comput. Surv.
31 (1999) 264323.
[118] T. Kohonen, T. Honkela, Kohonen Network, 2007.
[119] T. Kohonen, The self-organizing map, Proc. IEEE 78 (1990) 14641480.
[120] T. Kohonen, Self-organized formation of topologically correct feature maps,
Biol. Cybern. 43 (1982) 5969.
[121] S. Roberts, L. Tarassenko, Analysis of the Human EEG Using Self-organising
Neural Nets, IEE, London, UK, 1992, pp. 6/16/3.
[122] S. Roberts, L. Tarassenko, Analysis of the sleep EEG using a multilayer network with spatial-organization, IEE Proc. Radar Signal Process. 139 (1992)
420425.
[123] T. Kohonen, Self-organizing Maps, Springer, 2001.
[124] J.Y. Tian, J.Q. Liu, Automated sleep staging by a hybrid system comprising
neural network and fuzzy rule-based reasoning, in: 2005 27th Annual International Conference of the IEEE Engineering in Medicine and Biology Society,
vols. 17, IEEE, New York, 2005, pp. 41154118.
[125] S. Roberts, L. Tarassenko, New method of automated sleep quantication,
Med. Biol. Eng. Comput. 30 (1992) 509517.
[126] M. Golz, D. Sommer, T. Lembcke, B. Kurella, Classication of Pre-stimulus
EEG of K-complexes Using Competitive Learning Networks, vol. 1763, Verlag
Mainz, Aachen, Germany, 1998, pp. 17671771.
[127] D. Michie, D.J. Spiegelhalter, C.C. Taylor, Machine Learning, Neural and Statistical Classication, Ellis Horwood, 1994, pp. 289.
[128] R. Fisher, The use of multiple measurements in taxonomic problems, Ann.
Eugen. 7 (1936) 179188.
[129] L. Devroye, L. Gyr, G. Lugosi, A Probabilistic Theory of Pattern Recognition
(Stochastic Modelling and Applied Probability), Springer, 1996.
[130] V.N. Vapnik, Statistical Learning Theory, Wiley Inter-Science, 1998.
[131] F. Lotte, M. Congedo, A. Lecuyer, F. Lamarche, B. Arnaldi, A review of classication algorithms for EEG-based braincomputer interfaces, J. Neural Eng. 4
(2007) R1R13.

[132] S.P. Cho, H.S. Choi, H.K. Lee, K.J. Lee, Detection of EEG arousals in patients
with respiratory sleep disorder, in: S.I. Kim, T.S. Suh (Eds.), World Congress on
Medical Physics and Biomedical Engineering 2006, vol. 14, Pts 16, SpringerVerlag, Berlin, 2007, pp. 11311134.
[133] N. Acir, C. Guzelis, Automatic recognition of sleep spindles in EEG via radial
basis support vector machine based on a modied feature selection algorithm,
Neural Comput. Appl. 14 (2005) 5665.
[134] C.J.C. Burges, A Tutorial on support vector machines for pattern recognition,
Data Min. Knowl. Discov. 2 (1998) 121167.
[135] L. Rabiner, B. Juang, An introduction to hidden Markov models, IEEE ASSP
Mag. 3 (1986) 416.
[136] L.R. Rabiner, A tutorial on hidden Markov models and selected applications
in speech recognition, Proc. IEEE 77 (1989) 257286.
33
[137] A. Flexer, G. Dorffner, P. Sykacek, I. Rezek, An automatic, continuous and probabilistic sleep stager based on a hidden Markov model, Appl. Artif. Intell. 16
(2002) 199207.
[138] L.A. Zadeh, Fuzzy sets, Inf. Control 8 (1965) 338353.
[139] I. Gath, A.B. Geva, Unsupervised optimal fuzzy clustering, IEEE Trans. Pattern
Anal. Mach. Intell. 11 (1989) 773781.
[140] C.M. Held, J.E. Heiss, P.A. Estevez, C.A. Perez, M. Garrido, C. Algarin, P. Peirano,
Extracting fuzzy rules from polysomnographic recordings for infant sleep
classication, IEEE Trans. Biomed. Eng. 53 (2006) 19541962.
[141] H. Laufs, M.C. Walker, T.E. Lund, Brain activation and hypothalamic functional
connectivity during human non-rapid eye movement sleep: an EEG/fMRI
study its limitations and an alternative approach, Brain 130 (2007)
e75.

Art 1

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Art 1

Hochgeladen von

Copyright:

Verfügbare Formate

Biomedical Signal Processing and Control 10 (2014) 2133

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

Signal processing techniques applied to human sleep EEG signalsA

Faculty of Engineering and the Environment, University of Southampton, UK

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

The characteristic of the PSG signals to be analysed is rather

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

pre-processing which are common to many EEG analysis problems

to regard an artefact as any signicant deviation from the norm

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

to high resolution studies where the number of EEG electrodes

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

suitable for the analysis of non-stationary, noisy signals such as the

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

3.2.3. Parametric spectral estimation

are manifested by sudden changes in amplitude and frequency

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

real-time applications. In short, since its introduction, MP has been

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

The precise length of data for reliable estimation of ApEn

3.4.4. Lyapunov exponents

1, stage 2, stage 3 sleep and REM, or in automatic spindle detection,

4.1. Neural network (NN) classication (supervised learning)

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

new patterns (feedforward conguration). However, training them

4.3. Statistical classication

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

non-stationary signals by dividing the signal into segments and

feature classication. Each of these is supported by supplementary

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

[31] A. Isaksson, A. Wennberg, L.H. Zetterberg, Computer-analysis of EEG signals

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

S. Motamedi-Fakhr et al. / Biomedical Signal Processing and Control 10 (2014) 2133

Das könnte Ihnen auch gefallen