November 17, 2010 12:0 00258

International Journal of Neural Systems, Vol. 20, No. 6 (2010) 509–521

c World Scientific Publishing Company
DOI: 10.1142/S0129065710002589

ANALYSIS AND AUTOMATIC IDENTIFICATION OF SLEEP

STAGES USING HIGHER ORDER SPECTRA
RAJENDRA ACHARYA U
Department of Electronics and Computer Engineering
Ngee Ann Polytechnic, Singapore 599489
ERIC CHERN-PIN CHUA
Duke-NUS Graduate Medical School, Singapore
KUANG CHUA CHUA and LIM CHOO MIN
Department of Electronics and Computer Engineering
Ngee Ann Polytechnic, Singapore 599489
TOSHIYO TAMURA
Chiba University, Chiba, Japan

Electroencephalogram (EEG) signals are widely used to study the activity of the brain, such as to
determine sleep stages. These EEG signals are nonlinear and non-stationary in nature. It is difficult to
perform sleep staging by visual interpretation and linear techniques. Thus, we use a nonlinear technique,
higher order spectra (HOS), to extract hidden information in the sleep EEG signal. In this study, unique
bispectrum and bicoherence plots for various sleep stages were proposed. These can be used as visual
aid for various diagnostics application. A number of HOS based features were extracted from these plots
during the various sleep stages (Wakefulness, Rapid Eye Movement (REM), Stage 1-4 Non-REM) and
they were found to be statistically significant with p-value lower than 0.001 using ANOVA test. These
features were fed to a Gaussian mixture model (GMM) classifier for automatic identification. Our results
indicate that the proposed system is able to identify sleep stages with an accuracy of 88.7%.

Keywords: Sleep; EEG; bispectrum; entropy; higher order spectra; classifier.

1. Introduction (drowsiness state), and eyes begin to roll slightly in

Non-linear signal processing has been effectively
applied to electroencephalogram (EEG) signals
to study the dynamics of complex underly-
ing behaviour.5–7,10,11,20,21,30,34–38,42,43,46,49,53,59,62
These methods are superior to traditional linear
methods such as time domain analysis and power
spectral analysis.44 Sleep study is a very important
tool in the assessment of sleep-related disorders, such
as sleep apnea, schizophrenia, depression, and other
neural abnormalities.13
Typical features of the different sleep stages are
given below.28 During Wakefulness, the low ampli-
tude EEG is rapidly varying and beta waves (12–
30 Hz) are more prominent. The alpha (8–12 Hz) and
theta (4–7 Hz) waves are present in Stage 1 sleep
this state (slow eye movements). During Stage 2 sleep
(light sleep), the EEG amplitude becomes higher,
theta waves are more predominant, and K-complexes
appear. Usually the frequency ranges from 4 to 7 Hz
and lasts for a few minutes during this state. Deep
sleep (Stage 3) has very low frequency delta (2–4 Hz)
and theta waves. In this sleep stage, 20 to 50% of
EEG signals are delta waves and rest are theta waves.
In Stage 4, the frequency varies between 0.5 Hz to
2 Hz and more than 50% of EEG signals are delta
waves. In REM sleep, beta waves are more predomi-
nant and frequency will be greater than 12 Hz. REM
sleep is a highly activated state of the brain and
is accompanied by dreaming. Phasic bursts of eye
movements with occasional muscle twitches exists

509
November 17, 2010 12:0 00258
510 U. R. Acharya et al.
in this state. Sleep stages are characterized by a
cyclic alternating Non-rapid eye movement (NREM)
(Stages 1 to 4) and REM pattern.9,15,17,18,27,31
In the clinical context, assessment of sleep dis-
turbance by analysing sleep stage is required for
assessment of several sleep disorders, such as sleep
apnea [R&K]. The gold standard for doing so is
attended, in-laboratory polysomnography (PSG),
where sleep stages are obtained by manually scor-
ing EEG, electro-oculography (EOG) and electro-
myography (EMG) montages.60
However, resource constraints have restricted
access to attended, in-laboratory PSG in many
countries,27 and there is therefore significant interest
from professional bodies and sleep experts in simple
and low-cost portable monitors.52
To our knowledge, there are very few portable
monitors capable of providing full sleep staging infor-
mation, i.e. distinguishing between Wake, REM, and
the different levels of NREM sleep. Those that do
require all the five sleep staging signals, i.e. 2 EEGs,
2 EOGs and 1 EMG, making them cumbersome and
requiring considerable technical expertise to operate
in a home environment. Simplified systems employ-
ing actimetry are available, but they can only distin-
guish between Wake and Sleep, but cannot indicate
the specific stage of sleep.
A simple system capable of providing full sleep
staging information is therefore of potential clinical
value. An algorithm to separate the different sleep
stages in infants based on statistical analysis of the
spectral and nonlinear characteristics of the sleep
EEG recordings was developed.12 However, spectral
analysis of the EEG will not be able to capture
clearly the nonlinear nature of the sleep EEG signal.
A method for automatic detection and classifica-
tion of sleep stages using multichannel EEG was pre-
sented by Zhovna and Shallom.65 The author used
the cross correlation information existing between
different EEG signals. The classification phase was
performed using Kullback-Leibler (KL) divergence
and a classification accuracy rate of 93.2% was
obtained. Five EEG channels namely Pz , Cz , Pz , T3 ,
T4 were used. But, the effect of the number of chan-
nels and the location of electrodes on the classifica-
tion was not explored.
A method for the detection of sleep stages using
four steps: segmentation, parameter extraction, clus-
ter analysis, and classification was proposed using
EEG signals.40 The parameters of Hjorth, the har-
monic parameters and the relative band energy were
fed to a modified K-means algorithm. The authors
showed that the proposed method was able to find
‘similar’ segments and automate the detection of
sleep stages. However, with extra information of
ECG (electrocardiogram) or the EOG different sleep
stages can be discriminated more accurately.
A sleep EEG recognition neural network (SRNN)
to detect several important characteristic waves in
sleep EEG to diagnose sleep stages was developed.58
Their experimental results indicate that the pro-
posed NN model was more capable than other con-
ventional methods for detecting characteristic waves.
These important characteristics of sleep EEG wave
detected by SRNN model can be used for identifying
sleep stages.
The correlation dimensions were used to analyze
the EEG signal.41
The correlation dimension decreased from Wake-
fulness to stages 1–3 and increased during REM.
In each sleep cycle, the correlation dimensions
decreased for slow wave sleep, and increased for REM
sleep. However, in that work, only one nonlinear
parameter (correlation dimension) was used on a
single subject.
Several features have been proposed to detect the
hidden important dynamical properties of the phys-
iological phenomenon. Statistical characteristics of
biological signals change with time and are highly
irregular and non disorders like epilepsy.3,4,56 Vari-
ous nonlinear parameters like correlation dimension
for pathological signals has been shown to be useful
indicators of pathologies.33 These nonlinear parame-
ters have also been used to analyze the sleep stages.
Acharya et al., used nonlinear parameters: correla-
tion dimension, fractal dimension, largest Lyapunov
entropy, approximate entropy, Hurst exponent, phase
space plot and recurrence plots to identify differ-
ent sleep stages. 1 They proposed different range of
values for different non-linear parameters and pro-
posed unique recurrence plots for different sleep
stages. But their method did not identify sleep stages
automatically.
Accordingly, we present an automated system
that distinguishes between Wake, REM, and the dif-
ferent levels of NREM sleep, using only two chan-
nels of EEG. In this paper, we analyze and classify
the different sleep stages using higher order spectra
November 17, 2010 12:0 00258
features and a GMM classifier. Also, we propose dif-
ferent bispectrum and bicoherence plots for different
sleep stages.
2. Material and Methods
2.1. Subjects and laboratory procedures
Data from two cohorts were used for this study. For
the first cohort, twenty-five subjects were recruited
randomly from patients attending the Sleep Disor-
ders Clinic at St Vincent’s University Hospital, Ire-
land, for evaluation of suspected OSAS (21 M 4 F,
age 50 ± 10 years, BMI 31.6 ± 4.0 kg/m2 , AHI 24.1 ±
20.3 h−1 ). All subjects provided written, informed
consent. The subjects underwent standard overnight,
attended PSG. An experienced sleep technologist
subsequently performed sleep staging according to
the Rechtschaffen and Kales rules54 and annotated
the respiratory events.
In another study at University College Dublin,
Ireland, 14 subjects with no known medical con-
ditions were recruited from the general popula-
tion (12 M 2 F, age 27 ± 4 years, BMI 25 ±
4 kg/m2 ). All subjects provided written, informed
consent. Standard sleep staging signals were recorded
overnight using a set of Grass amplifiers (Astro-
Med Inc, USA) while subjects obtained their usual
sleep. Sleep staging was subsequently performed
using the Somnolyzer 24 × 7 system14 using the
Rechtschaffen and Kales rules.54 A recent valida-
tion study using the large Siesta database indicated
that the Somnolyzer achieves similar inter-rater reli-
ability with a human scorer as between two human
scorers.14
Two gold cup electrodes were placed at C4A1 and
C3A2 (standard 10/20 placement) for one night’s
sleep, from lights off to lights on to collect the EEG
data. Bandpass filter of frequency 0.3/35 Hz, 16 bit
ADC and sampling frequency of 128 Hz was used
for the St Vincent’s University Hospital dataset.
0.3/30 Hz frequency bandpass filter, 16 bit ADC and
sampling frequency of 125 Hz was used for the UCD
dataset.
2.2. EEG analysis
For each record, the two standard sleep staging EEG
signals (C4/A1 and C3/A2) were analysed. First, the
two channel EEG signals are averaged. Higher order
spectra (HOS) techniques were then applied to each
Higher Order Spectra-based Sleep Staging 511
epoch of 30 seconds to obtain various HOS features
(details provided in subsequent sections).
To compute bispectrum as in equation (1), the
FFT block size of 256 points (i.e. 2 seconds) was used
with an overlap of 128 point (i.e. 50%). This com-
putation is averaged over each track for the entire
length of the EEG signal.
2.3. Higher order spectra analysis
The EEG signals were analyzed using higher order
spectra that are spectral representations of higher
order moments or cumulants of a signal. The features
related to the third order statistics of the signal,
namely the bispectrum were extracted. The Bispec-
trum is the Fourier transform of the third order
correlation of the signal and is given by
B(f1 , f2 ) = E[X(f1 )X(f2 )X ∗ (f1 + f2 )], (1)
Where X(f ) is the Fourier transform of the sig-
nal x(nT ), f1 and f2 are the frequencies and E[·]
stands for the expectation operation. The expecta-
tion operation is the average over an ensemble of
realizations of a random signal. For deterministic
signals, the relationship holds without an expecta-
tion operation with the third order correlation being
a time-average. For deterministic sampled signals,
X(f ) is computed as the discrete Fourier trans-
form (DFT) with f being the normalized Nyquist
frequency (value ranges between 0 and 1).
2.4. Higher order spectral features
Different bispectral entropies and bicoherence plots
have been proposed for cardiac arrhythmia23 and
normal, epileptic and background epileptic EEG
signals.24 Bispectral entropies23 have been derived
to find the rhythmic nature of heart rate variabil-
ity (HRV) from bispectrum plots. The formulae for
these bispectral entropies are as follows:
Normalized Bispectral Entropy (BE 1):


Ent1 = − pn log pn , (2)
n
where
|B(f1 , f2 )|
pn = 
Ω |B(f1 , f2 )|
Ω = the region as in Fig. 1.
November 17, 2010 12:0 00258

512 U. R. Acharya et al.

The WCOB63 is given by:

 
Ω iB(i, j) jB(i, j)
f1m =  f2m = Ω (7)
Ω B(i, j) Ω B(i, j)
0.5 where i, j are the frequency bin index in the non-
redundant region.
f2 The
non-redundant
3. Automated Sleep Stage
region, Ω
Classification using Gaussian
Mixture Model
0.5 1 In this work, we used the Gaussian Mixture Model
f1 (GMM) classifier for classification. It is briefly
Fig. 1. Non-redundant region of computation of the bis- explained below:
pectrum for real signals. Features are calculated from this
region. Frequencies are shown normalized by the Nyquist 3.1. Gaussian Mixture Model (GMM)
frequency.
GMM is a statistical model which comprises a num-
ber of Gaussian functions, and has been widely used
Normalized Bispectral Squared Entropy (BE 2):

for automatic identification of bio-signals.30,32 It is
Ent2 = − qn log qn (3) used to approximate a continuous probability den-
n sity function from a multi-dimensional feature vec-
where tor. The Gaussian mixture distribution is given by:
  
|B(f1 , f2 )|2
K 

qn =  
2 p(x) = πk N x µk , , (8)
Ω |B(f1 , f2 )| 
k=1 k
Ω = the region as in Figure 1. 
where πk , µk , k are the mixing coefficients, mean
The normalization in the equations above ensures
and covariance respectively.
that entropy is calculated for a parameter that lies
The probability density of a single Gaussian com-
between 0 and 1 (as required of a probability) and
ponent of D dimensions is given by:
hence the entropies (Ent1 and Ent2) computed are    


also between 0 and 1.  1 1
g x µi , =   exp − (x − µi )

In this work, we used features related to  2
i (2π)D | i |
moments65 and the weighted centre of bispectrum

(WCOB)63 to characterize these plots. The features −1


related to the moments65 are: × (x − µi ) , (9)
The sum of logarithmic amplitudes of the bispec- i

trum: where (
) denotes the vector transpose. Maximum

Likelihood (ML) parameter estimation criterion
H1 = log(|B(f1 , f2 )|) (4)
was used to determine the solution for parame-
Ω
ters of GMM. The model parameters are evaluated
The sum of logarithmic amplitudes of diagonal
through training such that they maximize the like-
elements in the bispectrum:

lihood of the observations using the Expectation-
H2 = log(|B(fk , fk )|) (5) Maximization (E-M) algorithm.16,55,58
Ω Using K-means algorithm the initial estimates of
The first-order spectral moment of amplitudes of the parameters were obtained from a sample of the
diagonal elements in the bispectrum: training data. Randomly chosen initial means and
N

unit variances for the covariance matrix were used
H3 = k log(|B(fk , fk )|) (6) in this work. The diagonal covariance matrices were
k=1 used, because it was found to be more computa-
These features (H1 , H2 , and H3 ) can be used to clas- tionally efficient and performed better than the full
sify mental tasks from EEG signals. covariance matrix.16,47
November 17, 2010 12:0 00258

Higher Order Spectra-based Sleep Staging 513

4. Results where the bispectrum extended further away from

Figures 2 to 7 shows the bispectral and bicoher-
ence plots for the various sleep stages (Wakefulness,
Stages 1 to 4 sleep and REM) respectively. It can
be observed that these plots are unique for each
class, although Wakefulness and REM resemble each
other. This resemblance seems to suggest that dur-
ing REM there were more non-linear interactions of
higher frequency components, as seen from the graph
the centre.
Table 1 shows the features derived from HOS
analysis. These features were subjected to ANOVA
test and found to be statistically significant for dif-
ferent sleep stages (p < 0.0001). It can be seen
from Table 1 that the different entropies (Ent1 and
Ent2) are high for wakefulness, Stage 1 sleep and
REM sleep due to more variability (high frequency).
Stages 2 to 4 have lower values of entropies due to

(a) (b)

Fig. 2. Wakefulness plots: (a) bispectrum (b) bicoherence.

(a) (b)

Fig. 3. Stage 1 sleep plots: (a) bispectrum (b) bicoherence.

November 17, 2010 12:0 00258

514 U. R. Acharya et al.

(a) (b)

Fig. 4. Stage 2 sleep plots: (a) bispectrum (b) bicoherence.

(a) (b)

Fig. 5. Stage 3 sleep plots: (a) bispectrum (b) bicoherence.

lower variability (low frequency). The other parame-
ters mAmp, Wcobx, Wcoby, H1, H2 and H3 indicate
different ranges for various sleep stages.
We selected four HOS features namely, Ent1,
mAmp, Wcobx and H1 as the input to the GMM
classifier for classification. Ten-fold stratified cross
validation method was used to test the GMM classi-
fier. That is, the whole dataset was split into ten
parts such that each part contains approximately
the same proportion of class samples as the original
dataset. Nine parts of the data (training set) were
used for classifier development and the built classi-
fier was evaluated using the remaining one part (test-
ing data). This procedure was repeated ten times
(folds) using a different part for testing in each case.
Table 2 shows the confusion matrix comparing ref-
erence sleep stages and sleep stages estimated using
HOS statistics and the GMM classifier. It can be seen
from the confusion matrix that our proposed GMM
classifier is able to classify the unknown class with
November 17, 2010 12:0 00258

Higher Order Spectra-based Sleep Staging 515

(a) (b)

Fig. 6. Stage 4 sleep plots: (a) bispectrum (b) bicoherence.

(a) (b)

Fig. 7. REM sleep plots: (a) bispectrum (b) bicoherence.

an accuracy of 88.7%. The sensitivity, specificity and
positive predictive accuracy of the proposed system
are 89.16%, 87.13% and 96.02%, respectively.
5. Discussion
Table 4 shows the summary of comparison of meth-
ods used for the automated detection of sleep
stages. An automated method for sleep stage scor-
ing using hybrid rule- and case-based reasoning was
proposed.50 This method comprises signal processing
unit, rule-based scoring unit, and case-based scor-
ing unit. They obtained an average agreement rate
for normal recordings 87.5% and case-based scor-
ing enhanced this agreement rate by 5.6%. Hence,
the combination of rule-based reasoning and case-
based reasoning was more promising for an auto-
mated sleep scoring and can be used a good model
of the cognitive scoring process.
Computer-assisted sleep staging method using
polysomnograms was proposed based on segmen-
tation and self-organization (clustering).8 Results
November 17, 2010 12:0 00258

516 U. R. Acharya et al.

Table 1. Results of various HOS parameters for different sleep stages.

Features Wake Stage 1 Stage 2 Stage 3 Stage 4 REM p-value

Ent1 0.71378 0.65454 0.53345 0.45288 0.45197 0.56383 <0.0001

± 0.191 ± 0.153 ± 0.114 ± 7.508E-02 ± 7.391E-02 ± 0.165
Ent2 0.24550 0.28143 0.22111 0.16874 0.15579 0.22503 <0.0001
± 0.189 ± 0.164 ± 9.353E-02 ± 7.263E-02 ± 6.273E-02 ± 0.152
mAmp 1.54957E+12 3.46052E+11 6.94747E+11 1.77045E+12 1.51427E+12 1.54232E+12 <0.0001
± 3.771E+12 ± 9.348E+11 ± 1.017E+12 ± 2.499E+12 ± 1.591E+12 ± 3.790E+12
Wcobx 20.381 16.772 7.9618 5.7321 5.1614 11.831 <0.0001
± 14.8 ± 16.0 ± 7.38 ± 3.36 ± 3.72 ± 11.8
Wcoby 7.4714 6.8340 3.1523 2.8642 2.5992 4.6091 < 0.0001
± 5.45 ± 6.29 ± 1.72 ± 1.12 ± 1.30 ± 4.53
H1 42608 39484 39621 40413 40830 39604 <0.0001
± 2.331E+03 ± 1.974E+03 ± 1.529E+03 ± 1.999E+03 ± 1.687E+03 ± 2.646E+03
H2 700.79 665.89 666.97 678.05 686.72 667.48 <0.0001
± 32.9 ± 25.1 ± 20.0 ± 27.8 ± 24.4 ± 35.1
H3 22392 20876 20553 20885 21227 20775 <0.0001
± 1.163E+03 ± 1.023E+03 ± 800 ± 1.025E+03 ± 895 ± 1.270E+03

Table 2. Confusion matrix comparing reference sleep stages and sleep stages
estimated using HOS statistics and the GMM classifier (88.7% accuracy).

Reference sleep stages

W S1 S2 S3 S4 REM Total

Sleep Stages Estimated from GMM Classifier

W 1205 11 118 8 39 2 1383
S1 3 959 4 14 40 0 1020
S2 99 7 1785 0 148 55 2094
S3 0 43 6 131 21 0 201
S4 44 1 10 12 530 0 597
REM 0 0 4 0 11 888 903
Total 1351 1021 1927 165 789 945 6198

W: wake state; S1: Sleep 1 stage; S2: Sleep 2 stage; S3: Sleep 3 stage; S4: Sleep
4 stage; REM: Rapid Eye Movement stage.

Table 3. TP (True Positive), TN (True Negative), FP (False Positive), FN (False Negative), Sensitivity, specificity
and positive predictive accuracy (PPA) for the proposed system.

True True False False

Classifier positive (TP) negative (TN) positive (FP) negative (FN) Sensitivity Specificity PPA

GMM 4293 1205 178 522 89.16% 87.13% 96.02%

Note: TP: Number of abnormal detected as abnormal; TN: Number of normal detected as normal; FP: Number of
normal detected as abnormal; FN: number of abnormal detected as normal; Sensitivity: TP/(TP+FN); Specificity:
TN/(TN + FP); PPA; PPA (Positive Predictive Accuracy): TP/(TP + FP).
November 17, 2010 12:0 00258

Higher Order Spectra-based Sleep Staging 517

Table 4. Comparison of methods proposed in the literature for automated sleep stage detection.

Authors Method Accuracy (%)

Pfurtscheller et al., 199251
Park et al., 200050
Agarwal et al., 20018
Held et al., 200658
Sinha et al., 200861
Ebrahimi et al., 200829
Bresler et al., 200819
Acharya et al., 20101
Multilayer perceptron and a learning vector 65% to 80%
quantizer
Hybrid rule- and case-based reasoning 87.5%
Segmentation and self-organization 80.6%
Neuro-fuzzy classifier 83.9%
Second order Daubechies mother wavelet 96.84% in sleep, 93.68% in REM
coefficients for the EEG epochs and neural sleep, 95.52% in awake state
network
Wavelet packet coefficients and artificial neural 93%
networks
14 features extracted from two time series of 82% for light and 80% for deep
PAT amplitudes and inter-pulse periods sleep
HOS features and GMM 88.9%

showed an overall concurrence of 80.6% with man-
ual scoring of 20-s epochs according to Rechtschaffen
and Kales (RK) standard. The highest error occurred
during the identification of the highly transitional
Stage 1, 54% of which was misclassified into stage
2 or Wake. This method can easily be adapted for
sleep stages in animals or to generate new classifica-
tion schemes of human sleep.
Neuro-fuzzy classifier (NFC) to identify the sleep-
wake states and stages hass been developed for
healthy infants using polysomnographic recordings
having five classes: Wakefulness, REM, Stage 1,
Stage 2, and Stage 3–4.39 They showed a classifi-
cation efficiency of 83.9 +/− 0.4%. Results show
that the NFC can be used as a valuable tool for
implementing an automated sleep-wake classification
system.
Second order Daubechies mother wavelet coeffi-
cients for the EEG epochs (64 data) were extracted
from EEG signals for the training the neural network
and to classify sleep spindles (SS), rapid eye move-
ment (REM) sleep and awake (AWA) sates.61 The
ANN architecture used (64-14-3) was found effective
in differentiating the EEG power spectra from differ-
ent sleep-wake states (96.84% in SS, 93.68% in REM
sleep, 95.52% in AWA state). So, the combination of
wavelet coefficients and ANN can be used as a com-
putational tool for the sleep research due to its high
efficiency.
Wavelet packet coefficients and artificial neural
networks were used to classify the sleep stages using
EEG signals.29 Their results show that they were
automatically classify four sleep stages with a speci-
ficity of 94.4 ± 4.5%, a sensitivity of 84.2 + 3.9% and
an accuracy of 93.0 ± 4.0%.
An automatic algorithm based on the peripheral
arterial tone (PAT) signal to differentiate between
light and deep sleep stages was developed.19 The
algorithm was based on 14 features extracted from
two time series of PAT amplitudes and inter-pulse
periods (IPP). The sensitivity, specificity and agree-
ment of the automatic algorithm to detect 30 s
epochs of light and deep sleep stages were 66%, 89%,
82% and 65%, 87%, 80% for the training and val-
idation sets, respectively. This automatic detection
algorithm can be very useful for unattended ambu-
latory sleep monitoring assessing sleep stages when
EEG recordings are not available.
In our work, we have used novel HOS features
to extract the hidden complexities in the sleep EEG
signals. Our results show that the proposed system is
able to identify the unknown class with an accuracy
of 88.7%.
Sleep EEG signals are complex signals and has
some non-linear interaction in its frequency compo-
nents and phase coupling. The HOS methods per-
form better even when the physiological signals are
very noisy.25 Numerous experiments reported that
the HOS analysis (i.e. the 2D plots of bispectrum
and bicoherence) of various pathological signals are
unique and different HOS parameters have been used
to differentiate and classify them.25,26 These HOS
features were reported to have performance superior
than the second order measures (power spectrum).
November 17, 2010 12:0 00258
518 U. R. Acharya et al.
In this paper we have discussed these HOS features
for the identification of sleep stages.
We are proposing a general EEG staging algo-
rithm that can be used at home and laboratory. This
can be used as a portable algorithm to monitor sleep
staging at home, which will allow long term monitor-
ing without the transportation of logistics. The sleep
stages can be transmitted electronically to the labo-
ratory and can be reviewed during subject’s sched-
uled appointment.
6. Conclusion
Sleep EEG signals are highly non-linear and non-
stationary in nature. Hence, they are to be analyzed
by non-linear dynamics methods. In this paper, we
used HOS analysis to extract the salient features
from the sleep stages and fed it to the GMM clas-
sifier for automatic identification. We also proposed
unique bispectrum and bicoherence plots and ranges
of HOS features for each sleep stage. Our results
show that the proposed method is able to identify
the sleep stages with an accuracy of 88.7%, with
sensitivity and specificity of 89.16%, 87.13% respec-
tively. Our proposed method is computationally less
intensive and easy to use. The system is so easy to
use, even a nurse can use it to identify the sleep
stages. It can also be used to identify the efficacy of
the drug. However, with more diverse training data
and better HOS features the accuracy can be further
increased.
Acknowledgement
The authors are grateful to Professor Walter T
McNicholas of St Vincent’s University Hospital, Ire-
land and Professor Conor Heneghan of University
College Dublin, Ireland for providing the sleep data
for this study.
References
1. U. R. Acharya, O. Faust, N. Kannathal, T. Chua
and S. Laxminarayan, Non-linear analysis of EEG
signals at various sleep stages, Computer Methods
and Programs in Biomedicine 80(1) (2005) 37–45.
2. H. Adeli and S. Ghosh-Dastidar, Automated EEG-
based Diagnosis of Neurological Disorders — Invent-
ing the Future of Neurology, CRC Press Taylor &
Francis, Boca Raton, Florida (2010).
3. H. Adeli, Z. Zhou and N. Dadmehr, Analysis of EEG
records in an epileptic patient using wavelet trans-
form, Journal of Neuroscience Methods 123 (2003)
69–87.
4. H. Adeli, S. Ghosh-Dastidar and N. Dadmehr, A
wavelet-chaos methodology for analysis of EEGs
and EEG sub-bands to detect seizure and epilepsy,
IEEE Transactions on Biomedical Engineering 54(2)
(2007) 205–211.
5. H. Adeli, S. Ghosh-Dastidar and N. Dadmehr,
Alzheimer’s disease and models of computation:
Imaging, classification, and neural models, Journal
of Alzheimer’s Disease 7(3) (2005a) 187–199.
6. H. Adeli, S. Ghosh-Dastidar and N. Dadmehr,
Alzheimer’s disease: Models of computation and
analysis of EEGs Clinical EEG and Neuroscience
36(3) (2005b) 131–140.
7. H. Adeli, S. Ghosh-Dastidar and N. Dadmehr,
A spatio-temporal wavelet-chaos methodology for
EEG-based diagnosis of alzheimer’s disease, Neuro-
science Letters 444(2) (2008) 190–194.
8. R. Agarwal and J. Gotman, Computer-assisted sleep
staging, IEEE Trans Biomed Eng. 48(12) (2001)
1412–23.
9. H. W. Agnew, W. B. Webb and R. L. Williams,
Sleep patterns in late middle age males: An EEG
study, Electroenceph Clinical Neurophysiology 23
(1967) 168–71.
10. M. Ahmadlou, and H. Adeli, Wavelet- synchroniza-
tion methodology: A new approach for EEG-based
diagnosis of ADHD, Clinical EEG and Neuroscience
41(1) (2010) 1–10.
11. M. Ahmadlou and H. Adeli, Fuzzy synchroniza-
tion likelihood for automated EEG-based diagnosis
of attention-deficit/hyperactivity disorder, Clinical
EEG and Neuroscience 42(1) (2011).
12. P. Alexandra, T. Gyorgy, A. W. Wojbor, A. L.
Kenneth, S. S. Mark and Z. Anatoly, Automated
detection of neonate EEG sleep stages, Computer
Methods and Programs in Biomedicine 95(1) (2009)
31–46.
13. American College of Chest Physicians, SLEEP:
Study Lessons in Education and Evaluation of
Polysomnography 2, S. Karger Publishing (2006).
14. P. Anderer, G. Gruber, S. Parapatics, M. Woertz,
T. Miazhynskaia, G. Klosch, B. Saletu, J. Zeitlhofer,
M. J. Barbanoj, H. Danker-Hopfe, S. L. Himanen, B.
Kemp, T. Penzel, M. Grozinger, D. Kunz, P. Rap-
pelsberger, A. Schlogl and G. Dorffner, An e-health
solution for automatic sleep classification according
to rechtschaffen and kales: validation study of the
somnolyzer 24×7 utilizing the siesta database, Neu-
ropsychobiology 51(3) (2005) 115—133.
15. F. Bes, H. Schulz, Y. Navelet and P. Salzarulo, The
distribution of slow-wave sleep across the night: a
comparison for infants, children and adults, Sleep 14
(1991) 5–12.
November 17, 2010 12:0 00258
16. J. A. Bilmes, A Gentle Tutorial of the EM Algo-
rithm and its Application to Parameter Estima-
tion for Gaussian Mixture and Hidden Markov
Models, International Computer Science Institute
(1998).
17. E. O. Bixler, A. Kales, J. A. Jacoby, C. R. Soldatos
and A. Vela-Bueno, Nocturnal sleep and wakeful-
ness: effects of age and sex in normal sleeper, Int.
J. Neurosci. 23 (1984) 33–42.
18. D. L. Bliwise, Sleep in normal aging and dementia,
Sleep 16 (1993) 40–81.
19. M. Bresler, Differentiating between light and deep
sleep stages using an ambulatory device based on
peripheral arterial tonometry, Physiological Mea-
surement 29(5) (2008) 571–584.
20. N. Chakravarthy, S. Sabesan, K. Tsakalis and
L. Iasemidis, Controling synchronization in a neural-
level population model, International Journal of
Neural Systems 17(2) (2007) 123–138.
21. M. Chen, C. S. Jiang, Q. X. Wu and W. H. Chen,
Synchronization in arrays of uncertain delay neural
networks by decentralized feedback control, Inter-
national Journal of Neural Systems 17(2) (2007)
115–122.
22. M. Chiappalone, A. Vato, L. Berdondini,
M. Koudelka and S. Martinoia, Network dynamics
and synchronous activity in cultured cortical neu-
rons, 17(2) (2007) 87–103.
23. K. C. Chua, V. Chandran, U. R. Acharya and C. M.
Lim, Cardiac state diagnosis using higher order spec-
tra of heart rate variability, Journal of Medical &
Engineering Technology 32(2) (2008) 145–155.
24. K. C. Chua, V. Chandran, U. R. Acharya and C.
M. Lim, Analysis of epileptic EEG signals using
higher order spectra, Journal of Medical & Engineer-
ing Technology 33(1) (2009) 42–50.
25. K. C. Chua, V. Chandran, U. R. Acharya and C. M.
Lim, Application of higher order statistics/spectra in
biomedical signals — A Review, Medical Engineering
and Physics (2010a) (in press).
26. K. C. Chua, V. Chandran, U. R. Acharya and C. M.
Lim, Application of higher order spectra to identify
epileptic EEG, Journal of Medical Systems 2010b (in
press).
27. D. J. Dijk, T. L. Shanahan, J. F. Duffy, J. M. Ronda
and C. A. Czeisler, Variation of electroencephalo-
graphic activity during non-rapid eye movement and
rapid eye movement sleep with phase of circadian
melatonin rhythm in humans, Journal of Physiology
505 (1997) 851–858.
28. Dream Views: Staying up all night. The stages
of sleep. Retrieved from: http://www.dreamviews.
com/sleepstages.php (Last accessed on November
2009).
29. F. Ebrahimi, M. Mikaeili, E. Estrada and H. Naz-
eran, Automatic sleep stage classification based on
EEG signals by using neural networks and wavelet
Higher Order Spectra-based Sleep Staging 519
packet coefficients, 30th Annual International
Conference of the IEEE Engineering in Medicine
and Biology Society (2008) 1151–1154.
30. O. Faust, U. R. Acharya, C. M. Lim and H. C.
Sputh, Automatic identification of epileptic and
background EEG signals using frequency domain
parameters, International Journal of Neural Systems
20(2) (2010) 159–176.
31. I. Feinberg, R. L. Koresko and N. Heller, EEG sleep
patterns as a function of normal and pathological
aging in man, J Psych Res 5 (1967) 107–44.
32. M. Frean, M. Lilley and P. Boyle, Implement-
ing Gaussian process inference with neural net-
works, International Journal of Neural Systems
16(5) (2006) 321–327.
33. S. Ghosh-Dastidar, H. Adeli and N. Dadmehr,
Mixedband wavelet-chaos-neural network method-
ology for epilepsy and epileptic seizure detec-
tion, IEEE Transactions on Biomedical Engineering
54(9) (2007) 1545–1551.
34. S. Ghosh-Dastidar and H. Adeli, Improved spiking
neural networks for EEG classification and epilepsy
and seizure detection, Integrated Computer-Aided
Engineering 14(3) (2007) 187–212.
35. S. Ghosh-Dastidar, H. Adeli and N. Dadmehr, Prin-
cipal component analysis-enhanced cosine radial
basis function neural network for robust epilepsy and
seizure detection, IEEE Transactions on Biomedical
Engineering 55(2) (2008) 512–518.
36. S. Ghosh-Dastidar and H. Adeli, Spiking neural
networks, International Journal of Neural Systems
19(4) (2009a) 295–308.
37. S. Ghosh-Dastidar and H. Adeli, A new supervised
learning algorithm for multiple spiking neural net-
works with application in epilepsy and seizure detec-
tion, Neural Networks 22 (2009b) 1419–1431.
38. L. B. Good, S. Sabesan, S. T. Marsh, K. S. Tsakalis
and L. D. Iasemidis Control of synchronization of
brain dynamics leads to control of epileptic seizures
in rodents, International Journal of Neural Systems
19(3) (2009) 173–196.
39. C. M. Held, J. E. Heiss, P. A. Estévez, C. A. Perez,
M. Garrido, C. Algarı́n and P. Peirano, Extract-
ing fuzzy rules from polysomnographic recordings for
infant sleep classification, IEEE Trans Biomed Eng.
53(10) (2006) 1954–1962.
40. V. Hese, P. Philips, W. De Koninck, V. J. de
Walle, R. Lemahieu and I. Elis, Automatic detec-
tion of sleep stages using the EEG, Proceedings of
the 23rd Annual International Conference of the
IEEE Engineering in Medicine and Biology 2 (2001)
1944–1947.
41. T. Kobayashi, K. Misaki, H. Nakagawa, S.
Madokoro, H. Ihara, K. Tsuda, Y. Umezawa, J.
Murayama and K. Isaki, Non-linear analysis of the
sleep EEG, Psychiatry Clin Neurosci. 53(2) (1999)
159–161.
November 17, 2010 12:0 00258
520 U. R. Acharya et al.
42. M. A. Kramer, F. L. Chang, M. E. Cohen, D. Hud-
son and A. J. Szeri, Synchronization measures of the
scalp EEG can discriminate healthy from Alzheimers
subjects, International Journal of Neural Systems
17(2) (2007) 61–69.
43. H. Lee, A. Cichocki and S. Choi, Nonnegative
matrix factorization for motor imagery EEG clas-
sification, International Journal of Neural Systems
17(4) (2007) 305–317.
44. G. Mayer-Kress and S. P. Layne, Dimensionality of
the human electroencephalogram, in Perspectives in
biological dynamics and theoretical medicine, S. H.
Koslow, Annuls of New York Academy of Sciences,
(New York) 54 (1987) 62–87.
45. R. A. McCarthy and E. K. Warrington, Cogni-
tive Neuropsychology: A Clinical Introduction, San
Diego, LA: Academic Press (1990).
46. A. Montina, C. Mendoza and F. T. Arecchi, Role
of refractory period in homoclinic models of neu-
ral synchronization, International Journal of Neural
Systems 17(2) (2007) 79–86.
47. F. V. Nelwamondo and T. Marwala, Faults detec-
tion using gaussian mixture models, mel-frequency
cepstral coefficients and kurtosis, Systems, Man and
Cybernetics, (SMC’ 06), IEEE International Con-
ference 1 (2006) 290–295.
48. I. Osorio and M. G. Frei, Seizure abatement with
single DC pulses: Is phase resetting at play?, Inter-
national Journal of Neural Systems 19(3) (2009)
149–156.
49. H. Osterhage, F. Mormann, T. Wagner and K. Lehn-
ertz, Measuring the directionality of coupling: Phase
versus state space dynamics and application to EEG
time series, International Journal of Neural Systems
17(3) (2007) 139–148.
50. H. J. Park, J. S. Oh, D. U. Jeong and K. S. Park,
Automated sleep stage scoring using hybrid rule-
and case-based reasoning, Comp. Biomed Res. 33(5)
(2000) 330–349.
51. G. Pfurtscheller, D. Flotzinger and K. Matuschik,
Sleep classification in infants based on artificial neu-
ral networks, Biomed Tech (Berl). 37(6) (1992)
122–130.
52. Portable monitoring task force of the American
Academy of Sleep Medicine. Clinical guidelines for
the use of unattended portable monitors in the diag-
nosis of obstructive sleep apnea in adult patients.
Journal of Clinical Sleep Medicine 3(7) (2007)
737–747.
53. D. E. Postnov, L. S. Ryazanova, R. A. Zhirin,
E. Mosekilde and O. V. Sosnovtseva, Noise con-
trolled synchronization in Potassium Coupled Neural
Networks, International Journal of Neural Systems
17(2) (2007) 105–113.
54. A. Rechtschaffen and A. Kales A, A manual of stan-
dardized terminology, techniques, and scoring sys-
tem for sleep states of human subjects, Washington
DC: U.S. Government Printing Office (1968) NIH
Publication No. 204.
55. D. A. Reynolds and R. C. Rose, Robust text-
independent speaker identification using Gaussian
mixture speaker models, IEEE Transactions on
Speech and Audio Processing 3 (1995) 72–83.
56. J. L. Rossello, V. Canals, A. Morro and J. Verd,
Chaos-based mixed signal implementation of spik-
ing neurons, International Journal of Neural Systems
19(6) (2009) 465–471.
57. C. Seo, K. Y. Lee and J. Lee, GMM based on local
PCA for speaker identification, Electronics Letters
37 (2001) 1486–1488.
58. T. Shimada, T. Shiina and Y. Saito, Detection of
characteristic waves of sleep EEG by neural network
analysis, IEEE Trans Biomed Eng. 47(3) (2000)
369–379.
59. A. Shoeb, J. Guttag, T. Pang and S. Schachter,
Non-invasive computerized system for automatically
initiating vagus nerve stimulation following patient-
specific detection of seizures or epileptiform dis-
charges, International Journal of Neural Systems
19(3) (2009) 157–172
60. M. H. Silber, S. Ancoli-Israel, M. H. Bonnet,
S. Chokroverty, M. M. Grigg-Damberger, M. Hir-
shkowitz, S. Kapen, S. A. Keenan, M. H. Kryger, T.
Penzel, M. R. Pressman and C. Iber, The visual scor-
ing of sleep in adults, J Clin Sleep Med 3(2) (2007)
121–131.
61. R. K. Sinha, Artificial neural network and wavelet
based automated detection of sleep spindles, REM
sleep and wake states, J Med Syst 32(4) (2008) 291–
299.
62. X. Wang, Y. Chen, S. Bressler and M. Ding, Eval-
uating causal relations among multiple neurobiolog-
ical time series: Blockwise versus pairwise granger
causality, International Journal of Neural Systems
17(2) (2007) 71–78.
63. J. Zhang, C. Zheng, D. Jiang et al., Bispectrum anal-
ysis of focal ischemic cerebral EEG signal, in Pro-
ceedings of the 20th Annual International Conference
of the IEEE Engineering in Medicine and Biology
Society 20 (1998) 2023–2026.
64. I. Zhovna and I. D. Shallom, Automatic detection
and classification of sleep stages by multichannel
EEG signal modeling, Conf Proc IEEE Eng Med Biol
Soc. 1 (2008) 2665–2668.
65. S. M. Zhuo, J. Q. Gan and F. Sepulveda, Classifying
mental tasks based on features of higher-order statis-
tics from EEG signals in brain–computer interface,
Information Sciences 178(6) (2008) 1629–1640.
November 17, 2010 12:0 00258

Higher Order Spectra-based Sleep Staging 521

Appendix Selected EEG window. The bispectrum and Bico-

Figure 8 shows the snap shot of the graphical user
interface of our proposed automated sleep detection
system. There is a push button called “Browse for
file” is provided to upload the file to be analyzed.
The EEG signal to be analyzed is displayed at the
herence plot are displayed by clicking the “HOS
Plot” pushbutton. The unknown class of the EEG
signal will be displayed in the text box entitled
Result when clicked the “Diagnosis” push button.
In our present example it is Wake state.

Fig. 8. Snap shot of the graphical user interface of the proposed system.

The author has requested enhancement of the downloaded file. All in-text references underlined in blue are linked to publications on ResearchGate.