EEG feature extraction for classification of sleep stages

Abstract

Automated sleep staging based on EEG signal analysis provides an important quantitative tool to assist neurologists and sleep specialists in the diagnosis and monitoring of sleep disorders as well as evaluation of treatment efficacy. A complete visual inspection of the EEG recordings acquired during nocturnal polysomnography is time consuming, expensive, and often subjective. Therefore, feature extraction is implemented as an essential preprocessing step to achieve significant data reduction and to determine informative measures for automatic sleep staging. However, the analysis of the EEG signal and extraction of sensitive measures from it has been a challenging task due to the complexity and variability of this signal. We present three different schemes to extract features from the EEG signal: relative spectral band energy, harmonic parameters, and Itakura distance. Spectral estimation is performed by using autoregressive (AR) modeling. We then compare the performance of these schemes with the view to select an optimal set of features for specific, sensitive, and accurate neuro-fuzzy classification of sleep stages.

Full text

Abstract— Automated sleep staging based on EEG signal
analysis provides an important quantitative tool to assist
neurologists and sleep specialists in the diagnosis and
monitoring of sleep disorders as well as evaluation of
treatment efficacy. A complete visual inspection of the
EEG recordings acquired during nocturnal
polysomnography is time consuming, expensive, and
often subjective. Therefore, feature extraction is
implemented as an essential preprocessing step to
achieve significant data reduction and to determine
informative measures for automatic sleep staging.
However, the analysis of the EEG signal and extraction
of sensitive measures from it has been a challenging task
due to the complexity and variability of this signal.
In this paper we present three different schemes to
extract features from the EEG signal: Relative Spectral
Band Energy, Harmonic Parameters, and Itakura
Distance. Spectral estimation is performed by using
Autoregressive (AR) modeling. We then compare the
performance of these schemes with the view to select an
optimal set of features for specific, sensitive, and
accurate neuro-fuzzy classification of sleep stages
.
Keywords
— Sleep staging; EEG signal processing;
feature extraction; biomedical signal processing;
harmonic parameters; Itakura distance; sleep apnea.
I. INTRODUCTION
Electroencephalogram (EEG) is perhaps the most
important tool in studying sleep and sleep-related disorders
such as sleep apnea or insomnia. Sleep states are comprised
of two general stages: rapid eye mo vement (REM) and non-
rapid eye movement (NREM). NREM is in turn subdivided
into four stages: 1, 2, 3, and 4 according to Rechtschaffen
and Kales (RK) sleep scoring standard [1]. The awake state
is not formally a sleep state, but it is considered a state for
the purpose of sleep scoring.
Brain activity is divided into four main rhythms. Beta
waves are defined as low voltage (around 5 µV) and high
frequency waves (14 to 30 Hz, sometimes as high as 50 Hz).
Alpha waves, which occur during relaxed states, are regular
rhythms of 8 to 13 Hz with higher amplitudes than beta
waves. Theta waves are typically of even greater amplitude
and slower frequency than alpha waves. Their frequency
range is normally between 4 and 7 HZ. Delta waves, slowest
EEG rhythms, generally have the highest amplitude EEG
waveforms observed (about 300 µV) with all the frequencies
below 3.5 Hz.
Beta waves represent arousal, while alpha waves
represent non-arousal. The awake state is more related to the
beta waves than alpha waves. Stage 1, which is considered
to be a midway state between waking and sleep, is further
linked to the alpha and theta waves. Stage 2 is manifested by
the low voltage waves of stage 1 mixed with what are
known as the K-complex (a sharp, high voltage transient
wave which occurs spontaneously) and sleep spindle waves
(bursts of waves having a frequency of 12 to 15 Hz)
. Stage
3 sleep begins when low voltage background waves that
distinguish the spindles are replaced by high amplitude, low
frequency delta waves
. In the deep stage 4 sleep, spindles
drop out and the EEG signal consists almost of delta waves.
Table 1 shows the relationship among sleep stages and brain
activity [2, 3].
Table 1. Relationship among sleep stages and EEG rhythms.
Stage ECG Brain Activity Waveform Contained
Awake Beta, Alpha
Stage 1 Alpha, Theta
Stage 2 Alpha, Theta , K complex , spindle waves
Stage 3 Delta, spindle waves
Stage 4 Delta
During REM sleep, the brain activity is reversed from
Stage 4 to a pattern similar to stage 1. In general, REM sleep
is associated with visual dreaming.
II. MATERIALS AND METHODS
A. Data Collection
The data acquired from one volunteer subject with no
sleep disorders who was referred to our accredited Sleep
laboratory (Consultants Inc., Fort Worth, Texas) was used
for this study. A total of 7 hours and 4 minutes of sleep data
were recorded for this subject. The 10-20 Standard electrode
placement system was used for EEG recording. Specifically
recordings between C
1
and A
2
positions were used for this
study. The EEG signals were amplified using a Nihon
Kohden polygraph (Irvine, CA) and the data acquisition was
achieved using a Telefactor System (Conshocken, PA) for
polysomnography. One EEG channel was stored using a
EEG FEATURE EXTRACTION FOR CLASSIFICATION OF SLEEP STAGES
E Estrada, H Nazeran, P Nava, K Behbehani
*
, J Burk
**
, and E Lucas
**
Department of Electrical and Computer Engineering, The University of Texas at El Paso, El Paso, Texas, USA
* Biomedical Engineering Program, The University of Texas at Arlington, Arlington, Texas, USA
** Sleep Consultants Inc., Ft Worth, Texas, USA
Email: nazeran@ece.utep.edu
0-7803-8439-3/04/$20.00©2004 IEEE
196
Proceedings of the 26th Annual International Conference of the IEEE EMBS 
San Francisco, CA, USA • September 1-5, 2004

sampling frequency of 1000 Hz. An experienced sleep
specialist blind to the objective of this study scored the EEG
epochs of 30 seconds in duration based on the RK standard
method. Table 2 shows the percentage and the number of
scored segment for this patient. This information was saved
by the Sleep Laboratory into a binary file.
Table 2. Polysomnogram Information.
Stage 30 second scored epochs Percent %
Awake 185 21
Stage 1 52 6
Stage 2 314 36
Stage 3 84 10
Stage 4 128 15
REM 86 10
For simplicity and further processing this file was
restructured into two principal vectors: one containing the
EEG information (with F
s
= 1000 Hz), and the other
containing the Sleep Stage information (1 unit for
occurrence of each 30 seconds epoch of EEG).
B. EEG Preprocessing
The complete EEG vector was processed using a sixth
order Butterworth bandpass filter with cutoff (corner)
frequencies of 0.5 - 50 Hz. A zero-phase digital filter was
realized by filtering the EEG signals in both forward and
reverse directions resulting in its filtering by a 12
th
order
filter. EEG signals were then decimated by a factor of 10
resulting in signals with a new sampling frequency of 100
Hz.
C. Feature Extraction
For implementation of the feature extraction schemes
we used the following procedure to obtain the statistics for
each measure. First, we partitioned the EEG signal into
different epoch lengths considering the fact that the EEG
signals were scored in segments of 30-second epochs. In
addition, to track the sleep stage information provided by
the sleep specialist, partitioning was performed by taking
new epoch lengths as sub-partitions of 30 seconds. These
sub-partitions provided the flexibility to minimize the
probability of having 30-second segments with more than
one sleep stage present. Second, after running the feature
extraction algorithm, the extracted measures were collected
into different groups using the annotations in the sleep stage
vector. This allowed us to compute the mean, standard
deviation, maximum, and minimum values for each sleep
state.
Relative spectral energy band and harmonic Hjorth
methods [5] were used for feature extraction. These
measures required the computation of the power spectral
density of the EEG signals. This task could be achieved by
using parametric or non-parametric spectral modeling
methods. Autoregressive (AR) modeling was selected as it
provides smoother, more accurate, higher resolution spectra
of the EEG signals and the calculation of Itakura distance is
based on AR parameters. The disadvantage of using this
method is due to the need for selection of an appropriate
model order “p”. This selection could be based on a priori
knowledge or on different criteria (e.g. Akaike Information
Criterion or Minimum Description Length Criterion) [4].
C1. Autoregressive Spectral Estimation of EEG Signals
The autoregressive AR (p) process is an especial case of
an Autoregressive moving average ARMA (p,q) process
when q=0 [4]. The AR (p) process is generated when unit
variance white noise w(n) is passed through an all pole filter
of the form:
∑
=
−
+
=
p
k
k
p
zka
b
zH
1
)(1
)0(
)(
(1)
Also the autocorrelation sequence of this process satisfies
the Yule-Walker equations:
∑
=
≥=−+
p
l
xpx
kblkrlakr
1
2
0)0()()()(
(2)
Hence, solving for these equations we can obtain the a
p
(l)
coefficients. b(0) can be found as follows:
∑
=
+=
p
k
xpx
krkarb
1
2
)()()0()0( (3)
Using the estimates of the model coefficients, it is
possible to estimate the Power Spectrum:
2
1
2
2
)(1
)0(
)()(
∑
=
−
∧
∧
∧
=
+
=≈
p
k
jkw
p
jw
epoch
ez
eka
b
epzH
jw
(4)
where the P(f) is defined as follows:
)(
ˆ
)(
ˆ
2 fj
epoch
ePfP
π
= (5)
C2. Relative Percent Spectral Energy Band
For this analysis we computed the total power content
(TPC) of P(f), from 0.5 to 45 Hz. P(f) was divided into
seven different energy bands, and the respective power
energy bands (PEB) were calculated. The relative percent
spectral energy band (RPEB) was then expressed as:
100x
TPC
PEB
RPEB = (6)
197

Table 3 shows the partition of frequency range for each bin
or energy band [6].
Table 3. Spectral Energy Bands of EEG waves.
Band Bandwidth (Hz)
Delta 1 0.5-2.5
Delta 2 2.5-4
Theta 1 4-6
Theta 2 6-8
Alpha 8-12
Beta 1 12-20
Beta 2 20-45
C3. Harmonic Parameters
The harmonic parameters [5, 6], which are the
frequency versions of the Hjorth parameters are: the center
frequency, the bandwidth, and the value at the central
frequency. These parameters are defined as follows:
∫∫
=
H
L
H
L
f
f
f
f
c
dffPdfffPf )(/)( (7)
∫∫
−=
H
L
H
L
f
f
f
f
c
dffPdffPfff )(/)()(
2
σ
(8)
)(
cc
fPPf = (9)
where, f
L
and f
H
are set as 0.5 and 45 Hz, respectively. Since
the computation of P(f) involves discrete values, the above
formulas were approximated using summations as follows:
∑∑
==
=
H
L
H
L
f
ff
f
ff
c
fPfPff )(
ˆ
/)(
ˆ
(10)
∑∑
==
−=
H
L
H
L
f
ff
f
ff
c
fPfPfff )(
ˆ
/)(
ˆ
)(
2
σ
(11)
).'(
ˆ
cc
fPPf = (12)
In the above formulas, the f index spans from 0.5 to 45 with
increments of 0.5 Hz. The f
c
’ is the closest f index value to f
c.
C4. Itakura Distance
Itakura distance is used widely in speech processing
applications to measure the distance between 2 AR
processes [7, 8]. Here the Itakura distance was used to
measure the similarly of a base line EEG epoch (Awake,
Stage1, Stage2, Stage 3, Stage 4, REM) with the rest of the
epochs in the EEG vector. If we let the baseline epoch x[n]
be an AR process given by a
x
=[1-a
1
-a
2
…..-a
p
] and the
segment y[n] to be compared to it given by a
y
= [1 -a
1
-a
2
…..-a
p
], then the minimum square error (MSE) for the
baseline process is:
xx
T
xxx
apRaMSE )(
,
= (14)
where the R
x
(p) is the autocorrelation matrix for the baseline
epoch of size p +1. Similarly the MSE of the other processes
passing through the baseline model will be:
yx
T
yyx
apRaMSE )(
,
= (15)
The Itakura distance of the baseline to the other epochs is
defined as:
)/log(
,,
,
xxyxI
MSEMSEd
yx
= (16)
Furthermore, an analysis of how well y[n] is modeled
via the AR parameters of x[n] can be done, thus the new
Itakura distance is:
)
)(
)(
log()log(
,
,
,
yy
T
y
xy
T
x
yy
xy
I
apRa
apRa
MSE
MSE
d
xy
== (17)
Combining (16) and (17) we obtain the symmetric Itakura
distance as:
()
xyyxyx
III
ddd
,,,
2
1
' +=
(18)
III. RESULTS
Using the above mentioned features, a matrix of 11
features by 2,547 (10-second length) segments were
extracted. A total of 25,470 measures are now available to
be fed into a neuro-fuzzy system. Figure 1 shows the mean
Itakura distance between the sleep stages and the baseline
epoch (taken as the Awake state).
Figure 1. Mean Itakura Distance
198

Clearly the Itakura distance shows that when the subject
falls into a deeper sleep stage, the distance increases as a
consequence of the AR process changes. Additionally, the
central frequency reflects that the Awake state is related to
higher frequencies of the spectra, and Stage 4 is linked to
slow waves (Figure 2).
Figure 2. Central Frequency
The analysis of the data showed that the % relative
energy band of Delta 1 contained more spectral power.
Figure 3 below shows the % relative energy band of delta 1
waves for different sleep stages.
Figure 3. Relative spectral energy band of delta 1.
The remaining % relative energy bands for other
waveforms of the EEG signals are shown in Figure 4. It is
observed that the % relative energy band of beta 2 waves is
a good feature to be used to distinguish between different
sleep stages.
IV. DISCUSSION
The results demonstrate that the extracted features
provide promising possibilities to distinguish between
different sleep stages. It is also evident that REM is quite
difficult to separate from other sleep stages due to its
spectral overlap with those of the other stages. The EOG
signal may be a more discerning signal for the detection of
REM activity. Therefore, detection of the REM stage from
EEG signal analysis remains a challenging research topic
that warrants further investigation.
Figure 4. Percent relative spectral energy bands for different
waves in the EEG signals .
V. CONCLUSION
The Itakura distance and central frequency seem to
provide promising features for classification of sleep stages.
However, the high variance of these measures causes the
mean values of the features to overlap and makes sleep
staging by conventional statistics difficult. Therefore, neuro-
fuzzy classifiers are being developed to facilitate this
process.
REFERENCES
[1] Rechtschaffen A. Kales A. Rechtschaffen A. Kales A, eds. “A
Manual of standardized terminology, techniques and scoring
system for sleep stages of human subjects. Los Angeles: Brain
Information service/Brain Research Institute, 1968
[2] W. B. Meldenson “Human Sleep, Research and Clinical Care”
Plenum Medical Book Company New York and London”, 1987,
pp 6-12.
[3] J. G. Webster “Medical Instrumentation, Application and
Design” Third Edition , Wiley 1998, pp 165-171
[4] M. H. Hayes “Statistical Digital Signal Processing and
Modeling”, Wiley, 1996 pp. 194,198-199, 440, 447.
[5] P. Van Hese, W. Philips, J. De Koninck. R.Van de Walle, and I.
Lemahieu. “Automatic Detection of Sleep Stages Using the
EEG. Proceedings of the 23
rd
Annual EMBS International
Conference, October 25-28, Istanbul, Turkey, 2001, pp. 1944-
1947.
[6] K. Donohue and C. Scheib,\EEG fractal response to
anesthetic gas concentration." Available at:
www.engr.uky.edu/_donohue/eeg/pre1/EEGpre2.html.
[7] X. Kong, N. Thakor, and V. Goel, “Characterization of the EEG
Signal Changes Via Itakura Distance”, IEEE-EMBC and
CMEC, 1995, pp. 873-874
[8] J. Muthuswamy, and N. V. Thakor, “Spectral Analysis methods
for neurological signals” Journal of Neuroscience Methods,
1998, pp. 1-14.
199