Wavelet Transform Use For Feature Extraction and EEG Signal Segments Classification
Wavelet Transform Use For Feature Extraction and EEG Signal Segments Classification
Wavelet Transform Use For Feature Extraction and EEG Signal Segments Classification
Abstract—Segmentation, feature extraction and classification II. EEG SIGNAL PREPROCESSING AND
of signal components belong to very common problems in various SEGMENTATION
engineering, economical and biomedical applications. The paper
is devoted to the use of discrete wavelet transform (DWT) both Information content of EEG signals is essential for detec-
for signal preprocessing and signal segments feature extraction as tion of many problems of the brain and in connection with
an alternative to the commonly used discrete Fourier transform analysis of magnetic resonance images it forms one of the
(DFT). Feature vectors belonging to separate signal segments most complex diagnostic tools. To extract the most important
are then classified by a competitive neural network as one of
properties of EEG observations it is necessary to use efficient
methods of cluster analysis and processing. The paper provides
a comparison of classification results using different methods of mathematical tools [9], [10] to enable reliable and fast enough
feature extraction most appropriate for EEG signal components processing of very extensive data sets in most cases.
detection. Problems of multichannel segmentation are mentioned Digital filters can be used in the initial stage of EEG data
in this connection as well. processing to remove power frequency from the observed
Index Terms—Segmentation, change-point detection, feature signal and to reduce its undesirable frequency components.
extraction, classification, multichannel signal processing, discrete
wavelet transform, neural networks Fig. 1 presents a sample of a selected EEG channel comparing
results of its segmentation by an expert and by a selected
I. INTRODUCTION Bayesian method [2] detecting changes of its mean value
and variation. This approach has been used in this case for
The preliminary stage of multichannel signal classification a selected channel only even though further channels must be
includes its segmentation using principal component analysis taken into account in the real case as well.
and signal de-noising [1] in many cases followed by different
methods of change-points detection [2], [3]. This stage is fun- III. WAVELET ANALYSIS AND SIGNAL FEATURE
damental for further data processing both for one-dimensional EXTRACTION
and multi-dimensional signals. Wavelet transform forms a general mathematical tool for
Signal segments feature extraction forms the next step of signal processing with many applications in EEG data analysis
signal segmentation allowing combination of time-domain and [11], [12], [13], [14], [15] as well. Its basic use includes
frequency-domain signal features. Commonly used spectral time-scale signal analysis, signal decomposition and signal
representation of a signal based upon its all-pole model or compression.
its discrete Fourier transform provides the same frequency
resolution over the whole window function. To allow different (a) EXPERT AND COMPUTATIONAL SEGMENTATION − DEC: 1
4
resolution the wavelet transform [4], [5], [6] is often used pro-
2
compression of signal features as well. The method presented Fig. 1. Results of EEG signal segmentation presenting (a) signal segmenta-
in the paper is applied for an EEG signal analysis and its tion by an expert (red) and by a Bayesian method of change-point detection
segments classification into the proposed number of classes. (blue) and (b) a probability function to detect change-points
978-1-4244-1688-2/08/$25.00
2008
c IEEE
Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on July 18,2022 at 12:50:15 UTC from IEEE Xplore. Restrictions apply.
720 ISCCSP 2008, Malta, 12-14 March 2008
for integer values of m, k and the initial wavelet defined either −0.2
−0.4 0
by the solution of a dilation equation or by an analytical
8 8 0.5
expression [4], [5]. Both continuous or discrete signals can 4 50 4
0.2
0.3
0.4
2 0 2 0.1
be then approximated in the way similar to Fourier series and DILATION 1 −50
TIME
DILATION 1 0.0 FREQUENCY
discrete Fourier transform. In case of a sequence {x(n)}N −1
n=0
s
having N = 2 values it is possible to evaluate its expansion Fig. 2. The real part of the complex wavelet h(t)=(ejπ t−ejπ/2 t )/(jπ/2 t)
s−m−1 and the effect of its dilation to the corresponding spectrum compression
s−1
2 −1
x(n) = a0 + a2s−m−1 +k h(2−m n − k)
m=0 k=0
L−1
Wavelet transform coefficients can be organized in a matrix T {h(n)}n=0 = [l(L − 1), −l(L − 2), l(L − 3), · · · , −l(0)]
withits nonzero elements forming a triangle structure Corresponding sequences convolution and subsampling by two
a2s−1 a2s−1 +1 ··· as provides values
a2s−2 ··· a2s−1 −1 2 −1
· · ·
L−1
n−L+1
a4 a a6 a
a2 5 a3 7 p(n) = l(k) x(n − k) = x(j) l(n − j)
a1 k=0 j=n,n−1,...
a0
with each its row corresponding to a separate dilation co- L−1
n−L+1
efficient m. The set of N = 2s decomposition coefficients q(n) = h(k) x(n − k) = x(j) h(n − j)
{a(j)}N −1
j=0 of the wavelet transform is defined in the way
k=0 j=n,n−1,...
formally close to the Fourier transform but owing to the for n = L − 1, L + 1, ..., N − 1. Subsequent application of
general definition of wavelet functions they can carry different the same method provides signal analysis for different scales.
information. Using the orthogonal set of wavelet functions they Introducing decomposition matrices
are moreover closely related to the signal energy [5].
l(1) l(0) 0 0 ···
The initial wavelet can be considered as a pass-band filter l(3) l(2) l(1) l(0) · · ·
and in most cases half-band filter covering the normalized LN/2,N = ··· ···
··· ··· ···
frequency band 0.25, 0.5). A wavelet dilation by the factor 0 · · · l(L − 1) · · · l(0)
a = 2m corresponds to a pass-band compression. This general
property can be demonstrated for the harmonic wavelet func- h(1) h(0) 0 0 ···
h(3) h(2) h(1) h(0) · · ·
tion [5] and the corresponding scaling function by expressions HN/2,N = ···
··· ··· ··· ···
1
h(t) = (ejπ t − ejπ/2 t ) (2) 0 · · · h(L − 1) · · · h(0)
jπ/2 t
1 it is possible to decompose the initial signal x into two
l(t) = (ejπ/2 t − 1) (3) sequences p = L x and q = H x standing for subsampled
π/2 t
low-pass and high-pass signal components. The elements of
As both these functions are modified by the scaling index
vector q represent wavelet transform coefficients of the initial
m = 0, 1, · · · according to Eq. (1), the wavelet is dilated
level. Further wavelet coefficients can be obtained after the
and its spectrum compressed resulting in time and frequency
application of this process to signal p according to Fig. 3.
domain representation presented in Fig. 2. Similar approach
Resulting wavelet coefficients {a(n)}N −1
n=0 related to chosen
can be also applied for other wavelet functions defined in either
scales can then be used as signal features for its classification.
analytical or recurrent form.
The set of wavelets define a special filter bank which can
be used for signal component analysis and resulting wavelet
{x(n)}N −1
n=0 presented in Fig. 3 assumes a half-band low-pass
scaling sequence
L−1
{l(n)}n=0 = [l(0), l(1), l(2), · · · , l(L − 1)] Fig. 3. A pyramidal filter bank structure used to evaluate wavelet transform
and the complementary orthogonal wavelet sequence coefficients for a sequence {x(n)}N −1 s
n=0 and for values s = 4 and N = 2 = 16
Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on July 18,2022 at 12:50:15 UTC from IEEE Xplore. Restrictions apply.
ISCCSP 2008, Malta, 12-14 March 2008 721
0.8
1
0.8
of the hidden layer. Further improvement can be achieved by a
0.6 0.6 proper choice of signal segments for signal features evaluation.
0.4 0.4
30 30
0.2
0 20
0.2
0 20
B. Results
6
5
4
3
2
10
Time
20
40
10
Time
The algorithm presented above has been applied for clas-
1 60
Scale Frequency sification of a real EEG signal. Fig. 6 presents EEG signal
segments and their analysis by a harmonic wavelet transform
Fig. 4. Presentation of (a) a part of a one channel EEG signal plot and the resulting in features standing for scales 1, 2 and 3 respectively
result of its analysis (b) by the wavelet transform based on harmonic wavelet
functions and (c) by the short-time Fourier transform covering three frequency bands with different time-scale res-
olution. Results of signal classification into four classes by a
self-organizing neural network are given in Fig. 7 and 8 for
Fig. 4 presents the scalogram and spectrogram of the two selected signal features allowing a simple visualization
selected part of the EEG signal comparing results achieved by of segmentation results. Class boundaries were evaluated and
the DWT and DFT. It is obvious that owing to the principle presented in this case as well. Both features clustering and time
of the wavelet transform short time signal components can domain signal segments values show that signals of similar
be better detected and more precisely localized by the DWT structure belong to the same class.
comparing to results obtained by the DFT. Results of signal classification into four classes for different
features are compared in Tab. I. Each class i = 1, 2, · · · , C
IV. SIGNAL SEGMENTS CLASSIFICATION can be characterized by the mean distance of the column
feature vector pj,i belonging to its separate segments j for
Classification of signal segments into a given number of
j = 1, 2, · · · , Ni from the centre ci of individual classes using
classes using segments features can be achieved by various C Ni
statistical methods. Another approach presented further is
relation 1 1
crit = dist(pj,i , ci ) (4)
based upon the application of self-organizing neural networks. C i=1 Ni j=1
where C stands for the number of classes, Ni represents the
A. Principle number of segments belonging to class i and function dist
Two typical principles of signal classification are presented is used for evaluation of the Euclidean distance between two
in Fig. 5. In the first case general self-organizing maps employ
directly signal features as patterns for an input layer. The (a) SEGMENTED SIGNAL
number of output layer elements is equal to signal classes and
must be either defined in advance or it can be automatically 1
increased to create new classes [8]. During the learning process 0.5
0.3
0.2
0.1
a) b)
2 4 6 8 10 12 14 16
Segment
Fig. 5. Principles of signal classification using (a) signal features obtained Fig. 6. Results of feature extraction presenting (a) EEG signal segments and
separately and (b) signal features defined by a compression method (b) their wavelet features resulting from a harmonic DWT on scales 1, 2, 3
Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on July 18,2022 at 12:50:15 UTC from IEEE Xplore. Restrictions apply.
722 ISCCSP 2008, Malta, 12-14 March 2008
(b) SIGNAL CLASSIFICATION (c) CLASS A CLASS B CLASS C CLASS D The Mean (Lowest) Square Error (4 classes)
(1,2)
Feature Source
15
3 Set 1 Set 2 Set 3
14 (15 segments) (144 segments) (234 segments)
13
11
DFT 0.009 (0.006) 0.026 (0.011) 0.030 (0.005)
6
7 10
Harmonic DWT 0.026 (0.008) 0.044 (0.005) 0.039 (0.003)
4
12
1
14 9 5
10 7
2 1
2 4 9
ACKNOWLEDGEMENT
8 5
1113
15 8 6 3 12 Real data sets of EEG signals have been kindly provided
by the Neurocenter Caregroup in Rychnov nad Kněžnou. This
Fig. 7. Results of signal classification into four classes presenting (a) given Institute provided also notes to real applications of methods
signal, (b) the topology of signal segment features with class boundaries, proposed. The paper has been supported by the Research grant
and (c) signal segments divided into four classes and typical signal segments
closest to the final neuron values at the bottom of each class No. MSM 6046137306.
CLASS A: 0.043 CLASS B: 0.017 CLASS C: 0.008 CLASS D: 0.039
15
R EFERENCES
14
13
[1] S. V. Vaseghi, Advanced Digital Signal Processing and Noise Reduction,
John Wiley & Sons Ltd, West Sussex, 2006.
11
[2] W. J. Fitzgerald, J. J. K. O. Ruanaidh, and J. A. Yates, “Generalised
10
Changepoint Detection,” Tech. Rep., University of Cambridge Engineer-
1
ing Department, Cambridge, U.K., 1994.
5 7 [3] B. E. Brodsky and B. S. Darkhovski, Nonparametric Methods in
2 4 9 Change-Point Problems, Kluwer Academic Publishers, Boston, U.S.A.,
8 6 3 12
1993.
[4] I. Daubechies, “The Wavelet Transform, Time-Frequency Localization
Fig. 8. The topography of the two EEG signal segments features for their and Signal Analysis,” IEEE Trans. Inform. Theory, vol. 36, pp. 961–
classification into four classes with typical segment features at the bottom 1005, September 1990.
[5] D. E. Newland, An Introduction to Random Vibrations, Spectral and
Wavelet Analysis, Longman Scientific & Technical, Essex, U.K., third
edition, 1994.
vectors. It is obvious that classification parameters achieved [6] A. Procházka, J. Jech, and J. Smith, “Wavelet transform use in signal
both by the DFT and DWT provide similar results but features processing,” in 31st International Conference in Acoustics. 1994, pp.
of the dominant cluster are closer together in the case of the 209–213, Czech Technical University.
[7] S. Haykin, Neural Networks, A Comprehensive Foundation, Macmillan
DWT use in most cases. College Publishing Company, New York, 1994.
[8] D. I. Choi and S. H. Park, “Self-Creating and Organizing Neural
V. CONCLUSION Networks,” IEEE Trans. Neural Networks, vol. 5, no. 4, pp. 561–575,
July 1994.
The contribution presents the use of wavelet transform [9] Saeid Sanei and J. A. Chambers, EEG Signal Processing, Wiley -
Interscience, 2007.
for a given signal classification. Mathematical basis of the [10] M. Nixon and A. Aguado, Feature Extraction & Image Processing,
DWT and the following numerical experiments proved that Elsevier, Amsterdam, 2004.
signal analysis based on DWT coefficients can be used very [11] A. Glavinovitch, M. N. S. Swamy, and E. I. Plotkin, “Wavelet-Based
Segmentation Techniques in the Detection of Microarousals in the Sleep
efficiently for the estimation of signal segments features. The EEG,” in 48th Midwest Symposium on Circuits and Systems. 2005, pp.
paper presents results of a classification for signal segments 1302–1305, IEEE.
detected by the the Bayesian approach [2] for a selected [12] P. Johankhani, V. Kodogiannis, and K. Revett, “EEG Signal Classifi-
cation Using Wavelet Feature Extraction and Neural Networks ,” in
EEG channel. The following signal classification assumed the IEEE John Vincent Atanasoff 2006 International Symposium on Modern
knowledge of the range of the number of classes to apply Computing (JVA06). 2006, pp. 120–124, IEEE.
a self-creating classification method [8] to find their optimal [13] C. Dimoulas, G. Kalliris, G. Papanikolaou, and A. Kalampakas,
“Long-Term Signal Detection, Segmentation and Summarization Us-
value and to exclude the possibility of dead neurons. Results ing Wavelets and Fractal Dimension: A Bioacoustics Application in
of signal classification can be further improved by various Gastrointestinal-Motility Monitoring,” Comput. Biol. Med., vol. 37, no.
methods but one the most important problems is in the right 4, pp. 438–462, 2007.
[14] I. W. Selesnick, R. G. Baraniuk, and N. G. Kingsbury, “The Dual-Tree
definition of signal segment features using both its frequency- Complex Wavelet Transform,” IEEE Signal Processing Magazine, vol.
domain and time-domain properties and it seems that the DWT 22, no. 6, pp. 123–151, 2005.
can be used in this area very efficiently. Further studies will be [15] Nazareth P. Castellanos and Valeri A. Makarov, “Recovering EEG
Brain Signals: Artifact Suppression with Wavelet Enhanced Independent
also devoted to the use of principal component analysis to find Component Analysis,” Journal of Neuroscience Methods, vol. 158, no.
fundamental signal components from all observed channels for 2, pp. 300–312, 2006.
the following global segmentation.
Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on July 18,2022 at 12:50:15 UTC from IEEE Xplore. Restrictions apply.