Vous êtes sur la page 1sur 4

Channel Selection from EEG Signals and Application

of Support Vector Machine on EEG Data


Esen YÕldÕrÕm
Mustafa Turan Arslan Server Göksel Eraldemir Adana Science and Technology
Mustafa Kemal University øskenderun Technical University University
Hatay, Turkey Hatay, Turkey Adana, Turkey
mtarslan@mku.edu.tr sgoksel.eraldemir@gmail.com esenyildirim@gmail.com

Abstract—In this study, EEG data recorded during mental (OCD), using DWT. Guler et al. [11] extracted features using
arithmetic operations and silent reading were analyzed by discrete DWT from an EEG dataset, which consists of five sets; first
wavelet transform and feature vectors were obtained. The obtained two sets are recorded from five healthy volunteers with eyes
feature vectors are classified by Support Vector Machines (SVM). open and closed, third and fourth sets are recorded in seizure
Results are given for 26 channels, all recorded channels, and for 10 free intervals from epileptogenic zone and from the opposite
most effective channels. Correlation based feature selection based
algorithm is used for choosing the most effective channels.
hemisphere of the brain and the last set contains seizure
Decreasing the number of channels without compromising the activity. Total accuracy was reported as 98.68 by using
accuracy, is an important issue for real time applications for which a ANFIS for classification. Subasi et al. [12] applied PCA, ICA,
short analysis time is crucial. In this study, mental arithmetic and LDA feature selection methods to feature vectors obtained
silent reading tasks are classified with an accuracy of 90.71%, a from EEG signals and then performed classification analysis
precision rate of 91.03% and F-measure rate of 90.63% on the by applying the SVM classification method to feature vectors
average using 26 channels, whereas the accuracy, precision and F- whose dimensions were reduced. In another study, Panda et al.
measure were 90.44%, 90.61% and 90.08, respectively which were [13] performed a study to detect epilepsy from EEG signals
comparable to that of obtained using all channels, for reduced using DWT. In this study, EEG data recorded during cognitive
number of channels.
tasks were classified by means of Support Vector Machine
Index Terms— EEG, discrete wavelet transform (DWT), (SVM).
classification, support vector machine (SVM). The classification of EEG signals recorded during
cognitive processing has been among the topics of interest in
I. INTRODUCTION recent years. Kottaimalai et al. [14] first applied Principal
Component Analysis (PCA) on EEG data obtained from the
The brain, one of the most important organs in the body, University of Colorado database [12] to reduce the size of
performs its function in the body by producing electrical feature vectors and then applied Artificial Neural Networks
signals during each activity. Electrical signals produced by (ANN) to classify EEG data with reduced dimensions. They
brain can be measured by devices called EEG. The obtained reported 100% accuracy. Amin et al. [15] used the DWT
signals can be recorded in physical devices via special method to extract features from EEG signals collected from 8
software. EEG signals are evaluated in various aspects: the healthy male volunteers at Petronas Technology University.
time of occurrence, amplitude and frequency. EEG signals are Extracted feature vectors are classified by popular machine
known to be high dimensional non-stationary signals [1]. learning algorithms, i.e. SVM, Multilayer Layer Perceptron
Therefore classical signal processing techniques, such as (MLP), k-Nearest Neighbor (k-NN) and Naive Bayes (NB)
Fourier Transform which assumes stationarity within the methods. Extracted features were obtained using the total and
window, are inadequate for EEG signals processing and relative sub-band energies computed from A4 approximation
multiresolution analysis techniques, such as Wavelet coefficient and D1-D4 detail coefficients. It is observed that
Transform, have found many interesting applications in signal the obtained features are classified over accuracy of 98%
processing[2]–[5]. There are diverse studies using Discrete using DVM and MLP. Above 98 % accuracies with features
Wavelet Transform (DWT) and various classification extracted using A4 and D4 sub-bands were reported as the
algorithms in EEG signal analysis [6], [7]. Ocak [8] classification results of SVM and MLP.
automatically identified the epileptic seizure by means of EEG In this study, EEG signals, recorded from 18 healthy
signals using DWT. Adeli et al. [9] analyzed the EEG data of vocational school and engineering faculty students, while
an epileptic patient using the DWT. Hazarika et al. [10] performing mental mathematical operations and silent reading
classified three types of EEG data, which are normal control are used for classification. Previous studies on this database
subjects, patients diagnosed with schizophrenia (SCH) and resulted in accuracies of 89.6%, 90.6%, 89.1% using Bayesian
patients diagnosed with obsessive-compulsive disorder

978-1-5386-1880-6/17/$31.00 ©2017 IEEE


Network, J48 Decision Tree, Bayesian Network, respectively The 50 Hz network noise was automatically filtered by the
[16], [17]. In another study, addition-subtraction and EEG device when the data was collected from the students. A
multiplication-division groups were distinguished by 79.3% low-pass filter with a cut-off frequency of 120 Hz was applied
using k-NN algorithm [12]. This study, classifies these EEG to remove the data from the high frequency noises. In the
signals by means of SVM using the features extracted sorting of aggregated data, the first and last slides are not used
employing bior2.4 wavelets. to save the troubles that may occur due to synchronization.
II. MATERIAL AND METHODS B. Feature Extraction
EEG signals are segmented using 75% overlapping sliding
A. Collection of Data rectangular windows of 1 second length. In the study, to
EEG signals were obtained with 22 different electrodes extract the feature vectors from the EEG signals, the bior2.4
placed according to the international 10-20 system from 18 main wavelet function of the DWT method was used. The
volunteers, 11 vocational school students and 7 students from DWT method divides the EEG signal into approximation and
the engineering faculty, ages ranging from 18 to 25 with an detail coefficients using low and high pass filters. This method
average age of 20.39. has variable wavelet lengths that are wide for low frequencies
and narrow for high frequencies instead of a constant
sinusoidal wave. The bior2.4 wavelet is known as
biorthogonal or double barbed wavelet. These wavelets
contain a pair of scaling functions and associated scaling
functions for analyzing and synthesizing the signals due to
their double spikes [18]. Feature matrix is obtained using
statistical properties [19]; averages of the absolute values of
the coefficients, maximum of the absolute values of the
coefficients, minimum of the absolute values of the
coefficients and standard deviation of the coefficients.
C. Channel Selection
Analysis time is of great importance in real time systems.
On systems with many channels, the number of attributes to
be extracted will be very high, so the processing time of a
window can be a problem. Thus, reducing the number of
Fig.1. Placement of electrodes on the head channels without affecting the success rate will increase the
usability of the system. For this reason, it is important to find
EEG data were collected during a slide show, a total of 60 the most effective channels in classification. In this study,
slides; 30 verbal texts and 30 mathematical operations of Correlation Based Feature Selection (CFS) method was used
13.25 seconds each slide, on a computer screen. Subjects were to find the channels that provide the highest contribution to the
asked to read texts without moving their lips and solve classification accuracy. For channel selection, the following
mathematical operations mentally without using paper or pen. operations are performed in sequence;
Sample slides are given in Figure 2. 1. Selected attributes were found with the help of CFS.
2. Channels of the selected features were determined.
3. Numbers of selected features for each channel were
found.
4. These steps were repeated for each subject to
determine the most commonly used channels for all
subjects.
All channels were analyzed from the channel with the
highest number of features to the channel with the lowest
number of features, taking into consideration the number of
selected features in all subjects. 10 channels with the highest
selected features are selected as the most effective channels.
D. Classification
Fig.2.Sample slides a) Numerical slide example b) Verbal slide example 1) Support Vector Machine
Support Vector Machines is one of the most basic and
Students were informed about importance of focusing on common classification methods used for various tasks. This
making the operations or reading the test instead of feeling an method is proposed by Cortes and Vapnik [20] and is based on
obligation to get the correct answer or to rush to finish reading the principle of structural risk minimization. This method is a
the text. statistical and supervised learning algorithm. SVM uses core
functions to create a plane that will carry the data sets, that are Subject4 95.37 95.44 95.37
hard to separate linearly, into a high dimensional space and Subject5 95.26 95.26 95.26
maximizes the distance between classes [20]. The kernel Subject6 88.70 88.76 88.70
functions are obtained by internal multiplication of the support Subject7 97.33 97.35 97.33
vectors and the feature vectors. In this study, we use Radial Subject8 92.59 92.91 92.58
basis kernel functions, which are the most commonly used Subject9 91.59 91.59 91.59
core functions for SVM[13], [21] Subject10 93.07 93.08 93.07
Subject11 84.74 84.78 84.74
K ( x, x i ) = Φ ( x ).Φ ( x i ) (1) Subject12 87.56 87.69 87.54
Subject13 95.70 95.81 95.70
The non-linear decision function is used to find out which Subject14 97.33 97.34 97.33
class the sample belongs to. Subject15 95.30 95.36 95.29
Subject16 82.07 82.07 82.07
f ( x) = ¦ i =1α i y K ( x, x i ) + d
N
(2) Subject17 94.33 94.70 94.32
i
Subject18 95.81 95.88 95.81
Here are the Lagrange multipliers, are the desired Average 90.71 91.03 90.63

outputs and are the support vectors where


obtained from training samples. The classification results of 26-channel EEG data are
given in Table II. The highest accuracy was obtained with
III. EXPERIMENTAL STUDY 97.33% for the subject 7 and subject 14. The lowest accuracy
was 67.30% for subject 3. The average classification
EEG signals from 18 subjects were collected and then
performance of EEG data with 26 channels was 90.71%. In
feature vectors were extracted from 26 channels using bior2.4
analysis of 26-channel EEG data, precision and F-measure
wavelet function which is one of the main functions of the
were 91.03% and 90.63%, respectively on the average. In a
DWT method.
previous study, Eraldemir et al [16] obtained 26 channel
Processing time is very important in real time applications.
feature vectors using bior2.4 wavelet from EEG signals on the
For this reason, the analysis process to be performed must be
same dataset. The obtained features vectors were classified by
as fast as possible. Thus, reducing the number of features is of
Bayesian Network and J48 Decision Tree. Bayesian Network
importance. Therefore, reducing the number of channels used
classified with accuracy of 89.6% whereas J48 Decision Tree
in feature extraction would help rapid processing. In this
classified with accuracy of 90.6%.
study, 10 most effective channels, that make the best
contribution to the classification accuracy, are determined. TABLE III. ANALYSIS RESULTS OF 10-CHANNEL EEG DATA WITH SVM
Table I shows the names of the 10 most effective channels
we used in our study and the total number of selected features Subject Accuracy Precision F-Measure
using CFS for those channels. The obtained set of feature Subject1 92.96 93.09 93.65
vectors is normalized to the value range [-1, 1]. Then SVM, Subject2 84.81 86.13 82.57
which is a statistical and supervised algorithm, is applied to the Subject3 69.11 69.73 66.87
normalized data set. In order to test the performance of this Subject4 95.56 95.57 95.32
method, a 10-fold cross-validation technique was used. Subject5 95.30 95.30 95.25
Subject6 87.44 87.45 87.55
TABLE I. CHANNEL NAMES AND FEATURE NUMBERS Subject7 96.22 96.32 95.67
Channel Name Number of Features Subject8 93.22 93.25 93.53
Fp1-Fp2 113 Subject9 89.96 89.97 89.86
O1-O2 69 Subject10 92.33 92.34 92.16
T5-T6 52
Subject11 86.59 86.67 87.16
T3-A1 51
T6-A2 49 Subject12 85.30 85.62 84.14
Fp1-A1 43 Subject13 95.30 95.33 95.66
T4-A2 42 Subject14 96.56 96.56 96.49
Fp1-O1 42 Subject15 94.96 95.01 94.58
F8-A2 40 Subject16 82.15 82.15 82.04
F7-F8 39 Subject17 94.81 95.04 93.96
Subject18 95.37 95.44 94.91
TABLE II. ANALYSIS RESULTS OF 26-CHANNEL EEG DATA WITH SVM Average 90.44 90.61 90.08
Subject Accuracy Precision F-Measure
Subject1 93.33 93.47 93.33 The classification results with reduced channels are shown
Subject2 85.33 86.76 85.19 in Table III. SVM yielded the highest performance for subject
Subject3 67.30 70.19 66.08 14 with the accuracy of 96.56% and the lowest performance
for subject 3 with the accuracy of 69.11%. The average Expert Syst. Appl., vol. 36, no. 2, pp. 2027–2036, Mar. 2009.
classification performance of EEG data with 10 channels of 18 [9] H. Adeli, Z. Zhou, and N. Dadmehr, “Analysis of EEG records
students was 90.44%. On reduced data, precision and F- in an epileptic patient using wavelet transform,” J. Neurosci.
measure were 90.61% and 90.08%, respectively on the Methods, vol. 123, no. 1, pp. 69–87, Feb. 2003.
average. On the same dataset, Eraldemir et al. [22] made [10] N. Hazarika, J. Z. Chen, A. C. Tsoi, and A. Sergejew,
channel selection from EEG signals using db4 wavelet. 26- “Classification of EEG signals using the wavelet transform,”
channel EEG data were classified with accuracy of 88.8% by Signal Processing, vol. 59, no. 1, pp. 61–72, 1997.
Bayesian Networks, whereas the accuracy for only 10 most [11] I. Guler and E. D. Ubeyli, “Adaptive neuro-fuzzy inference
effective channels was 90%. Results show that, SVM system for classification of EEG signals using wavelet
classification has the highest accuracy for this dataset. coefficients,” J. Neurosci. Methods, vol. 148, no. 2, pp. 113–
121, Oct. 2005.
IV. CONCLUSION [12] A. Subasi and M. Ismail Gursoy, “EEG signal classification
using PCA, ICA, LDA and support vector machines,” Expert
In this study, EEG signals from 18 volunteers, studying in Syst. Appl., vol. 37, no. 12, pp. 8659–8666, Dec. 2010.
vocational school and engineering faculty, were recorded.
[13] R. Panda, P. S. Khobragade, P. D. Jambhule, S. N. Jengthe, P.
From this bioelectrical measurements, 26 channel feature
R. Pal, and T. K. Gandhi, “Classification of EEG signal using
vectors were obtained by means of bior2.4 wavelet. By wavelet transform and support vector machine for epileptic
analyzing 26-channel feature vectors, the most effective 10 seizure diction,” in 2010 International Conference on Systems in
channels were determined. The average accuracy with 26 Medicine and Biology, 2010, pp. 405–408.
channels was found to be 90.71%, whereas a better average [14] R. Kottaimalai, M. P. Rajasekaran, V. Selvam, and B.
accuracy of 90.44% was obtained for the most effective 10 Kannapiran, “EEG signal classification using Principal
channels. Results show that, SVM has a better classification Component Analysis with Neural Network in Brain Computer
accuracy than other algorithms with feature vectors obtained Interface applications,” in 2013 IEEE International Conference
from EEG signals using bior2.4 wavelet. ON Emerging Trends in Computing, Communication and
Nanotechnology (ICECCN), 2013, pp. 227–231.
REFERENCES [15] H. U. Amin et al., “Feature extraction and classification for EEG
[1] G. Mayer-Kress and S. Layne, “Dimensionality of the human signals using wavelet transform and machine learning
electroencephalogram,” Ann NY Acad Sci, vol. 504, no. 1, pp. techniques,” Australas. Phys. Eng. Sci. Med., vol. 38, no. 1, pp.
62–87, 1987. 139–149, Mar. 2015.
[2] M. Coskun and A. østanbullu, “Analysis of EEG signals by FFT [16] S. G. Eraldemir and E. Yildirim, “Comparison of wavelets for
and Discrete Wavelet Transform,” in XIV. Akademic Techonolgy classification of cognitive EEG signals,” in 2015 23nd Signal
Conference Reports, 2012, pp. 323–328. Processing and Communications Applications Conference
(SIU), 2015, pp. 1381–1384.
[3] S. Savas, N. Topaloglu, and M. Yilmaz, “Data mining and
application examples in Turkey,” J. Sci. Istanbul Commer. [17] S. G. Eraldemir and E. Yildirim, “Classification of simple text
Univ., vol. 11, no. 21, pp. 1–23, 2012. reading and mathematical tasks from EEG,” in 22nd Signal
Processing and Communications Applications Conference
[4] C.-C. Lai and C.-C. Tsai, “Digital Image Watermarking Using
(SIU), 2014, pp. 180–183.
Discrete Wavelet Transform and Singular Value
Decomposition,” IEEE Trans. Instrum. Meas., vol. 59, no. 11, [18] M. Misiti, M. Yves, G. Oppenheim, and J.-M. Poggi, “Wavelet
pp. 3060–3063, Nov. 2010. toolbox,” The MathWorks Inc., vol. 15, no. 21, 1996.
[5] R. J. Martis, U. R. Acharya, and L. C. Min, “ECG beat [19] I. Guler and E. D. Ubeyli, “Performing Statistical Operations on
classification using PCA, LDA, ICA and Discrete Wavelet Feature Vectors Extracted from EEG Signals by Wavelet
Transform,” Biomed. Signal Process. Control, vol. 8, no. 5, pp. Transform,” in Electrical-Electronics-Computer Engineering
437–448, Sep. 2013. Symposium (Eleco), 2004, pp. 230–234.
[6] R. E. J. Yohanes, Wee Ser, and Guang-bin Huang, “Discrete [20] C. Cortes and V. Vapnik, “Machine Learning,” in Machine
Wavelet Transform coefficients for emotion recognition from Learning, vol. 20, Boston: Kluwer Academic Publishers, 1995,
EEG signals,” in 2012 Annual International Conference of the pp. 273–297.
IEEE Engineering in Medicine and Biology Society, 2012, pp. [21] A. Subasi and M. I. Gursoy, “EEG signal classification using
2251–2254. PCA, ICA, LDA and support vector machines,” Expert Syst.
[7] P. Ghorbanian et al., “Identification of Resting and Active State Appl., vol. 37, no. 12, pp. 8659–8666, 2010.
EEG Features of Alzheimer’s Disease using Discrete Wavelet [22] S. Eraldemir, E. Yildirim, S. Yildirim, and Y. Kutlu,
Transform,” Ann. Biomed. Eng., vol. 41, no. 6, pp. 1243–1257, “Classification and Channel Selection for Feature Selection
Jun. 2013. from EEG signals,” in Innovations and Applications in
[8] H. Ocak, “Automatic detection of epileptic seizures in EEG Intelligent Systems Symposium (ASYU), 2014.
using discrete wavelet transform and approximate entropy,”

Vous aimerez peut-être aussi