Human Emotion Classification from EEG Signals ...

2014 International Conference on Medical Biometrics

Human Emotion Classification from EEG Signals using Multiwavelet Transform Varun Bajaj

Ram Bilas Pachori

Discipline of Electronics and Communication Engineering PDPM Indian Institute of Information Technology, Design and Manufacturing Jabalpur, 482005 Jabalpur India Email: [email protected]

Discipline of Electrical Engineering Indian Institute of Technology Indore 452017 Indore, India Email: [email protected] classifier and support vector machines (SVMs) respectively. The time-frequency based features have been used as input to SVM classier for classification of three emotional states and obtained average classification accuracy was 63% [8]. The combination of surface Laplacian (SL) filtering, wavelet transforms (WT) and linear classifier has been used for emotion classification from EEG signals [9]. The reported classification accuracies are 83.04% and 79.17% for k-nearest neighbors (kNN) and linear discriminant analysis (LDA) respectively. The short time Fourier transform based features have been used as input for SVM classifier for emotion classification [10]. The experimental results provided the classification accuracy of 82.29% by considering four emotions using SVM classifier. The higher order crossing based features have been used for emotion classification from EEG signals with 83.33% classification accuracy for six emotions [11]. Time domain and frequency domain based features have been used for emotion classification using EEG signals with 66.5% classification rates for four emotions [12]. The spectrogram, Zhao-Atlas-Marks and Hilbert-Huang spectrum based features have been used for classification of arousal and neutral with 86.52% classification rate [13].

Abstract—In this paper, we propose new features based on multiwavelet transform for classification of human emotions from electroencephalogram (EEG) signals. The EEG signal measures electrical activity of the brain, which contains lot of information related to emotional states. The sub-signals obtained by multiwavelet decomposition of EEG signals are plotted in a 3-D phase space diagram using phase space reconstruction (PSR). The mean and standard deviation of Euclidian distances are computed from 3-D phase space diagram. These features have been used as input features set for multiclass least squares support vector machines (MC-LS-SVM) together with the radial basis function (RBF), Mexican hat wavelet and Morlet wavelet kernel functions for classification of emotions. The proposed features based on multiwavelet transform of EEG signals with Morlet wavelet kernel function of MC-LS-SVM have provided better classification accuracy for classification of emotions. Keywords—Electroencephalogram (EEG) signal, Multiwavelet transform, Classification of emotions, Multiclass least squares support vector machines (MC-LS-SVM), Phase space reconstruction (PSR).

I.

I NTRODUCTION

The electroencephalogram (EEG) signal contains valuable information related to the different emotional states, which help us to understand physiology and psychology of the human brain [1]. The emotions measured by EEG signals are more advantageous because it is difficult to influence the electrical activity of the brain intentionally. The emotion classification using EEG signals can help us to develop and improve brain computer interface (BCI) system [2]. The BCI system can be used for assisting the physically disabled and impaired people to interact with the real world [3].

In this paper, we propose new features based on multiwavelet decomposition for classification of emotions from EEG signals. The features extracted from Euclidian distance of 3-D phase space representation of sub-signals obtained by multiwavelet decomposition have been used as input feature set for classification of emotions. The paper is structured as follows: Section II presents the experimental setup, preprocessing, the multiwavelet transform, Euclidian distance based features from 3-D phase space reconstruction and MCLS-SVM classifier. The experimental results and discussion for the classification of emotions from EEG signals based on the proposed method are provided in Section III. Finally, Section IV concludes the paper.

The parameters extracted from EEG signals are useful for recognition and classification of emotions. The chaoticity of different emotional states of brain activity using EEG signals can be measured by the help of Kolmogorov entropy and the principal Lyapunov exponent parameters [4]. The fractal dimension has been used as input features for classification of human emotions [5]. The statistical and energy based features extracted by using discrete wavelet transform of EEG signals have been used for classification of human emotions [6].

II. A. Experimental Setup

The EEG signals were collected from 8 healthy subjects (4 males and 4 females; age 20-35 years) during audio-video stimulus. The subjects were undergraduate students or employees from Indian Institute of Technology Indore, India. A 16channel EEG module (BIOPAC system, Inc.) in compliance to the international 10-20 system was used for recording of EEG

Recently, event related potential and event related oscillation based features have been used as input feature set for classification of emotions [7]. The achieved classification rates were 79.5% and 81.3% for mahalanobis distance (MD) based 978-1-4799-4013-4/14 $31.00 © 2014 IEEE DOI 10.1109/ICMB.2014.29

M ETHODOLOGY

125

[φ1 , φ2 , ...., φr ]T , where T represents the transpose, and the associated r-wavelets Ψ(t) = [ψ1 , ψ2 , ...., ψr ]T satisfies the following matrix dilation and matrix wavelet equation [22-23]: Φ(t) = H[k]Φ(2t − k) (1)

signals. The sampling frequency of EEG signals was 1000 Hz with bipolar montage for this study. The different ways of inducing emotions are: visual includes images and pictures [14], recalling of past emotional events [15], audio may be songs and sounds [16], audiovideo includes film clips and video clips [17, 18]. Many of researchers have used visual stimulus for inducing emotions. It has been shown that audio-visual stimulus is more effective for inducing emotions than the visual stimulus [18]. This study examined four basic emotional states following a 2D valence-arousal emotion model [19-20], including happy, neutral, sadness, and fear. In the experiment, 3 audio-video stimulus with 5 trials of each emotion from eight subjects with F3/F4 channel of EEG signals were used. The 480 EEG signals were used with the duration of 2 seconds. The example of the EEG signals of different emotional states is shown in Fig. 1.

k

Ψ(t) =

G[k]Φ(2t − k)

(2)

k

Where the coefficients H[k] and G[k] are r ×r matrices instead of scalars. The matrices G[k] and H[k] are low-pass filter and high-pass filters for multiwavelet filter bank respectively. In current work, three famous multiwavelets are employed, which are Gernoimo-Hardin-Massopust (GHM) [22], ChuiLian (CL) [24] and SA4 [25]. Figures 2-5 show the third level sub-band signals obtained by multiwavelet decomposition of EEG signals shown in Figures 1(a)-1(d) respectively. D. Euclidian Distance based Features from 3-D Phase Space Reconstruction The concept of phase space is an important tool for characterizing high-dimensional dynamic systems. In order to extract the nonlinear dynamics of the sub-signals obtained by multiwavelet, we use the phase spaces reconstruction (PSR) [26-27]. A dynamic system can be illustrated using phase space diagram, which essentially provides a coordinate of the system where these coordinates use as variables for mathematical representation of the system. Where the x-axis, y-axis, and zaxis plot Xs , Xs+1 , and Xs+2 respectively. The phase space reconstruction of sub-signals Xs can be expressed as: Yd = [Xs , Xs+τ , ..., Xs+(m−1)τ ]

Fig. 1. Example of EEG signals of different emotional states: (a) happy, (b) neutral, (c) sad, and (d) fear.

(3)

where s = 1, 2, ..., S and d = 1, 2, ..., S − (m − 1), m and τ denote the dimension of the phase space and the delay time, respectively. Figures 6-9 show the plots of the third level sub-signals obtained by multiwavelet decomposition of EEG signals shown in Figures 2-5 in a 3-D phase space reconstruction respectively. Mean of Euclidean distance (MED): The mean of Euclidian distance of 3-D phase space is defined as:

B. Pre-processing EEG signals are recorded from the different positions on the scalp. They are contaminated with noises due to power line and external interferences and artifacts. The 8th order, band pass, Butterworth filter with a bandwidth 0.5-100 Hz has been used for removing noises and artifacts. The 50 Hz notch filter is employed to remove the power-line contamination. The required pre-processing for multiwavelet transform includes generation of vectored input stream and pre-filtering. The vectored input stream can be obtained by many ways [21]. In this paper, this vectored input stream is obtained from the commonly used repeated row pre-processing scheme. The matrix-valued multiwavelet filter bank requires multiple streams of input instead of one as decided by multiplicity. This can be accomplished by pre-filtering operation on the input stream to produce the required multiple streams.

S−2

M ED =

E(s)

s=1

(4) S−2 Standard deviation of Euclidean distance (SED): The standard deviation of Euclidian distance of 3-D phase space is expressed as: ⎛ S−2 ⎞ 12 2 (E(s) − M ED) ⎟ ⎜ ⎜ s=1 ⎟ ⎜ ⎟ SED = ⎜ (5) ⎟ S − 2 ⎝ ⎠

C. Multiwavelet Transform Where E(s) is the Euclidean distance between the origin (0, 0, 0) and the point (Xs , Xs+1 , Xs+2 ) of phase space reconstruction. It can be defined as: 2 2 + Xs+2 (6) E(s) = Xs2 + Xs+1

The multiwavelet is the extension of idea of scalar wavelet where multiple scaling functions and associated multiple wavelets are used instead of single. The multi-wavelet can be considered as vector-valued wavelets which satisfy the condition of two scale relation with involvement of matrices rather than scalars. The vector-valued scaling function Φ(t) =

126

Fig. 6. The 3-D phase space diagram of sub-band signals obtained by third level multiwavelet decomposition of EEG signal shown in Fig. 2.

Fig. 2. The third level sub-band signals obtained by multiwavelet decomposition of EEG signal shown in Fig. 1(a)


Fig. 3. The third level sub-band signals obtained by multiwavelet decomposition of EEG signal shown in Fig. 1(b)


Fig. 4. The third level sub-band signals obtained by multiwavelet decomposition of EEG signal shown in Fig. 1(c)


Fig. 5. The third level sub-band signals obtained by multiwavelet decomposition of EEG signal shown in Fig. 1(d)

127

Ωki = yik y k gkT (x)gk (xi ) + γ −1 I

E. Multiclass Least Squares Support Vector Machines The least squares support vector machines (LS-SVM) are a group of supervised learning methods that can be used for classification of two data set [28-30]. For multiclass case we k i=P,k=m have considered the input training data {xi }i=P i=1 , {yi }i=1,k=1 is the output training data of the k th output unit for pattern i. The derivation of the MC-LS-SVM can be expressed as [31]:

1 = [1, ..., 1] bM = [b1 , ..., bm ] αi,k = [α1,1 , ..., αP,1 ; ...; α1,m , ..., αP,m ] Where Kk (x, xi ) = gkT (x)gk (xi ) is kernel function, which satisfy Mercer condition [31]. The decision function of MCLS-SVM is expressed as [32]: P (k) f (x) = sign αik yi Kk (x, xi ) + bk (11)

m 1

P m γ 2 (m) wkT wk + ei,k Minimize JLS (wk , bk , ei,k ) = 2 2 i=1 k=1 k=1 (7) the applied equality constraints are:

i=1

⎧ 1 T yi [w1 g1 (xi ) + b1 ] = 1 − ei,1 , i = 1, 2, ..., P ⎪ ⎪ ⎪ 2 T ⎪ ⎪ ⎨ yi [w2 g2 (xi ) + b2 ] = 1 − ei,2 , i = 1, 2, ..., P . . ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ .m T yi [wm gm (xi ) + bm ] = 1 − ei,m , i = 1, 2, ..., P

The radial basis function (RBF) kernel for MC-LS-SVM can be expressed as [31]: −||x − xi ||2 (12) Kk (x, xi ) = exp 2σk2

(8)

Where P denotes the number of training dataset. The ei,m and bk denotes the classification error and the bias, respectively. wk and γ are the weight vector of k th classification error and the regularization factor, respectively. The gk (.) is a nonlinear function and it maps the input space into a higher dimensional space. The Lagrangian multipliers αi,k can be defined for (7) as: (m) (k) L(m) (wk , bk , ei,k ; αi,k ) = JLS − αi,k {yi [wkT gk (xi ) (9)

which provides the following conditions for optimality: ⎧ P ⎪ (k) ⎪ ∂L ⎪ = 0, −→ w = αi,k yi gk (xi ) ⎪ k ∂wk ⎪ ⎪ ⎪ i=1 ⎪ ⎪ P ⎨ (k) ∂L (10) αi,k yi = 0 ∂bk = 0, −→ ⎪ ⎪ i=1 ⎪ ⎪ ∂L ⎪ ⎪ ⎪ ∂ei,k = 0, −→ αi,k = γei,k ⎪ ⎪ ⎩ ∂L = 0, −→ y (k) [wT g (x ) + b ] = 1 − e k i,k i k k i ∂αi,k

YM

⎥ ⎢ ⎥ ⎢ ⎥ , ..., ⎢ ⎥ ⎢ ⎦ ⎣

(m)

y1 . . . (m) yP

R ESULTS AND D ISCUSSION

In the proposed method, the GHM, CL, and SA4 multiwavelets have been employed for obtaining the third level decomposed sub-signals of the EEG signals. The phase space reconstruction of sub-signals is more useful for higher dimensional representation, which facilitates to compute the Euclidian distance based features from decomposed sub-signals obtained by multiwavelet decomposition of EEG signals. The mean and standard deviation of Euclidian distance of 3-D phase space reconstruction is summarized the complexity of decomposed sub-signals obtained by multiwavelet decomposition of EEG signals. These features have been used as an input to the MC-LS-SVM classifier with the RBF kernel, Mexican hat and Morlet wavelet kernel function. The trial and error approach was used for selection of optimal kernel parameters.

with the following matrices: ⎡

Similarly, The kernel function of Morlet mother wavelet for MC-LS-SVM can be expressed as [33]: d (xl − xli ) xl − xli 2 Kk (x, xi ) = cos ω0 exp − ak 2a2k l=1 (15) where xli is the lth component of ith training data. III.

where i=1, 2, ..., P and k=1, 2, ..., m. Elimination of wi and ek,i provides the linear system as: T 0 bM 0 YM = αM YM ΩM 1

⎤

l=1

The kernel function of Mexican hat mother wavelet for MCLS-SVM can be expressed as [33]: d (xl − xli )2 xl − xli 2 Kk (x, xi ) = 1− exp − (14) a2k 2a2k l=1

i,k

+bk ] − 1 + ei,k }

⎧⎡ (1) ⎪ y1 ⎪ ⎪ ⎪ ⎢ ⎨⎢ . = blockdiag ⎢ ⎢ . ⎪ ⎪ ⎣ . ⎪ ⎪ ⎩ (1) yP

where σk controls the width of RBF function. The multidimensional wavelet kernel function for MC-LSSVM can be given as [32]: l d x − xli ψ Kk (x, xi ) = (13) ak

⎤⎫ ⎪ ⎪ ⎥⎪ ⎪ ⎥⎬ ⎥ ⎥⎪ ⎦⎪ ⎪ ⎪ ⎭

The classification performance of the MC-LS-SVM classifier for emotion classification can be determined by computing the classification accuracy, 10-fold cross-validation, and confusion matrix. The classification accuracy can be defined as

ΩM = blockdiag{Ω1 , ..., Ωm }

128

TABLE II.

T HE CONFUSION MATRIX OF M ORLET WAVELET KERNEL FUNCTION OF THE MC-LS-SVM CLASSIFIER FOR CLASSIFICATION OF EMOTION FROM EEG SIGNALS WITH CL MULTIWAVELETS .

the ratio of the number of epochs correctly classified to the total number of epochs. In 10-fold cross-validation a dataset, Y is randomly divided into 10 disjoint subsets Y1 , Y2 , ..., Y10 of nearly uniform size of each class. Then, the method is repeated 10 times and at every time, the test set is formed from one of the 10 subsets and remaining 9 subsets are used to form a training set. Then the average error across all 10 trials is computed in order to obtain the final classification accuracy. A confusion matrix contains information about the actual and predicted classifications performed by a classification method. Confusion matrix provides the common misclassifications in the classification of emotions from EEG signals.

Happy Neutral Sad Fear Accuracy (%)

Table I shows the classification accuracy (%) for RBF kernel, Mexican hat and Morlet wavelet kernel functions of the MC-LS-SVM classifier for emotion classification with GHM, CL, and SA4 multiwavelets. It is clear from the Table I that multiwavelet decomposition of EEG signals has provided better classification accuracy as compared to without multiwavelet decomposition of EEG signals. The confusion matrix of Morlet wavelet kernel function of the MC-LS-SVM classifier for classification of emotions with CL multiwavelet decomposition of EEG signals is shown in Table II. The proposed features extracted from sub-signals obtained by CL multiwavelet decomposition of EEG signals with Morlet wavelet kernel function of MC-LS-SVM classifier provides 91.04% accuracy for classification of emotions. IV.

[2]

[3]

[4]

[5]

[6]

[7]

[8] TABLE I.

T HE CLASSIFICATION ACCURACY (%) WITH DIFFERENT KERNELS OF THE MC-LS-SVM CLASSIFIER FOR EMOTION CLASSIFICATION FROM EEG SIGNALS USING DIFFERENT MULTIWAVELETS .

Without Multiwavelet GHM

SA4

CL

Happy

Neutral

Sad

Fear

Total

(Parameters)

(%)

(%)

(%)

(%)

Accuracy

RBF (σk =1)

76.94

Mexican hat (ak =1) Morlet (ω0 =0.7, ak =1) RBF (σk =1)

77.5

70.27

67.22

71.38

71.59

78.33

71.11

69.72

70.55

72.42

Mexican hat (ak =1) Morlet (ω0 =0.7, ak =1)

71.11

68.05

71.66

84.16

91.66

90.83

85.83

88.12

69.44

68.33

78.33

71.31

85.83

90.83

91.67

86.67

88.75

RBF (σk =1)

88.33

Mexican hat (ak =1) Morlet (ω0 =0.7, ak =1) RBF (σk =1)

87.5

75.00

73.06

65.83

75.35

88.33

75.00

82.50

80.00

81.46

87.50

86.67

88.33

96.67

89.79

Mexican hat (ak =1) Morlet (ω0 =0.7, ak =1)

70.00

80.33

79.17

[9]

71.94

69.17

[10]

[11]

79.46

78.33

76.67

78.33

93.33

81.66

88.33

86.67

90.83

98.33

91.04

Sad 3.33 5.83 90.83 0.00 90.83

Fear 0.00 0.00 1.67 98.33 98.33

R EFERENCES [1]

C ONCLUSION

Kernel Function

Neutral 5.83 86.67 7.50 0.00 86.67

Classification of EEG Signals based on Nonlinear and Nonstationary Signal Models”, project no. SR/FTP/ETA-90/2010 is greatly acknowledged. This research work was carried out at Discipline of Electrical Engineering, Indian Institute of Technology Indore, Indore, India.

We have developed multiwavelet decomposition based features for classification of emotions from EEG signals. The multiwavelet offers the advantage of simultaneously exhibit orthogonality, short support, symmetry, and second order approximation. The mean and standard deviation of Euclidian distances based on 3-D phase space reconstruction of subsignals have been useful to measure the emotional changes of brain. Finally, we conclude that multiwavelet based features are effective for automatic emotion classification. The classification results indicated that Morlet wavelet kernel function of MC-LS-SVM classifier with CL multiwavelet had provided 91.04% accuracy for classification of emotions from EEG signals.

Multiwavelet

Happy 88.33 6.67 5.00 0.00 88.33

[12]

[13]

ACKNOWLEDGEMENTS [14]

Financial support obtained from the Department of Science and Technology (DST) India, project titled “Analysis and

129

M.K. Kim, M. Kim, E. Oh, and S.P. Kim, “A review on the computational methods for emotional state estimation from the human EEG,” Computational and Mathematical Methods in Medicine, Article ID 573734, 2013. H-II. Suk and S.W. Lee, “A novel bayesian framework for discriminative feature extraction in brain-computer interfaces,” IEEE Trans. Patt. Anal. Mach. Intell., vol. 35(2), pp. 286-299, 2013. M. Murugappan, M. Rizon, R. Nagarajan, and S. Yaacob, “Inferring of human emotional states using multichannel EEG,” European Journal of Scientific Research, vol. 48(2), pp. 281-299, 2010. L.I. Aftanas, N.V. Lotova, V.I. Koshkarov, V.L. Pokrovskaja, S.A. Popov, and V.P. Makhnev, “Non-linear analysis of emotion EEG: calculation of Kolmogorov entropy and the principal Lyapunov exponent,” Neurosci. Lett., vol. 226(1), pp. 13-16, 1997. R. Boostani and M.H. Moradi, “A new approach in the BCI research based on fractal dimension as feature and Adaboost as classifier,” J. Neural Eng., vol. 1(4), pp. 212-217, 2004. M. Murugappan, N. Ramachandran, and Y. Sazali, “Classification of human emotion from EEG using discrete wavelet transform,” J. Biomed. Science Eng., vol. 3, pp. 390-396, 2010. C.A. Frantzidis, C. Bratsas, C.L. Papadelis, E. Konstantinidis, C. Pappas, and P.D. Bamidis, “Toward emotion aware computing: an integrated approach using multichannel neurophysiological recordings and affective visual stimuli,” IEEE Trans. Inf. Technol. Biomed., vol. 14(3), pp. 589597, 2010. G. Chanel, J.J.M. Kierkels, M. Soleymani, and T. Pun, “Short-term emotion assessment in a recall paradigm,” Int. J. Human-Computer Studies, vol. 67(8), pp. 607-627, 2009. M. Murugappan, R. Nagarajan, and S. Yaacob, “Combining spatial filtering and wavelet transform for classifying human emotions using EEG signals,” Journal of Medical and Biological Engineering, vol. 31(1), pp. 45-51, 2010. Y.P. Lin, C.H. Wang, T.P. Jung, T.L. Wu, S.K. Jeng, J.R. Duann, and J.H. Chen, “EEG-based emotion recognition in music listening,” IEEE Trans. Biomed. Eng., vol. 57(7), pp. 1798-1806, 2010. P.C. Petrantonakis and L.J. Hadjileontiadis, “Emotion recognition from EEG using higher order crossings,” IEEE Trans. Inf. Tech. Biomed., vol. 14(2), pp. 186-197, 2010. X.W. Wang, D. Nie, and B.L. Lu, “EEG-based emotion recognition using frequency domain features and support vector machines,” in Neural Information Processing of Lecture Notes in Computer Science, 2011, pp. 734-743. S.K. Hadjidimitriou and L.J. Hadjileontiadis, “Toward an EEG-based recognition of music liking using time-frequency analysis,” IEEE Trans. Biomed. Eng., vol. 59(12), pp. 3498-3510, 2012. P.C. Petrantonakis and L.J. Hadjileontiadis, “Adaptive emotional information retrieval from EEG signals in the time-frequency domain,” IEEE Trans. Sig. Proc., vol. 60(5), pp. 2604-2616, 2012.

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24] [25]

[26] [27] [28] [29] [30]

[31]

[32]

[33]

G. Chanel, J. Kronegg, D. Grandjean, and T. Pun, “Emotion assessment: arousal evaluation using EEG’s and peripheral physiological signals,” in B. Gunsel, A. M. Tekalp, A. K. Jain, B. Sankur (Eds.), Multimedia Content Representation Classification and Security. Springer Lectures Notes in Computer Sciences, 2006, pp. 530-537. J. Kim and E. Andre, “Emotion recognition based on physiological changes in music listening,” IEEE Trans. Pattern. Anal. Mach. Intell., vol. 30(12), pp. 2067-2083, 2008. S. Koelstra, C. Muhl, M. Soleymani, J.S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras, “DEAP: A database for emotion analysis using physiological signals,” IEEE Trans. Affective Comput., vol. 3(1), pp. 18-31, 2012. M. Murugappan, M.R.B.M. Juhari, R. Nagarajan, and S. Yaacob, “An investigation on visual and audiovisual stimulus based emotion recognition using EEG,” J. Medical Eng. Informatics, vol. 1(3), pp. 342356, 2009. R. Cowie, E. Douglas-Cowie, S. Savvidou, E. McMahon, M. Sawey, and M. Schroder, “’Feeltrace’: an instrument for recording perceived emotion in real time,” in Proc. ISCA Workshop, Speech and Emotion, Northern Ireland, 2000, pp. 19-24. E. Niedermeyer and F.L. Silva, “Electroencephalography basic principles, clinical applications, and related fields,” Baltimore, MD Williams and Wilkins, 1993. V. Strela, P.N. Heller, G. Strang, P. Topiwala, and C. Heil, “The application of multiwavelet filter banks to image processing,” IEEE Trans. Image Process., vol. 8(4), pp. 548-563, 1999. J.S. Geronimo, D.P. Hardin, and P.R. Massopust, “Fractal functions and wavelet expansions based on several functions,” J. Approximation Theory, vol. 78, pp. 373-401, 1994. J. Qumar and R.B. Pachori, “A novel technique for merging of multisensor and defocussed images using multiwavelets,” in Proc. IEEE Tencon Conference, Melbourne, Australia, 2005, pp. 1733-1738. C.K Chui and J. Lian, “A study of orthonormal multi-wavelets,” Applied Neumrical Methamatics, vol. 20, pp. 273-298, 1996. J.Y. Tham, L. Shen, S.L. Lee, and H.H. Tan, “A general approach for analysis and application of discrete multiwavelet transforms,” IEEE Trans. Signal Process., vol. 48(2), pp. 457-464, 2000. F. Taken, “Detecting stange attractor in turbulance,” in Proc. Dyamics System and Turbulance, 1980, pp. 366-381. H. Kantz and T. Schreiber, “Nonlinear Time Series Analysis,” Cambridge University Press, 1997. V. Vapnik, The Nature of Statistical Learning Theory, New York: Springer-Verlag, 1995. J.A.K. Suykens and J. Vandewalle, “Least squares support vector machine classifiers,” Neural Process. Lett., vol. 9(3), pp. 293-300, 1999. V. Bajaj and R.B. Pachori, “Classification of seizure and nonseizure EEG signals using empirical mode decomposition,” IEEE Trans. Info. Tech. Biomed., vol. 16(6), pp. 1135-1142, 2012. J.A.K. Suykens and J. Vandewalle, “Multiclass least squares support vector machines,” in Proc. of the International Joint Conference on Neural Networks, Washington DC, USA, 1999, pp. 900-903. Y. Xing, X. Wu, and Z. Xu, “Multiclass least square wavelet support vector machines,” in IEEE International Conference on Networking, Sensing and Control, Sanya, 2008, pp. 498-502. V. Bajaj and R.B. Pachori, “Automatic classification of sleep stages based on the time-frequency image of EEG signals,” Comput. Meth. Prog. Bio., vol. 112(3), pp. 320-328, 2013.

130