International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 INTERNATIONAL JOURNAL OF ELECTRONICS AND

D 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

COMMUNICATION ENGINEERING & TECHNOLOGY (IJECET)

ISSN 0976 6464(Print) ISSN 0976 6472(Online) Volume 4, Issue 5, September October, 2013, pp. 73-79 IAEME: www.iaeme.com/ijecet.asp Journal Impact Factor (2013): 5.8896 (Calculated by GISI) www.jifactor.com

IJECET
IAEME

WAVELET BASED EPILEPSY DETECTION IN EEG USING NEURAL NETWORK

Shah Aqueel Ahmad Royal Institute of Tech. & Science, Hyderabad 501503, India Dr. Syed Abdul Sattar Royal Institute of Tech. & Science, Hyderabad 501503, India

Dr. D. Elizabath Rani GITAM University, Vishakapatnam, AP, India

ABSTRACT One of the many challenges in automated detection of seizure is to draw a line of demarcation between seizure activity and non seizure activity. To accomplish this task, identification of related features and their extraction from EEG plays a key role. The aim of this work is to develop a new method for automatic detection and classification of EEG patterns using Discrete wavelet transform and neural network. A wavelet based technique has been developed to extract features, energy covariance, entropy, power spectral density min and maximum. These feature have been applied to neural network for training and classification of epileptic and non epileptic. The good accuracy is achieved. Keywords: Electroencephalogram (EEG) signals, epileptic seizures, discrete wavelet transform (DWT), artificial neural network (ANN). 1. INTRODUCTION Brain is the most complex of all the other organs of human body. It generates electrical signals to directly or indirectly control the entire body. The electrical activity of the brain generated by millions of neurons is recorded by medical technique known as electroencephalogram (EEG). An EEG channel is formed by taking the difference between potentials measured by placing two electrodes, and records the summed potential of neurons. EEG is measured using 10-20 electrode system. EEG is used for diagnosis of brain disorders like coma, to verify brain death, to monitor anesthetic depth etc. One of the main applications of EEG lies in diagnosing epilepsy. A patient suffering from epileptic attack has a distinctly different EEG as compared to excessive neural activity in the brain.
73

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

The detection of epilepsy, which includes visual scanning of EEG recordings for the spikes and seizures, is very time consuming, especially in the case of long recordings. In addition, biosignals are highly subjective so disagreement on the same record is possible. These EEG signal parameters extracted and analyzed using computers, are highly useful in diagnostics. The EEG signals, like most biological signals, are inherently difficult to quantify. These signals can be characterizes as non stationary. The EEG signals, like most biological signals, are inherently difficult to quantify. These signals can be characterized as non stationary. 2. EXPERIMENTAL DATA The EEG data used in the work is available at the Department of Epileptology, University of Bonn . The complete data set consists of five sets (Z, O, F, N, S), each containing 100 single channel EEG signals. Z is obtain from healthy subject with eyes closed and S is taken from ictal subject These segments of EEG signals were cut out from continuous multi-channel EEG recordings and that too after taken care for artifacts like muscle activity or eye movements. Each signal segment is of 23.6 s duration and Sampling frequency is 173.61 Hz, so each segment contains N = 4096 Samples. 3. DISCRETE WAVELET TRANSFORM Wavelet transform can be defined as a spectral estimation technique in which any general function can be expressed as a sum of an infinite series of wavelets. The decomposition of the signal results in a set of coefficients called wavelet coefficients. The decomposition is done by low pass and high pass filter in time domain. DWT successfully analyses the multi-resolution signal at different frequency bands, by decomposing the signal into approximation and detail information. Here Data is first preprocessed by removing dc component from the signal thereby achieving different levels of decomposition for Daubechies order-2 wavelet with a sampling frequency of 173.6 Hz on each signal of 4096 samples. Figure 1 shows the decomposition of the EEG signal x(n) at Nyquist frequency which is the maximum useful frequency i.e. half of sampling frequency 86.8 Hz. Signal is subjected to six level decomposition where each stage of this scheme consists of two digital filters. The first filter, h[n] is the high pass filter or it is the discrete mother wavelet and the second, g[n] is low-pass filter. After the first level of decomposition, the EEG signal (0-86.81Hz), is decomposed into its lower resolution components, CA1 (0-43.4) Hz and higher resolution components, CD1 (43.4-86.81 Hz). by CAi and CDi where i is 1,2,3.... at each level of decomposition. The data after decomposition to various levels has approximation and detail coefficients at each level of different frequency resolution. These frequencies are to be then recombined in order to achieve subbands delta, theta, alpha, beta, gamma at their perspective frequencies. The delta band is of frequency range 0-4 Hz so lower resolution components i.e. approximation coefficients CA7 (0-2.7 Hz) and CA8 (2.7- 4.05 Hz) are to be recombined. But the problem arises when the number of samples of CA7 and CA8 mismatch. CA7 has 128 samples while CA8 has 64 samples. For this up-sampling of CA8 is done to 128 samples and then the DELTA band is reconstructed successfully by recombination of CA7 and up-sampled CA8 through Inverse Discrete Wavelet Transform using Daubechies order-2. Similarly THETA subband is of frequency range 4-8 Hz so CA9 and up-sampled CD8 are recombined. After recombining all coefficients falling in their respective subbands, delta, theta, alpha band with 256 samples and beta with 1024 and gamma with 4096 samples are achieved as shown in Table 1.

74

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

Table 1. Subbands with their Frequency Range, Recombining Coefficients and no of Samples Subband DELTA THETA ALPHA BETA GAMMA Frequency 0-4 Hz 4-8 Hz 8-12 Hz 12-30 Hz > 30 Hz Coefficients CA7,CA8(up-sampled) CD8(up-sampled), CA9 CD9, CA11(up-sampled) CD11(up-sampled), CD10, CD4(doubly up-sampled) Samples 256 256 256 1024,CD6, CA 4 4096 CD1

4. FEATURE EXTRACTIONS In this study, the discrete wavelet transform is used as a primary computational tool for extracting features of the epileptic EEG signals at different resolutions. Decomposition of Epileptic and Non- Epileptic data into Delta, Theta, Alpha, Beta, Gamma subbands are shown in figure 1 and Figure 2 resp. It is apparent from figures that the amplitude of gross epileptic signal is considerably higher than non-epileptic one. Also the amplitudes of subbands are significantly high in case of epileptic data, especially in gamma subband. 1. Covariance Covariance is defined as a measure of the dispersion of a set of data points around their meanvalue. Assume some random variable X that have the sample values of each EEG subband signal. Let the sample valueof X is Xi= {X1,X2, ....... Xn}. Where i represent a sample set from the subbands delta, theta, alpha, beta, and gamma. The covariance can thus be expressed as

where is the mean value of the set X and n is the number of samples in the EEG dataset. 2. Energy The energy of the signal is defined as the sum of squared modulus of the sample values of any signal. The wavelet based energy of each decomposed subbands such as delta, theta, alpha beta and gamma can be calculated using the formula

where X is the samples values in each subbands and N is the total number of samples. 3. Power Spectral Density (PSD) The PSD is the amount of power per unit frequency as a function of frequency. Periodogram is commonly used for computing PSD. This is computed by squared modulus of the Fourier transform of the time series of the signal.
75

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

||2 Min(w) =

The Maximum and minimum values are estimated from the PSD of each EEG subbands can be considered as feature for classification. 4. Entropy Entropy is a numerical measure of the randomness of a signal. Entropy can act as a feature and used to analyze psychological time series data such as EEG data.

Entropy is the statistical descriptor of the variability within the EEG signal and can be a strong feature for epilepsy detection. The above statistical features are extracted from each subband i.e.Delta, Theta, Alpha, Beta, Gamma of each set to represent the time-frequency distribution of the EEG signals and are nomenclatured as 1. Covariance of the coefficients in each subband (y1, y2, y3, y4) 2. Energy of the wavelet coefficients in each subband (E1,E2,E3, E4) 3. Power Spectral Density (max) and Power Spectral Density (min) of each subband (PSDmax1, PSDmax2, PSDmax3,PSDmax4, ,PSDmin1, PSDmin2, PSDmin3, PSDmin4) 4. Entropy of wavelet coefficients in each subband (Eadelta,Eatheta ,Eaalpha,EaBeta) The same thing is reflected in the figure 9.

Fig. 2: Block Diagram of Epileptic Seizure Classification systems

76

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

5. Proposed Algorithm Step 1: Read EEG Signal Step 2: Using Wavelet Transform extract Delta, Theta, Alpha and Beta Step 3: Extract Energy E from Delta, Theta, Alpha, Beta. Step 4: Extract the Covariance Step 5: Extract the Entropy Step 6: Extract the Minimum Power Spectral Density. Step7: Extract the features of maximum Power spectral Density Step 8: Input the data of extracted features in Neural Network. Step 9: Train the Neural Network Step 10: Apply Threshold value and perform classification of epilepsy Step 11: End 6. NEURAL NETWORKS Artificial neural networks are computing systems made up of large number of simple, highly interconnected processing elements (called nodes or artificial neurons) that abstractly emulate the structure and operation of the biological nervous system. . Learning in ANNs is accomplished through special training algorithms developed based on learning rules presumed to mimic the learning mechanisms of biological systems. There are many different types and architectures of neural net-works varying fundamentally in the way they learn. In this paper, feed forward back propagation neural network considered. From the decomposed EEG signal the features calculated for each sub band a total of 20 features are considered for classification. Neural network is designed with 20 input nodes and one output node. The network is trained with parameters a Lavenberg Marquardt algorithm with momentum factor included (TRAINLM)was used for training. The stopping criterion was specified to be 0.001 Root Mean Square Error (RMSE). The training was stopped when the RMSE between the network outputs and the targets was lesser than or equal to 0.001. The learning rate was fixed at 0.5. The number of training epochs was fixed uniformly at 1000. . The output from threshold detector value is greater than threshold value the subject considered as epileptic otherwise normal. The threshold value of 0.5 is considered. In this work. Feed Forward back propagation is used. 7. RESULTS AND DISCUSSIONS Each EEG signal is decomposed using DWT .from each sub band the five statistical features Energy, Covariance, Entropy, Power Spectral Density min, and Power Spectral Density max calculated and average values are tabulated in Table-I. This table shows the deviation of values for different subjects. In each sub band the variation of values of features indicting the Presence of epilepsy Features of two different Subjects and each having 300 sets. Epoch duration of 7 second is considered for each type dataset. Therefore 600 feature sets are generated from each subject and the total feature sets are 12000. The target value are fixed as per the type of the subjects. For epileptic subject is assigned a maximum value of 0.9 and normal subject is assigned 0.1. These feature sets are arranged in random manner before neural network training. The trained network is simulated with normal and epileptic subject's data. The performance of the classification is evaluated in terms of specificity, sensitivity and classification accuracy .The performance analysis is tabulated in Table II.

77

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

The Fig2 shows the plot of classification performance of Feed Forward Back Propagation neural network.

Fig. 3: Performance analysis of Epileptic Seizure Classification systems

Work Proposed Method Zakir Hussain and K. S Rao(ICINT 2012) , IPCSIT vol 37 2012[7] Nidal Rafiuddin , Y.U Khan ICMSPCT 2011[3] Abibullaev Berdakh IEEE ICIT 2009[8]

Specificity 100

Sensitivity 99.8

Accuracy 99.899

98.59

97.69

98.32

96.5

96.49

94.39

95.8

Table. 2: Performance evaluation Neural network classifier

78

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 5, September October (2013), IAEME

8. CONCLUSION Diagnosing epilepsy is a difficult task requiring observation of the patient, an EEG, and gathering of additional clinical information. An artificial neural network that classifies subjects as having or not having an epileptic seizure provides a valuable diagnostic decision support tool for neurologists treating potential epilepsy, since differing etiologies of seizures result in different treatments. In this work, classification of EEG signals was examined. The features are extracted using wavelet transform technique. Five features , Energy, Covariance, Entropy and PSD min and max from each sub band of EEG are extracted The generated data from wavelet technique are given to ANN for training of normal and abnormal EEG conditions. The ANN is used to discriminate between the two tasks with a success rate of 99.89%. Further research is needed to find more elaborated memory architectures and its appropriate training algorithms. Neural networks as classifiers have here a high potential because they can compute in real time with a high numbers of features. This characteristic enable the development and construction of transportable devices, improving substantially the quality of life of epileptic patients intractable by medication and that must learn to live with seizures. 9. REFERENCES Abdulhamit Subasia, Ergun Ercelebi "Classification of EEG signals using neural network and logistic regression." In Computer Methods and Programs in Biomedicine (2005) 78, 8799. [2] D. Najumnissa, S. ShenbagaDevi " ntelligent identification and classification of epileptic seizures using wavelet transform" Int. J. Biomedical Engineering and Technology, Vol. 1, No. 3, 2008,293-313. [3] Y Nidal Rafiuddin1Yusuf Uzzaman Khan and Omar Farooq, Feature Extraction and Classification of EEG for Automatic Seizure Detection, IEEE International Conference on Multimedia, Signal Processing and Communication Technologies, pp. 184-187, 2011. [4] EEG time timeseries (epilepticdata)(2005,Nov.). http://www.meb.unibonn.de/epileptologie/science/physik/eegdata.html. [5] J. Gotman., Automatic recognition of epileptic seizures in the EEG, Clinical Neurophysiology, vol. 54, pp. 30 540, 1982. [6] Y.U. Khan, J. Gotman, Electroencephalogram Wavelet based automatic seizure detection iintracerebral, linical Neurophysiology, vol. 114, pp. 899-908, 2003. [7] Sheikh.Jakeer Husain and K.S Rao Epileptic Seizures classification from EEG Signal using Neural Network IPCSIT vol.37 (2012) IACSIT press, Singapore ICINT 2012. [8] Abibullaev Berdakh, Seo Hee Don, Epileptic Seizures Detection using Continuous Time Wavelet Based Artificial Neural Networks2009 IEEE sixth International Conference on Information Technology: New Generation. [9] Alexandros T.Tzallas, M.G. Tsipouras Epileptic Seizure Detection in EEGs Using TimeFrequency Analysis IEEE TRANSACTION ON IT IN BIOMEDICINE, VOL.13, NO.5, SEPTEMBER 2009. [10] G. Hemalatha, Dr.B. Anuradha and V. Adinarayana Reddy, Efficient Extraction of Evoked Potentials from Noisy Background EEG, International Journal of Electronics and Communication Engineering & Technology (IJECET), Volume 4, Issue 1, 2013, pp. 216 - 229, ISSN Print: 0976- 6464, ISSN Online: 0976 6472. [11] Imteyaz Ahmad, F Ansari and U.K. Dey, A Review of EEG Recording Techniques, International Journal of Electronics and Communication Engineering & Technology (IJECET), Volume 3, Issue 3, 2012, pp. 177 - 186, ISSN Print: 0976- 6464, ISSN Online: 0976 6472.
79

[1]