Detection and classification of multiclass epileptic seizures exploiting EMD-wavelet analysis of EEG signals

Abstract

Epileptic seizure is often interpreted by the abnormalities in the brain activity and Electroencephalogram (EEG) is a promising tool for identification of Epileptic seizure. Signal processing methods try to modei visual information into few parameters, thus decision making becomes more accurate compared to the method based on visual observation of EEG, a source of misinterpretation in disease treat-. ment. Researchers have used different signal processing and machine learning algorithms to extract features for seizure detection and classification . Since, EEG is a non-stationary signal, empirical mode. decomposition. (EMD), and discrete wavelet transform (DWT) have the potential to perform better than the conventional timefrequency analysis method. However, detection and classification of multi class EEG epilepsy originated from different parts and state of the brain in the stringent conditions is still a challenging ta~k. EMD analysis of theEEG signals is performed and the temporal energy contehts of the IMFs is analyzed to select the dominant IMF. Since the dominant IMF vary for different EEG recordings, a mismatch may be produced between training and testing data even for the same class. Therefore, histogram analysis of the dominant IMFs of all classes is performed to develop a criterion for selecting appropriate number of IMFs for each EEG class. For a better time-frequency resolution and more discriminatory behavior, DWT analysis is cllIried 'out on the selected IMFs:" Analyzing the parameters, namely normalized energy, Fourier spectrum and cross-correlation coefficient, only the 4th Level DWT. coefficients of selected IMFs are found reasonable for feature computation. Finally, HOS, such as variance, skewness amd kurtosis of these coefficients are proposed to constitute the feature vector. The reduced feature vector is found effective for detectIng and ciassifying multi-class EEGepilepsy when fed to diferent state-of-the art classifiers, in stringent conditions, such as reduced training data as well as random. selection of training and testing dataset.