Multiclass epileptic seizure activity classification exploiting statistical modeling of band-specific DWT coefficients of EEG signals

Abstract

Epileptic seizure is often interpreted by the abnormalities in the brain activity and Electroencephalogram (EEG) is a promising tool for identi cation of Epilep-tic seizure. Signal processing methods try to model visual information into few paramters, thus decision making becomes more accurate compared to the meth-ods based on visual observation of EEG where sometimes misinterpretation takes place in disease treatment. Researchers have used di erent signal processing and machine learning algorithms to extract features for seizure activity detection and classi cation. Since EEG is a non-stationary signal, Discrete Wavelet Transform (DWT) has the potential to perform better than conventional time-frequency anal-ysis method. However, detection and classi cation of multiclass EEG signals of epileptic seizure activity originated from di erent parts and state of the brain in the stringent condition is still a challenging task. DWT of the EEG signals is performed and band-speci c gamma and theta DWT coe cients have been cho-sen. A statistical model has been employed to summarize information in Discrete Wavelet Transform (DWT) coe cients and thus form e ective feature set utiliz-ing the parameters of the proposed statistical probability density function (PDF). Rather than taking discrete parameter as feature like wavelet energy or entropy, it is found more rational to use statistical modeling parameters as features since they are being taken from the shape of the entire data class and representing the class in more consistent way. Gaussian statistical model has been found t for this purpose based on visual inspection of superimposed plots of empirical and Gaussian PDFs, cumulative distribution functions (CDFs) in probability-probability (p-p) plot and K-S test result. The goodness of features has been justi ed by one way ANOVA test, Geometrical Separability Index and Bhattacharyya Distance parameters. The feature set is found e ective and e cient for detecting and classifying multi-class EEG signals of epileptic seizure activity when fed to di erent state-of-the-art clas-si ers in stringent condition random selection of training and testing dataset. The performance parameters (accuracy, sensitivity and speci city) achieved using pro-posed scheme are found almost 100% (maximum accuracy of 100% for 3-class and 93% for 5-class) for multi-class classi cation problems and outperformed the stat-of-the-art strategies.