A Machine Learning Framework for Epileptic Seizure Detection by Analyzing EEG Signals

Abstract

Electroencephalography (EEG) signals are non-stationary and mixed with artefacts. A clinical finding through observation is relatively difficult and may lead to misinterpretations. In particular, epilepsy is the brain neural disorder which is hard to be diagnosed by visual observation of EEG signals. In an attempt to avoid such key issues, automated detection of epilepsy is proposed by analyzing EEG signals in a systematic way to support the clinical decision making process. Initially the EEG signal data is preprocessed by removing signal noise and artefacts by adopting selective threshold denoising method of Discrete Wavelet Transform (DWT). Distinct statistical features are mined from each signal sub bands through multiscale approximation. The dimensionality of the signal features are reduced by using kernel based robustified Principal Component Analysis (PCA). A two-class Support Vector Machine (SVM) nonlinear classifier is used for classifying the ictal and interictal EEG signals with its two variants namely Polynomial Kernel and Radial Basis Function kernel. The performance of the various classification experiments are determined by computing of sensitivity (SEN) and specificity (SPE) and accuracy (ACC). The 5-fold cross validation is exercised to assess the performance of the classifier. Classification accuracy of 99.6% is obtained with the proposed model and outperforms similar benchmarking classification works reported recently.