Bimodal Gaussian PDF of the Dominant IMFs of EEG Signals for Seizure Activity Classification

Abstract

In this paper, time-frequency domain operation exploiting Empirical Mode Decomposition (EMD) is performed on EEG signals to classify normal activity of healthy person, non-seizure (inter-ictal) and seizure (ictal) activity of seizure patients. Analyzing power spectral density estimation in the set of Intrinsic Mode Functions (IMFs) by periodogram, it is seen rational to specify some prominent IMFs to use in classification problem. Bimodal distribution which is two components Gaussian statistical model is utilized in specified leading IMFs to summarize information in it and propose feature set using the modeling parameters of two component Gaussian probability density function (PDF). Empirical and Gaussian PDF have been plotted along with Kolmogorov-Smirnov (K-S test) goodness-of-fit hypothesis test and found Gaussian statistical model most effective to feed modeling parameters of Gaussian PDF in SVM and ANN classifier for normal, inter-ictal and ictal stages of seizure classification. The proposed technique is accomplished with making higher sensitivity, specificity and accuracy compared to that made by an up-to-date method while wide range of simulations were performed using the same benchmark EEG dataset and classifier.