A new approach for feature extraction of EEG signal using GARCH variance series

In this paper, we present a new feature extraction method for Electroencephalogram (EEG) signals classification, based on GARCH modeling of wavelet coefficients. First, the EEG signals are decomposed into the frequency sub-bands using discrete wavelet transform (DWT). GARCH model can capture the important statistical properties of wavelet coefficients. We demonstrate that Autoregressive Conditional Heteroscedastcity (GARCH) effect exists in wavelet coefficients of EEG signals. Then GARCH variance series are calculated from each sub-band of wavelet coefficients. After that, a set of statistical features are extracted from variance series to represent of variance series. We use linear discriminate analysis (LDA) for feature selection. Resultant features are classified by multilayer perceptron (MLP) with three discrete outputs: healthy volunteers, epilepsy patients during seizure-free interval and epilepsy patients during seizure. Experimental results show that by using GARCH model on the wavelet coefficients, correct classification rate (CCR) is improved.