Automated Detection of Seizure and Nonseizure EEG Signals Using Two Band Biorthogonal Wavelet Filter Banks

Abstract

The automated feature identification and classification of nonseizure and seizure electroencephalogram (EEG) is very useful for the diagnosis of epilepsy. In this chapter two band biorthogonal wavelet filter banks are used for classification of nonseizure and seizure EEG signals, and their classification accuracy has been evaluated. The energy or the bispectral phase entropies of the wavelet subbands can be used to discriminate nonseizure and seizure EEG signals. We compare the performance of energy measure and the bispectral phase entropies to discriminate EEG signals. We compare the classification accuracy of thirty biorthogonal filter banks with respect to the regularity orders of the synthesis and analysis of low pass filters and the number of wavelet decompositions. It is found that the energy measure performs better than the bispectral phase entropy for the cases for which the regularity order is greater than or equal to five independent of the wavelet decomposition level. For the fifth and sixth levels of wavelet decomposition, it is found that the energy measure always performed better than the bispectral phase entropy measure independent of the regularity of the filter bank. For the energy measure, the filter banks with higher regularity orders are found to perform better than the filter banks with lower regularity orders at almost all the decomposition levels. However, for the bispectral phase entropy measure, the filter banks with lower regularity orders are found to perform better than the filter banks with higher regularity orders for most of the decomposition levels. The highest classification accuracies obtained from the bispectral phase entropies and the energy measure is 96.4% and 98.2% respectively.