Complexity Based Multilevel Signal Analysis for Epileptic Seizure Detection

Abstract

The electroencephalogram (EEG) signal is a signal produced by a complex biological system. Thus, a signal complexity analysis can be useful for analyzing the EEG signal. Many studies have shown the vast development of signal complexity analysis in the EEG. The most commonly used methods were the entropy and fractal dimension measurement. These methods were able to perform well in the epileptic EEG seizure detection system. They were suitable for time, frequency, and wavelet domain signal processing. The use of wavelet analysis, such as discrete wavelet transform (DWT) and wavelet packet decomposition (WPD), was quite famous. In many studies, the feature extraction process was performed in the DWT or WPD process sub-band signal. One of the developments of WPD was called as the multilevel wavelet packet entropy (MWPE), which produced less features than that of WPD. This study developed a new method based on WPE, which used signal complexity measurement at each level as in MWPE. The seizure detection process in this study was started with a channel selection method to reduce the processed channels. EEG signals from selected channels were then decomposed using a five-level of wavelet packet decomposition (WPD), producing 32 wavelet coefficients. The feature extraction process was performed using the entropy and fractal dimension for all 32 sub-bands, that were segmented using a ten-minute non-overlapping window. A support vector machine (SVM) was used to classify the feature set into a seizure and normal conditions. The system was evaluated using the CHBMIT EEG dataset, which was recorded from 24 patients having a total of 198 seizure events. The highest average accuracy of 91% was achieved by using multilevel wavelet higuchi fractal dimension (MWHF) analysis. This indicates that the use of fractal based measurement has a good opportunity to be implemented in epileptic seizure detection and prediction system.