Discrete Wavelet Transform and Sample Entropy-Based EEG Dimensionality Reduction for Electroencephalogram classification

Abstract

Multiple Electroencephalogram (EEG) channels are required for brain pathologies recognition, such as epilepsy. So far, analyzing all these channels leads to an over-dimensional issue, which impedes the desired performance. Consequently, a variety of studies have proposed static channel selection algorithms to characterize the most pertinent channels. However, these selected channels cannot adapt with unpredictable data within processed EEG signals. Thus, in this paper, we propose a dynamic channel selection algorithm based on discrete wavelet transform (DWT) combined With sample entropy. Firstly, we explored the complexity of CHB–MIT EEG signals through sample entropy and DWT. Brainwaves of Well–known neurological interest namely; Delta, Theta, Alpha, Beta and and Gamma, span the [0:64] Hz frequency range. We calculated the Sample Entropy (SE) of a 1–level DWT coefficients for normal and ictal EEG signals. Results show that Sample entropy values fall abruptly during seizure periods. Since, all the sample entropy values fall when seizing, the channel with minimum value in comparison of other channels is the closer one to the epilepsy source also known as epileptogenic area. Consequently, The channel at a minimum value of sample entropy was selected to be processed further with a 5-levels DWT, for pre–ictal, ictal and normal EEGs. In order to evaluate the selected channels, variance–to–mean ratio, standard deviation (STD) and Kurtosis of DWT coefficients Were used as an input feature vector for a Naive Bayes classfier. Normal and ictal EEGs were processed for epilepsy detection, and pre–ictal and normal EEGs for prediction. Classfication results confirmed the effectiveness of the developed algorithm.