Characterisation of epileptic seizure EEG signals using wavelet transform principal component analysis and optimization methods

Abstract

The study on characterisation of epileptic Seizure EEG signals was carried out to address the need for an automatic system to allow timely medical intervention using advanced methodology. In the present work, epileptic seizures have been analysed using wavelet based methods to identify and analyse the presence or absence of seizures using EEG data. The EEG data (N=170) for the study were recorded using standard data acquisition protocol. The 10 second scalp EEG data used in this study is sampled at a rate of 500 Hz after filtering between 1 and 70 Hz. The EEG was decomposed into sub-bands using Discrete Wavelet Transform (DWT). Further, the features obtained from the significant wavelet were used for classification. 55 features were extracted out of the 5 sub bands and were subjected to PCA to reduce the dimensionality further. The PCA based features were quantitatively analysed. The PCA based features was used for classification with Back Propagation Algorithm (BPA), Radial Basis Function Neural network (RBFNN), Particle Swarm Optimisation Neural Network (PSONN) and Adaptive Neuro Fuzzy Inference System (ANFIS) classifiers to discriminate normal and seizure EEG. Results revealed that the percentage variance of Quadratic Spline wavelet (QSW) is 52.2% with a maximum contribution of above 75% of the first five Principal Component. Results show the ability of the Wavelet Transform and RBFNN method to identify epileptic seizures. It is concluded that the Quadratic Spline Wavelet is significant for extracting features for EEG signals. Energy, Minimum, Standard Deviation and Entropy are the significant features for seizure EEG. RBFNN based classification using Wavelet transform and PCA derived features pertaining to normal and seizure EEG appear to be efficient for automated epileptic seizure analysis. newline newline