Epileptic-seizure onset detection using PARAFAC model with cross-wavelet transformation on multi-channel EEG

Finding components from multi-channel EEG signal for localizing and detection of onset of seizure, is a new approach in biomedical signal analysis. Tensor-based approaches are utilized to fit the components into multi-dimensional arrays in recent works. We initially decompose EEG signals into Beta band using discrete wavelet transform (DWT). We compare patient templates with normal template for cross-wavelet analysis to obtain Wavelet cross spectrum (WCS) and Wavelet cross coherence coefficients. Next we apply parallel factorization (PARAFAC) modeling, a three-way tensor-based representation in channel, frequency and time-points dimensions on features. Finally, we utilize the ensemble classifier for detecting seizure-free, onset and seizure classes. The clinical dataset for this work comprises of 5 normal subjects and 6 epileptiform patients. The classification performances of WCS features on PARAFAC model for Seizure detection using Ensemble Bagged-Trees classifier obtains 82.21% accuracy, while for Wavelet Coherence features, it provides higher 84.76% accuracy. The results have been compared with well-known Fine Gaussian SVM, Weighted KNN and Ensemble Subspace KNN classifiers. The aim is to analyze data over three dimensions namely, time, frequency and space (channels). This EEG based analysis is significant and effective as an automatic method for detection of seizure before its actual manifestation.