A novel robust diagnostic model to detect seizures in electroencephalography

Abstract

Identifying seizure patterns in complex electroencephalography (EEG) through visual inspection is often challenging, time-consuming and prone to errors. These problems have motivated the development of various automated seizure detection systems that can aid neurophysiologists in accurate diagnosis of epilepsy. The present study is focused on the development of a robust automated system for classification against low levels of supervised training. EEG data from two different repositories are considered for analysis and validation of the proposed system. The signals are decomposed into time-frequency sub-bands till sixth level using dual-tree complex wavelet transform (DTCWT). All details and last approximation coefficients are used to calculate features viz. energy, standard deviation, root-mean-square, Shannon entropy, mean values and maximum peaks. These feature sets are passed through a general regression neural network (GRNN) for classification with K-fold cross-validation scheme under varying train-to-test ratios. The current model yields ceiling level classification performance (accuracy, sensitivity & specificity) in all combinations of datasets (ictal vs non-ictal) in less than 0.028s. The proposed scheme will not only maximize hit-rate and correct rejection rate but also will aid neurophysiologists in the fast and accurate diagnosis of seizure onset.