A Fuzzy Statistical Perspective for Empirical Evaluation of EEG Classification Models for Epileptic Seizures

Electroencephalography (EEG) signals are a combination of complex pattern sequences, which are periodic in nature. These pattern sequences include a gamma waves that indicates deep thinking behaviour, a beta wave sequence that indicates busy and active mind status, an alpha wave segment which indicates reflective and restful behaviour, a theta wave which is an indicative of drowsiness, and a delta wave which indicates sleeping & dreaming conditions. Features like frequency changes, amplitude changes, pattern changes, etc. are used to identify chronic, ischemic and other diseases related to the brain. In order to classify these wave patterns into brain diseases like epilepsy, a series of high complexity signal processing operations are needed to be executed in tandem. These operations include signal pre-processing, feature extraction, feature selection, classification into epileptic & non-epileptic seizure and post-processing. A large variety of algorithms are developed by researchers for each of these operations. Performance of these algorithms varies largely w.r.t. the number of leads used for EEG capture, filtering efficiency, feature extraction & selection efficiency, and classifier efficiency. Thus, it becomes ambiguous for researchers and system designers to select the best possible algorithm set for their application. In order to reduce the ambiguity, this text provides a comprehensive comparison of a wide variety of epileptic & non-epileptic seizure classification system models. These models are statistically compared on the basis of overall accuracy, delay of decision making, precision, recall, f-measure and field of application. It is observed that convolutional neural network (CNN) based models outperform other models in terms of general-purpose performance, while specialized CNN models must be used for application specific deployments.