A New Feature Definition for Classification of Multi-channel EEG Signals and its Application to Epileptic Seizure Detection

In this study, a new feature definition for classifying multi-channel EEG waveforms is introduced. New definition relies upon using the second-order statistics, e.g. autocorrelation, as a random process that represents temporal behavior of EEG signals. A spatially invariant representation of multi-channel time-series associated to EEG waveform components with class labels is obtained based on respective statistics. As an application of proposed feature vector description, a simple multivariate Gaussian classifier is designed to identify normal and epileptic EEG waveforms. Experiments with a publicly available dataset indicate that the proposed method with randomly selected lag vectors of random length within chosen ranges yields high classification success in statistical terms.