Statistical spectral feature extraction for classification of epileptic EEG signals

Discrimination of epileptic activity in the electroencephalogram (EEG) signals continuously recorded from the brain may facilitate the effective and accurate diagnosis of epilepsy. This paper proposes a new statistical method combined with a simple classification algorithm that can discriminate epileptic EEG signals from normal signals. The statistical method extracts most significant spectral features by maximizing statistical distance between the epileptic and the normal power spectrums. The power spectrum density of EEG signals is estimated by the multi-taper method. A linear algorithm based on the Fisher discriminant analysis classifies the selected spectral features as either the epileptic or the normal class from the EEG recordings. The results demonstrate that our method could reach >99.6% classification accuracy while its computational complexity appears to be much lower than the previously proposed methods that exhibited similar classification performances. It is suggested that our method may be readily implemented in real time with high accuracy so that it can provide an on-line monitoring tool for clinical epilepsy diagnosis.