Detecting epileptic seizure activity in the EEG by independent component analysis

Manually reviewing EEG (Electroencephalogram) recordings, for detection of electrographical patterns, is a time consuming business. Therefore, the ability to automate the classification of interesting electrographical patterns is a good supplement to the wide range of detection algorithms currently used for EEG analysis. This paper presents the development of an algorithm for the detection and classification of epileptic activity in the EEG using Independent Component Analysis (ICA). Detection and classification of epileptic activity is achieved by an algorithm that searches for paroxysmal activity in the EEG. The number of underlying components and their activity is combined in one detection measure (absolute sum). When paroxysmal activity is detected, the estimated components are grouped in physiologically relevant clusters. These clusters are used to reconstruct a clean EEG signal, which can be evaluated again by the detection measure. The algorithm was designed using artificial EEG data constructed from known ICs. From evaluations it is seen that the absolute sum is a good measure to detect paroxysmal activity, resulting in a 74 % sensitivity and a 19 % selectivity. It is likely that these values will improve substantially if the estimated mixing matrix is used for the rejection and classification of artifacts. From this work it is seen that this ICA based algorithm has some great possibilities to detect and discriminate epileptic activity from several kinds of artifacts. This project gives a promising base for the development of a system that automates the classification of interesting electrographical patterns using ICA decompositions.