4. Classification of EEG graphoelements with supervised and unsupervised learning algorithms

Abstract

The electroencephalogram (EEG) provides markers of brain disturbances in the field of epilepsy. In short duration EEG data recordings, the epileptic graphoelements may not manifest. The visual analysis of lengthy signals is a tedious task. It is necessary to track the EEG activity on the computer screen and to detect the epileptiform graphoelements and the other pathological activity. The automation of the process is needed. We will compare the EEG wave classification both by supervised and unsupervised learning algorithms. The feasibility to detect the changes in the microstructure of epileptic activity will be verified. The procedure is based on multichannel adaptive segmentation, feature extraction and classification of graphoelements. To take into account the non-stationary behavior of the signal, the features were extracted from segments detected by adaptive segmentation. The features included amplitude variance, parameters describing duration, number of segments, power in the frequency bands, signal entropy, number of zero crossings, nonlinear operator and others. The classification was performed by (fuzzy) cluster analysis and by artificial neural network using genetic algorithm. Adaptive segmentation extracts the discriminative features much better, than segmentation with fixed segment boundaries. The temporal profiles were plotted for EEG channels to reflect the macro-structure of the signal. The graphoelements of relevant class were highlighted by a color. Learning with a teacher was more successful than unsupervised algorithms.