Clustering Epileptiform Discharges in the Interictal Electroencephalogram with Topography Preserving Maps

Abstract

Topography preserving maps have proved useful in the clustering of paroxysmal events in the Electroencephalogram (EEG) - in particular epileptiform events (EEvs) and artefacts. With the aim of enhancing performance of pre-existing systems, a novel variant of Kohonen’s Self Organising Feature Map (SOFM) is considered. Realistic, synthetic EEvs have been generated using a 3-sphere head model, superimposed on true EEG. Pre-processing by means of Principal Component Analysis has allowed dimensionality reduction of the synthetic, interictal 25 channel EEG. This was clustered employing an Adaptive Subspace variant of the SOFM. The resulting clusters were interpreted to allow classification. This has permitted the development of a scheme to automatically detect and extract features from EEG traces, which offer results comparable with those in the literature over the synthetic data.