Application of the biologically inspired network for electroencephalogram analysis

Abstract

Architecture of a neural network combining automatic feature extraction with the minimized amount of network training acquired by means of employing of a multistage training procedure is investigated. The network selects prototypical signals and calculates features based on the similarity of a signal to prototypes. The similarity is measured by the prognosis error of the linear regression model. The network is applied for the meaningful paroxysmal activity vs. background classification task and provides better accuracy than the methods using manually selected features. Performance of several modifications of the new architecture is being evaluated. © Springer-Verlag 2001.