Brain Electrographic State Detection Using Combined Unsupervised and Supervised Neural Networks

Abstract

This work describes the development of a new approach to NN processing of sleep-related brain electrographic signals, using two sequentially combined unsupervised (Kohonen layer, KNN) and supervised (Widrow-Hoff layer, WHNN) algorithms. Twelve parameters extracted from physiological data (EEG, EMG and EOG, obtained from unrestrained rats through several sleep-waking periods), were first processed by a KNN, that detected different signal patterns. These patterns were further examined by an EEG expert, who identified them as belonging to one of the known sleep-waking stages, or to transitional and/or unknown signal combinations. Selected outputs of the KNN, classified in this way, formed the input vectors to a WHNN, that allowed fast and reliable tracking of changes in these states (both known and newly detected) during prolonged periods of time. Such an approach can represent an important aid for simultaneous exploration, detection and also temporal following of electrographic events along the sleep-waking cycle.