Study of micro-arousals using neural network analysis of the EEG

This paper describes the use of neural networks to analyze the electroencephalogram (EEG) from patients with recurrent micro-arousal episodes. A bank of bandpass filters and AR modelling have been used separately to represent EEG data from 7 chronic patients, divided into three different classes, each class correspondly to a different a sleep stage. A radial basis function (RBF) network has been trained to classify normal sleep EEG data from a 2-class set and from a 3-class set for automatic micro-arousal scoring of test data. Wakefulness and deep-sleep form the 2-class set, while light/dreaming sleep is included in the 3-class set. The results are compared with the visual scores assigned by an expert. The high percentage (88%-100%) of matches between the automatic and the visual scores demonstrates the ability of neural networks to recognize large and well-defined micro-arousals. The 2-class RBF with an AR model of the EEG as inputs shows a better performance in terms of sensitivity, but the 2-class RBF with a bank of filters as input data follows more accurately the micro-arousal transitions. Detection of less-obvious micro-arousals and correlation with day-time performance are among the objectives of the further work.