Automatic recognition of rapid eye movement (REM) sleep by artificial neural networks

Abstract

Artificial neural networks are well known for their good performance in pattern recognition. Their suitability for detecting REM sleep periods on the basis of preprocessed EEG data in humans under clinical conditions was tested and their performance compared with the manual evaluation. A single channel of the EEG signal was analysed in time periods of 20 s and preprocessed into a vector of six real numbers, which served as input to the network. EOG and EMG information was ignored. Backpropagation was used as a learning rule for the network, which consisted of 12 neurons and 39 synapses. Training datasets were put together from the input vectors and the corresponding sleep stages were scored manually. In working mode different networks were compared in terms of the rate of misclassified time periods for data not belonging to the training sets. The indicator function of REM sleep was well approximated by the network output in the course of the night, which was especially true for REM onsets. The average rate of correctly classified time periods was 89%. The errors were analysed and suggestions for improvements developed.