An automatic procedure for the online recognition of REM sleep appears to be a necessary tool for selective REM sleep deprivation in depressive patients. To develop such a procedure we applied an artificial neural network to preprocessed single-channel EEG activity. EOG and EMG information was purposely not provided as input to the network. A generalized back-propagation algorithm was used for computer simulation. The sleep profile scored manually according to Rechtschaffen and Kales served as the desired output during the training period and as standard for the judgement of the network output during working mode. Polysomnographic recordings from 5 healthy subjects were pooled to train the network, whereas second-night EEG recordings from the same subjects were used as independent working data sets. We further applied the network to the data of 5 depressive patients without medication and 6 depressive patients treated with amitriptyline. For these groups between 84.9 and 88.6% out of all time periods consisting of 20 s of continuous EEG activity were correctly classified. The indicator function of REM sleep was well approximated by the network output in the course of the night. Especially the REM onset was excellently recognized. The inclusion of patient data in the training set yielded a different network, which was evaluated and compared.