Comparative Study of Deep Learning Based Sleep Scoring Methods

The paper is devoted to the analysis of multichannel biomedical signals acquired in the sleep laboratory. The whole data set includes 29 overnight records of electroencephalogram (EEG), electrooculogram (EOG), electrocardiogram (ECG), sound and movement data observed with a sampling frequency of 200 Hz, and breathing records (Flow) acquired with a sampling frequency of 10 Hz, among others. The methodology used for their processing includes their filtering, feature extraction and classification using both standard and deep learning classification methods as well. The goal of the paper is (i) in presentation of the deep learning machine learning in the frequency domain without any specification of features, (ii) in comparison with results obtained by the classical approach based on the two layer neural network model and initial specification of signal features, and (iii) in comparison of separation of the Wake sleep stage from the REM, NonRem1, NonREM2, and NonREM3 sleep stages using the deep learning method. The best separation accuracy of 92.12 % (with the loss value 0.19) was achieved for the separation of the Wake and NonREM3 stages for a single EEG channel and data segments 30s long. Results suggest that the deep learning strategy can help with sleep stages classification in the clinical environment.