Deep Learning With Convolutional Neural Networks for Sleep Arousal Detection

Sleep arousal influences the quality of sleep and causes health problems. Polysomnography (PSG), a group of biological signals, is often used to diagnose sleep arousal. But it is costly for sleep experts to identify sleep arousal via PSG. Thus, we designed an automatic algorithm to analyze PSG for sleep arousal identification. According to the nature of sleep arousal, we selected electroencephalogram (EEG) from PSG for further analysis. To extract frequency domain information, Welch algorithm was applied to EEG to obtain power spectral density (PSD). And then PSD was fed into a 34-layer convolutional neural network (CNN) for further feature extraction and classification. Shortcut connections were employed across every two convolutional layer to speed up the training process and realize identity mapping. Our model was trained on 900 subjects' PSGs and validated on 94 subjects' PSGs. And 989 subjects' PSGs were used as test set. Our method achieved an AUPRC of 0.10 on the full test set. There is some room for improvement comparing with other methods.