Automatic Sleep Staging Based on Deep Neural Network Using Single Channel EEG

Abstract

Sleep staging is the first step for sleep research and sleep disorder diagnosis. The present study proposes an automatic sleep staging model, named ResSleepNet, using raw single-channel EEG signals. Most of the existing studies utilize hand-engineered features to identify sleep stages. These methods may ignore some important features of the signals, and then influence the effect of sleep stage classification. Instead of hand-engineering features, we combine feature extraction and classification into an algorithm based on residual network and bidirectional long short-term memory network. In the proposed method, we develop a 22-layer deep network to automatically learn features from the raw single-channel EEG and classify sleep stages. Residual network can learn time-invariant features, and bidirectional long short-term memory can add learned transition rules among sleep stages to the network. The model ResSleepNet is tested on the Sleep-EDF database. We perform 10 experiments and get average overall accuracy of 90.82% and 91.75% for 6-state and 5-state classification of sleep stages. Experimental results show the performance of our model is better than the state-of-the-art sleep staging methods, and it yields high detection accuracy for identifying sleep stage S1 and REM. In addition, our model is also suitable for extracting features from other signals (EOG, EMG) for sleep stage classification.