Automatic Sleep Staging Method Based on Multi-Channel Fusion

Sleep staging is the division and recognition of different stages in the sleep process, which is of great significance for sleep assessment and disease diagnosis. To make manual sleep staging less cumbersome and expensive, many automated schemes have been developed. In this paper, we propose a new approach to sleep staging using the Transformer for the main task, in order to improve the performance of automatic staging. The model converts the original EEG signal, takes the timefrequency images of multiple channels as the model input, and uses the encoder for single-channel feature extraction. The features of each channel are mapped to the feature fusion module (SE, Squeeze and Excitation), and the multi-head attention mechanism replaces LSTM for mining the fusion features. Thus, the prediction performance is improved. The prediction results on the public dataset SleepEDF-20 showed that the average accuracy was achieved 87.3%.