Automatic Sleep Staging Based on Data Augmentation with Diffusion Model and Transformer Network

The quality of sleep directly affects human health, making it essential to accurately analyze sleep stages. With the help of deep learning algorithms, automatic sleep staging improves the efficiency of diagnosis, but the precision is hindered due to the lack of electroencephalogram (EEG) signals and unavoidable noises when collecting them. In this research, we used the Diffusion Model to achieve data augmentation on EEG signals, compensating for the effects of noises. Then we applied the graph convolutional neural network (GCN) to extract features from the EEG time-frequency images. Instead of convolutional neural network (CNN) and recurrent neural networks (RNN), we implemented transformer mechanism to categorize sleep stages, which can better capture the relations between epochs of EEG signals and reflect the general reaction of brains. In the 5-stage sleep staging on public datasets Sleep-EDF-20 (±30mins),we gained a collective accuracy of 97.1%, an MF1 score of 95.5% and the kappa value was 0.96, which was competitive in relation to state-of-art techniques on the same datasets. This research may help develop a precise sleep staging mechanism and be used in actual medical treatments.