Mixed Multi-Branch Sequential Fusion Model Based on Graph Convolutional Network and Transformer for Automatic Sleep Staging

Although deep learning methods have been improved a lot since it was introduced, there is not a method to summarize and conclude the relationship between PSG records. Comprehensive correlations are one of the most significant features people use to classify different kinds of sleeping stage. Electrocardiogram (EEG), Electromyography (EMG), and Electrocardiogram (ECG) can reflect correlated physiological conditions. As a result, it is important that we need a specific mechanism to extract and make use of these correlations to further the understanding and analysis of sleeping stage. In this paper, a new framework for sleeping stage is proposed, which consists of two parts. Time domain signals are processed by graph convolutional networks (GCN). Frequency domain signals are processed based on transformers. The experimental results show the excellent performance compared to mainstream methods.