MCASleepNet: Multimodal Channel Attention-Based Deep Neural Network for Automatic Sleep Staging

Sleep staging is significant for the capture of sleep patterns and the assessment of sleep quality. Although previous studies attempted to automatically detect sleep stages and achieved high classification performance, several challenges remain: 1) How to correctly classify the sleep stages end-to-end. 2) How to capture the representations and sleep transition rules effectively. 3) How to capture the sleep features adaptively from multimodal data for sleep staging. To address these problems, a multimodal channel attention-based sleep staging network named MCASleep-Net is proposed. Specifically, the proposed network is a mixed model composed of a dual-stream convolutional neural network (CNN) structure, a solitary long short-term memory (LSTM) module and a multimodal channel attention module. The dual-stream CNN structure is designed for the extraction of sleep representation characteristics from multimodal data. Meanwhile, the LSTM module is developed to learn transition rules between sleep stages. In addition, the multimodal channel attention module is developed to extract meaningful features of specific sleep stages from multimodal sleep data, electroencephalogram (EEG) and electrooculogram (EOG) signals. The sleep staging experiment demonstrates that the MCASleepNet proposed in this study exceeds the state-of-the-art baseline methods, with an overall accuracy of 89.1% and a macro F1 score of 84.4 on the publicly available Sleep-EDF dataset.