Multivariate Sleep Stage Classification using Hybrid Self-Attentive Deep Learning Networks

Recently, significant efforts have been made to explore comprehensive sleep monitoring to prevent sleep-related disorders. Multivariate sleep stage classification has garnered great interest among researchers in health informatics. In this paper, we propose HybridAtt, a unified hybrid self-attentive deep learning network, to classify sleep stages from multivariate polysomnography (PSG) records. HybridAtt is an end-to-end model that explicitly captures the complex correlations among biomedical channels and the dynamic relationships over time. By constructing a new multi-view convolutional representation module, HybridAtt is able to extract hidden features from both channel-specific and global views of the heterogeneous PSG inputs. In order to enhance feature representation, a new fusion-based attention mechanism is also proposed to integrate the complementary information carried by each feature view. To evaluate the performance of our model, we carry out experiments on a benchmark PSG dataset. Experimental results show that the proposed HybridAtt model achieves better performance compared to ten baseline methods, demonstrating the effectiveness of HybridAtt in the task of sleep stage classification.