Convolutional Transformer with Domain Adversarial Learning for Multi-Channel Sleep Stage Classification

Automatic sleep staging plays an essential role in the diagnosis of sleep disorders. While significant advancements have been achieved in the field of automatic sleep staging research, three challenges persist: (1) The problem of how to effectively extract and utilize the features of multi-channel physiological signals has not been well addressed. (2) The temporal correlation of sleep signals is very strong, so how to design the network to learn the correlation between local temporal features becomes significant. (3) Due to there are many differences in physiological signals between different individuals, this difference seriously affects the generalization of model training methods. To address the aforementioned challenges, the Convolutional Transformer with Domain Adversarial Learning for Multi-Channel Sleep Stage Classification (CT-DAL) is proposed. This method enables automated extraction of spatio-temporal features from multi-channel signals and utilizes multi-head attention to learn the correlations between these diverse features. Finally, by integrating Domain Adversarial Learning into the network, the common features between different individuals can be learned to enhance model’s generalization capability. In this study, the efficacy of each module is demonstrated through ablation experiments, and the experimental results from baseline comparisons provide compelling evidence that the proposed model outperforms all baseline models.