Transusleepnet: A Scalable Multi-Modal Learning Architecture for Sleep Scoring Based on Transformer and U-Net

Abstract

Sleep staging research belongs to the essential category of brain neuroscience, which provides early data evaluation and technical help for the diagnosis of numerous disorders. In the past 10 years, the application of deep learning has set off waves in different fields, neural networks have also become increasingly favored tools for analyzing physiological time series data in this context. Although the previous methods using convolutional neural networks and recurrent neural networks have achieved high performance in sleep staging attempts, they still face several problems: 1) How to effectively capture the deep significant differences in different sleep stages in EEG data; 2) How can deep learning be used to improve the classification model's capacity to capture fine spatial features. To address the aforementioned concerns, we devised TransUSleepNet, which is a sleep staging network that not only uses multimodal signals as input but can also be extended to a single EEG channel as input to achieve great results. The model is a time series full convolution network based on the U-Net architecture, which combines multi-scale convolution operations and attention mechanisms to expand the receptive field. Its distinguishing advantage is the ability to lock the key elements of network architecture learning on the description of multi-scale, omni channel, and detailed spatial features in the sleep process. Finally, we analyze and assess the overall performance of our model structure on standard open-source datasets, and the outcomes indicate the model's commendable efficacy in sleep staging, yielding satisfactory performance levels.

Note:
Funding Information: This study was supported by the Jilin Province Science and Technology Development Plan project of China under Grant N0.20240304052SF.

Conflict of Interests: Authors declare that there is no conflict of interest in this paper.

Keywords: Sleep Stage Classification, u-net, encoder-decoder, Transformer, multimodal signals, single-channel EEG