Automatic Sleep Stage Classification Method based on Transformer-in-Transformer

In this paper, a transformer-in-transformer-based sleep stage classification method extracting global information along both time and frequency axes is proposed. Global information along the frequency axis contains important information of the signal, such as harmonics. However, existing deep learning-based sleep staging methods have been focused on capturing global information over time. Inspired by transformer-in-transformer for music information retrieval, which is capable of capturing both time and frequency axes along dependency, we propose to adopt transformer-in-transformer architecture into sequence-to-sequence scheme of automatic sleep staging. Experimental results on SleepEDFX dataset confirm that the proposed method achieves improved performance compared to existing deep learning-based sleep staging methods.