Semi-supervised learning with dynamic threshold for sleep staging and its application on DoC patients

Sleep staging is crucial for evaluating sleep quality and diagnosing sleep-related health issues. Current sleep staging methods primarily rely on supervised learning. However, due to inherent individual differences among subjects, the performance of these established models may encounter challenges when applied to the new subject. To address this issue, a Semi-Supervised Dynamic Threshold sleep staging method (SSDT) is proposed, which modulates its learning strategy in accordance with the progression of class learning. This mutual feedback between the learning state and the learning strategy ultimately enhances the performance of the target model. The effectiveness of SSDT was validated on the Sleep- EDF dataset. Specifically, the accuracy for the Fpz-Cz channel reached 87.49%, and the Macro-average F1 score (MF1) score was 77.02%. For the Pz-Oz channel, the accuracy was 84.55%, and the MF1 score achieved 72.50%. Further analysis shows that adaptively adjusting the number of samples in the minority class can significantly boost performance, leading to an 8.86% increase in accuracy for the minority class N1 in the Fpz-Cz channel. Additionally, we explored its application in patients with disorders of consciousness (DoC). Our findings indicate that SSDT can enhance sleep staging performance in DoC patients, with a 26.76% average improvement in sleep staging accuracy. The research presents a novel and effective semi-supervised learning approach to sleep staging.