Deep subdomain adaptation subject-specific sleep staging framework with iterative self-training

• Published in: Computer methods and programs in biomedicine Vol. 271; p. 108996
• Main Authors: Lyu, Juntong, Chen, Ziyang, Shi, Wenbin, Yeh, Chien-Hung
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.11.2025
• Subjects: Batch normalization; EEG; Self-training; Sleep staging; Unsupervised domain adaptation; Unsupervised domain adaptation; Self-training; EEG; Batch normalization; Sleep staging
• ISSN: 0169-2607; 1872-7565; 1872-7565
• DOI: 10.1016/j.cmpb.2025.108996

•A novel deep subdomain adaptation subject-specific sleep staging framework is proposed to tackle the domain shift problem in sleep classification.•A self-iterative training scheme was proposed to stabilize and improve the performance of discrepancy-based training for sleep staging.•An adaptive domain-specific batch normalization module was proposed to align the statistics of the source and target domains.•Experiments on two sleep EEG datasets show that the proposed framework is effective and reliable. Sleep staging is pivotal in assessing sleep quality and diagnosing sleep-related disorders. Although previous efforts in sleep classification have achieved considerable success, individual differences arising from factors such as age, gender, and ethnicity continue to pose significant challenges to the generalization capability of deep neural networks, compromising their performance in subject-specific sleep staging tasks. To address this challenge, we proposed a novel framework, DDAST, which leverages a discrepancy-based learning framework to effectively solve the domain shift problem inherent in the unlabeled target domain of sleep staging. First, we designed an adaptive domain-specific batch normalization to merge statistical information from the source domain (training data) into the target domain (testing data), especially for the small target data size condition. This reduces the uncertainty in estimating moments of the target domain, thereby improving the classification of target domain data. Second, we combined the self-training scheme with a discrepancy-based unsupervised learning strategy to develop a cross-subject sleep staging framework, which utilizes target domain pseudo-labels to align the fine-grained distributions of the source and target domains effectively. The proposed framework was evaluated on two datasets through cross-validation experiments, achieving an accuracy of 89.7 % and 84.3 % on the MASS-SS3 and ISRUC-S3 datasets, respectively, outperforming other baseline methods. The effectiveness of different modules of the proposed framework was verified through ablation experiments. Visualization of feature representation also reveals a better alignment between the source and target domains after applying the proposed method, which suggests the proposed framework can effectively solve the domain shift problem in subject-specific sleep staging tasks. This study presents a domain adaptation framework targeting subject-specific sleep classification, which holds promise in sleep-related disorders diagnosis as well as clinical sleep monitoring.