DCAL: Dual-Temporal Contrastive Adversarial Learning for Cross-Subject EEG Sleep Analysis
IEEE Transactions on Automation Science and Engineering · 2026 · 23 · 13859-13873
DOI: 10.1109/tase.2026.3716751 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Automation Science and Engineering, 2026. Verfasst von Yuxin Yin und 4 weiteren Autor:innen. Die Arbeit nennt 54 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Key-Area Research and Development Program of Guangdong Province, Science and Technology Program of Guangzhou.
Abstract (Originalsprache)
Accurate automated sleep analysis is essential for sleep health assessment, primarily comprising two key tasks: sleep staging and the detection of obstructive sleep apnea-hypopnea syndrome (OSAHS). However, cross-subject generalization remains a key challenge, as inter-individual electroencephalogram (EEG) variability introduces substantial feature shifts, while differences in sleep architecture and OSAHS event frequency give rise to label shifts. This paper proposes a Dual-Temporal Contrastive Adversarial Learning network (DCAL), a unified unsupervised domain adaptation framework designed to enhance the generalizability of both sleep staging and OSAHS detection using single-channel EEG. Specifically, DCAL holistically addresses the concomitant challenges of feature and label shifts. A Dual-contrastive Adversarial Learning (DAL) module is introduced to mitigate feature distribution discrepancies by synergizing temporal contrastive learning with an adversarial training paradigm, thereby learning robust, domain-invariant representations. Concurrently, a Learnable Re-Weighting (LRW) module is introduced to rectify the label distribution bias often induced by pseudo-labeling, proposing a novel mechanism that incorporates information gain to achieve cross-domain label consistency. Extensive experiments on a public dataset (UCDDB) and our newly collected OSAHS-PSG dataset validate the superiority of our proposed DCAL, demonstrating consistent performance improvements of 1%–7% across key metrics compared to state-of-the-art methods. Moreover, DCAL remains effective under incomplete or imbalanced data, highlighting its practical utility for real-world sleep monitoring and clinical diagnostics. Note to Practitioners—The motivation for this work stems from the operational challenge of scaling automated sleep monitoring systems to broad and diverse patient populations. A critical bottleneck in current deployments is that significant discrepancies in both feature and label distributions across individuals severely hinder the generalization of pre-trained models to new subjects. Consequently, standard algorithms often require costly re-calibration or manual labeling when applied to new users. To address this, we present a domain adaptation framework tailored for single-channel EEG that allows the system to self-adapt to a new user’s physiological patterns without human supervision. By mathematically harmonizing the signal characteristics and event probabilities between the training population and the target user, our method enhances diagnostic accuracy for both sleep staging and apnea detection. From an engineering perspective, this solution offers a pathway to reduce reliance on extensive annotated datasets for every new deployment scenario while maintaining high reliability on minimal hardware configurations. Although the current framework demonstrates robustness on clinical datasets, future integration into wearable edge devices will necessitate further optimization for computational efficiency and validation against the signal artifacts typical of uncontrolled, non-clinical environments.
Förderung
- National Natural Science Foundation of China
- Key-Area Research and Development Program of Guangdong Province
- Science and Technology Program of Guangzhou
- GJYC Program of Guangzhou
Zitieren
APA
Yin, Y., Chen, B., Xu, Z., Chen, C. L. P., & Zhang, T. (2026). DCAL: Dual-Temporal Contrastive Adversarial Learning for Cross-Subject EEG Sleep Analysis. IEEE Transactions on Automation Science and Engineering, 23, 13859-13873. https://doi.org/10.1109/tase.2026.3716751
Vancouver
Yin Y, Chen B, Xu Z, Chen CLP, Zhang T. DCAL: Dual-Temporal Contrastive Adversarial Learning for Cross-Subject EEG Sleep Analysis. IEEE Transactions on Automation Science and Engineering. 2026;23:13859-13873. doi:10.1109/tase.2026.3716751
BibTeX
@article{Yin2026,
author = {Yin, Yuxin and Chen, Bianna and Xu, Zihua and Chen, C. L. Philip and Zhang, Tong},
title = {DCAL: Dual-Temporal Contrastive Adversarial Learning for Cross-Subject EEG Sleep Analysis},
journal = {IEEE Transactions on Automation Science and Engineering},
year = {2026},
volume = {23},
pages = {13859--13873},
doi = {10.1109/tase.2026.3716751}
}
EUVIMED