Personalized Sleep Staging Leveraging Source-free Unsupervised Domain Adaptation

TL;DR

This work tackles the problem of sleep staging generalization to unseen individuals by introducing SF-UIDA, a source-free unsupervised individual domain adaptation framework. It treats each new subject as a distinct target domain and uses a two-stage alignment—subject-specific adaptation via Sequential Cross-View Contrasting and subject-specific personalization with teacher-student pseudo-labeling—to convert a source-trained model into a personalized one without accessing source data. The approach demonstrates state-of-the-art performance on three public sleep datasets, with rapid per-subject adaptation (~40 seconds) and clear improvements over non-personalized domain adaptation. The method preserves privacy and offers a practical, plug-and-play solution for clinical deployment in sleep clinics.

Abstract

Sleep staging is crucial for assessing sleep quality and diagnosing related disorders. Recent deep learning models for automatic sleep staging using polysomnography often suffer from poor generalization to new subjects because they are trained and tested on the same labeled datasets, overlooking individual differences. To tackle this issue, we propose a novel Source-Free Unsupervised Individual Domain Adaptation (SF-UIDA) framework. This two-step adaptation scheme allows the model to effectively adjust to new unlabeled individuals without needing source data, facilitating personalized customization in clinical settings. Our framework has been applied to three established sleep staging models and tested on three public datasets, achieving state-of-the-art performance.

Figures (6)

• Figure 1: (a) Traditional Domain Adaptation for a group of target (new) individuals and (b) Source-free Unsupervised Personalized Customization for each target (new) individual.
• Figure 2: Illustration of the two-step alignment strategy in the SF-UIDA Framework: aligning the marginal probability distribution $\mathcal{P}_\mathcal{T}(x)$ and the class conditional probability distribution $\mathcal{P}_\mathcal{T}(x\mid{y})$ for individual target domains
• Figure 3: The architecture of the proposed SCC module.
• Figure 4: Comparison with traditional domain adaptation paradigm (i.e., non-personalized domain adaptation).
• Figure 5: Average time cost per individual (seconds).