Research map / Label-efficient and robust learning / Domain adaptation and generalization
Personalized Sleep Staging Leveraging Source-free Unsupervised Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SF-UIDA, a source-free, subject-by-subject unsupervised domain adaptation framework for sleep staging that combines sequential cross-view contrasting with confident pseudo-label fine-tuning via a teacher model.
Problem
Sleep staging models generalize poorly to new unseen subjects because individual discrepancies are ignored, and conventional domain adaptation needs a batch of target subjects and source data, which is impractical and raises privacy concerns.
Approach
Treats each target subject as its own individual domain and adapts a pretrained source sleep model per subject without source data, using a sequential cross-view contrasting task (forward and reversed sequence views predicting each other via MMD) followed by EMA-teacher pseudo-label fine-tuning on confident sequences only.
Key results
- SF-UIDA averages 76.4% ACC and 68.7% MF1 across three datasets and three backbones, versus 66.2% / 56.1% for source-only and 75.0% / 67.3% for the best source-free baseline C-SFDA.
- The SSP (pseudo-label personalization) module alone reaches 75.0% ACC / 67.7% MF1 average; SSA alone reaches 74.8% / 66.8%; combined gives the best results.
- Per-individual personalization takes about 40 seconds on average per subject.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SF-UIDA instead of C-SFDA | Sleep staging models evaluated across ISRUC, HMC, and SleepEDF · Average | ACC: 75.0 -> 76.4 | Helped |
| Use SF-UIDA instead of C-SFDA | Sleep staging models evaluated across ISRUC, HMC, and SleepEDF · Average | MF1: 67.3 -> 68.7 | Helped |
| Add subject-specific adaptation to the source-only model (ablation) | SF-UIDA · Average | ACC: 66.2 -> 74.8 | Helped |
| Add subject-specific adaptation to the source-only model (ablation) | SF-UIDA · Average | MF1: 56.1 -> 66.8 | Helped |
| Add subject-specific personalization to the source-only model (ablation) | SF-UIDA · Average | ACC: 66.2 -> 75.0 | Helped |
| Add subject-specific personalization to the source-only model (ablation) | SF-UIDA · Average | MF1: 56.1 -> 67.7 | Helped |
| Add subject-specific personalization alongside subject-specific adaptation (ablation) | SF-UIDA with subject-specific adaptation · Average | ACC: 74.8 -> 76.4 | Helped |
| Add subject-specific adaptation alongside subject-specific personalization (ablation) | SF-UIDA with subject-specific personalization · Average | MF1: 67.7 -> 68.7 | Helped |
Surprising findings
- Per-subject personalization outperforms traditional joint domain adaptation over all target subjects on all baselines and datasets, despite using only target data.
Limitations
- Evaluated only on three lightweight CNN-RNN sleep models; complex models such as UTime or SleepTransformer were not tested.
- Relies on hand-set hyperparameters (confidence threshold 0.8, N_c=15, T=17) without reported sensitivity analysis in the text.
- Evaluated only on sleep staging; generalization to other EEG tasks is left to future work.
Experimental setup
Datasets: ISRUC HMC SleepEDF-153
Benchmarks: ISRUC HMC SleepEDF-153
Metrics: ACC MF1
Baselines: Source Only CPC SimSiam AdaContrast CoTTA C-SFDA traditional (non-personalized) domain adaptation
Builds on: CPC TS-TCC SimSiam AdaContrast CoTTA
Closest papers by meaning
- Generalizable Sleep Staging via Multi-Level Domain Alignment (AAAI 2024 · similarity 0.90)
- SleepFM: Multi-modal Representation Learning for Sleep Across Brain Activity, ECG and Respiratory Signals (ICML 2024 · similarity 0.85)
- Subject-Relative Micro-Motion and Sleep Dynamics for Near-Infrared Video Sleep Staging (NeurIPS 2026 · similarity 0.84)
- Robust Sleep Staging over Incomplete Multimodal Physiological Signals via Contrastive Imagination (NeurIPS 2024 · similarity 0.84)
- SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG (ICLR 2025 · similarity 0.83)
- SleepVST: Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.