Research map / Label-efficient and robust learning / Domain adaptation and generalization
Generalizable Sleep Staging via Multi-Level Domain Alignment
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.
Introduces generalizable sleep staging as a multi-source domain generalization task and proposes SleepDG, which aligns epoch-level (mean and covariance) and sequence-level (inter-epoch Pearson correlation) features across five public PSG datasets.
Problem
Automatic sleep staging models trained and tested within one dataset generalize poorly to unseen sleep datasets with different populations, channels, and equipment.
Approach
Treat each sleep dataset as a domain and learn domain-invariant features with two alignment losses: epoch-level alignment of first- and second-order feature statistics, and sequence-level alignment of the expected inter-epoch Pearson correlation matrices, combined with an autoencoder reconstruction loss in a CNN-Transformer encoder.
Key results
- SleepDG reaches 75.03% average ACC and 69.64% average MF1 across five leave-one-dataset-out targets, versus 67.48% / 61.93% for DeepSleepNet and 68.49% / 62.60% for TinySleepNet... (per-method averages as reported; SalientSleepNet 68.91% ACC).
- Among DG baselines, CORAL is strongest at 72.27% ACC and 67.52% MF1; BASE reaches 68.76% ACC and 63.34% MF1.
- Ablation: epoch-level alignment alone gives +3.57% ACC over BASE; sequence-level alignment alone gives +2.37%; full model gives +6.27% ACC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add autoencoder (ablation) | BASE · Unseen datasets (Avg) | ACC: +0.97 | Helped |
| add autoencoder (ablation) | BASE · Unseen datasets (Avg) | MF1: +1.09 | Helped |
| add epoch-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | ACC: +3.57 | Helped |
| add epoch-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | MF1: +4.07 | Helped |
| add sequence-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | ACC: +2.37 | Helped |
| add sequence-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | MF1: +3.16 | Helped |
| add autoencoder and epoch-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | ACC: +4.42 | Helped |
| add autoencoder and epoch-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | MF1: +4.71 | Helped |
| add autoencoder and sequence-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | ACC: +3.10 | Helped |
| add autoencoder and sequence-level feature alignment (ablation) | BASE · Unseen datasets (Avg) | MF1: +3.85 | Helped |
| add epoch-level and sequence-level feature alignment with autoencoder (ablation) | BASE · Unseen datasets (Avg) | ACC: +6.27 | Helped |
| add epoch-level and sequence-level feature alignment with autoencoder (ablation) | BASE · Unseen datasets (Avg) | MF1: +6.30 | Helped |
Surprising findings
- DANN gives the smallest gain over BASE among the DG methods (71.12% vs 68.76% ACC), consistent with prior observations that adversarial training helps DA more than DG.
- Architectural complexity alone does not yield reliable cross-dataset gains: several non-DG models such as U-Time, AttnSleep, and EnhancingCE barely improve over DeepSleepNet.
Limitations
- Uses only single-channel EEG and EOG, with a fixed set of five datasets and only 150 recordings from SHHS and P2018.
- Sequence-level alignment relies on a heuristic (Pearson correlation expectation) to approximate the conditional distribution, and the prior assumption that P(Y|H) is identical across datasets is not verified.
Experimental setup
Datasets: SleepEDFx HMC ISRUC SHHS P2018
Benchmarks: SleepEDFx HMC ISRUC SHHS P2018
Metrics: accuracy macro-F1
Baselines: DeepSleepNet U-Time AttnSleep ResnetMHA TinySleepNet EnhancingCE SalientSleepNet BASE MMD CORAL IRM DANN
Builds on: DeepDG CORAL MMD DANN IRM SalientSleepNet
Closest papers by meaning
- Personalized Sleep Staging Leveraging Source-free Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.90)
- SleepFM: Multi-modal Representation Learning for Sleep Across Brain Activity, ECG and Respiratory Signals (ICML 2024 · similarity 0.85)
- Robust Sleep Staging over Incomplete Multimodal Physiological Signals via Contrastive Imagination (NeurIPS 2024 · similarity 0.85)
- Subject-Relative Micro-Motion and Sleep Dynamics for Near-Infrared Video Sleep Staging (NeurIPS 2026 · similarity 0.83)
- SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG (ICLR 2025 · similarity 0.81)
- SleepVST: Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers (CVPR 2024 · similarity 0.81)
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.