SleepBand: Single-Source Domain Generalization for Sleep Staging via Physiologically Structured Spectral Modeling

2026-07-06

AI Summary

This work addresses the poor generalization of sleep staging models under the challenging setting of single-source annotation without access to domain labels or target data. To this end, the authors propose SleepBand, a novel framework that, for the first time, incorporates physiological priors into single-source domain generalization. Specifically, it employs learnable Morlet wavelet filters to explicitly model key oscillatory rhythms—such as slow waves and sleep spindles—and integrates band-structured integration with a recalibration mechanism to anchor learned representations to neurophysiological features, thereby reducing reliance on dataset-specific artifacts. Trained end-to-end, SleepBand achieves state-of-the-art performance across five public datasets in the single-source domain generalization setting, and its learned filters exhibit strong alignment with established neurophysiological knowledge.

Abstract

Generalizing sleep staging models to unseen datasets is challenging, and typical domain generalization (DG) methods often rely on multiple source domains or domain labels that are rarely available in practice. We tackle the stricter and more practical setting of single-source domain generalization: training on a single labeled source dataset, without domain labels or access to target data. We present SleepBand, a physiology-guided framework that embeds oscillatory priors via a learnable Morlet filter bank and a structured integration-and-recalibration pipeline. This anchors representations to domain-invariant sleep rhythms (e.g., slow waves, spindles), reducing reliance on dataset-specific artefacts. On five public datasets, SleepBand achieves state-of-the-art SDG performance and remains competitive under leave-one-domain-out (multi-source) DG. Analyses show that the learned filters align with canonical neurophysiology and that robustness stems from focusing on narrowband, physiologically meaningful cues. Our results suggest that principled, physiology-aware inductive biases are a promising path for robust single-domain sleep staging. Code is available at https://github.com/lzcn/sleep-band

Problem

• sleep staging
• domain generalization
• single-source
• physiological modeling
• unseen datasets

Innovation

• single-source domain generalization
• physiologically structured spectral modeling
• learnable Morlet filter bank
• sleep staging
• domain-invariant representations