AdaSleep: A Robust Source-Free Domain Adaptation Framework for Unobtrusive Sleep Staging via Belt and Radar

Unobtrusive sleep staging based on respiratory signals enables long-term sleep monitoring, yet real-world deployment is often hindered by domain shifts caused by demographic heterogeneity, environmental noise, and differences in sensing modality. Privacy constraints can make source data inaccessible, motivating Source-Free Domain Adaptation (SFDA), which uses only a source-pretrained model and unlabeled target data. However, existing self-training based SFDA methods are vulnerable to confirmation bias: signal artifacts can induce erroneous pseudo-labels that are subsequently reinforced during adaptation. To address these challenges, we propose AdaSleep, a robust framework for adapting a source-pretrained model using unlabeled target data. AdaSleep introduces a consistency-based self-training strategy built on a Teacher-Student architecture. By aggregating predictions across stochastically augmented views, AdaSleep reduces sensitivity to random perturbations and produces more stable pseudo-labels. It further applies a Quality Gate to restrict model updates with reliable pseudo-labels, thereby reducing error propagation. Extensive evaluations across four real-world datasets spanning abdominal belt and radar sensing in hospital and home environments show that AdaSleep outperforms the compared SFDA baselines, with average improvements of 3.67% in accuracy and 3.90% in Macro-F1 over the source-only baseline. These results support AdaSleep as a reliable adaptation strategy for privacy-sensitive unobtrusive sleep monitoring.