WaveCrossNet: interpretable cross-modal automatic sleep staging for OSA patients

Obstructive sleep apnea (OSA) patients show highly fragmented sleep with frequent stage transitions, challenging existing automatic sleep staging in feature extraction, multimodal correlation modeling, and interpretability. Current methods rely on fixed-threshold denoising and standard positional encoding, suffering from obvious module stacking and a lack of design tailored to OSA's non-stationary signals. To address these issues, this paper proposes WaveCrossNet, an enhanced Transformer-based model for OSA sleep staging. Unlike simple module stacking, WaveCrossNet integrates adaptive wavelet thresholding and a dual-channel CNN into the Transformer's temporal modeling framework in an end-to-end, learnable manner. First, an energy-adaptive subband thresholding mechanism selectively denoises while preserving transient events such as micro-arousals. Second, a position-aware attention with dynamic cross-modal fusion explicitly models the complementary roles of EEG and EOG across sleep stages (e.g., N1, REM). Third, a hard example mining loss alleviates class imbalance, where minority stages like N1 account for less than 5% of OSA data. Attention weight visualization links feature correlations to physiological events (e.g., slow eye movements in N1, high-amplitude slow waves in N3), enhancing interpretability. Evaluated on the Sleep-EDFx healthy dataset and 200 SHHS OSA records with subject-independent splits, WaveCrossNet achieves 84.9% accuracy on the OSA dataset and outperforms state-of-the-art methods on healthy data. By jointly optimizing denoising, representation learning, and temporal modeling, WaveCrossNet provides an interpretable, end-to-end solution for accurate OSA sleep staging.