HAMCE-Net: Heterogeneity-Aware Multimodal Cross-Epoch Network for Sleep Apnea Detection

Sleep Apnea Syndrome (SAS) is clinically diagnosed from overnight polysomnography (PSG), but manual scoring of apnea and hypopnea is labor-intensive and variable across scorers. Deep learning–based methods have improved automation, yet often underexploit heterogeneous physiological responses across PSG channels and neglect the temporal asynchrony between airflow reduction and oxygen desaturation. We propose Heterogeneity-Aware Multimodal Cross-Epoch Network for sleep apnea(SA) detection (HAMCE-Net), a cross-epoch framework for epoch-level classification of normal breathing, hypopnea, and apnea from heterogeneous PSG channels. HAMCE-Net processes two-minute windows comprising four consecutive 30 s epochs of airflow (Flow), thoracoabdominal effort (Abdo), and blood oxygen saturation (SpO2) signals, and consists of three main modules. A Periodic Respiratory Signal Encoder (Periodic Encoder) performs modality-specific, within-recording amplitude variation-aware feature extraction of Flow and Abdo with channel-specific convolutions, amplitude-aware channel interaction, and attention-guided segment aggregation to obtain physiologically meaningful epoch-level respiratory representations. In parallel, an SpO2Patch-Based Temporal Encoder (SpO2Encoder) uses a patch-based temporal Transformer to model multi-epoch desaturation dynamics and event-wise variability in the delay between airflow-limiting respiratory events and subsequent oxygen desaturation. A Multimodal Fusion and Cross-Epoch Decision Network (MFCED) then fuses respiratory and SpO2embeddings and applies bidirectional cross-epoch sequence modeling to produce per-epoch three-class predictions, explicitly accommodating cross-modal response delays and encouraging temporally coherent outputs. On the Multi-Ethnic Study of Atherosclerosis (MESA) PSG cohort, HAMCE-Net achieves strong discriminative performance for segment-level SA detection, while ablation analyses demonstrate that the Periodic Encoder, the SpO2Encoder, and the MFCED module are each well adapted to their target modalities and contribute reliably to the overall model performance.