Dual-Stream Cross-Differential Transformer for obstructive sleep apnea detection via biosignal fusion of ECG and SpO2

Obstructive sleep apnea (OSA) affects over one billion adults worldwide and remains severely underdiagnosed due to the limited accessibility of polysomnography(PSG). Automated OSA detection using minimal biosignals, such as electrocardiogram (ECG) and pulse oximetry saturation (SpO 2 ), offers a scalable alternative for large-scale screening, yet existing multimodal approaches often struggle to robustly capture transient and asynchronous cardiorespiratory interactions under physiological noise. In this study, we propose a lightweight Dual-Stream Cross-Differential Transformer (DSCDT), which integrates a differential attention mechanism into a dual-stream architecture to suppress common-mode noise and emphasize pathophysiologically relevant cross-modal dynamics. The model employs parallel encoders for ECG and SpO 2 to preserve modality-specific characteristics, followed by cross-differential attention to enhance informative interactions between signals. Evaluated on two public datasets (PhysioNet Apnea-ECG and PSG-Audio) under standard minute-level detection protocols, DSCDT achieves an accuracy of 98.35%, a recall of 98.81%, and an F1-score of 98.09% on the PhysioNet Apnea-ECG dataset. Furthermore, patient-level severity stratification analysis reveals an extremely strong correlation (R 2 > 0.94) between predicted and ground-truth Apnea-Hypopnea Index (AHI) values, achieving near-perfect discriminative performance (AUC ≥ 0.99) across standard clinical severity thresholds. Notably, the model maintains this performance with only 3.20 million parameters and 2.35 GFLOPs per one-minute segment, resulting in lower computational overhead compared to heavy CNN–RNN baselines. Ablation and masking experiments further demonstrate the robustness of the proposed framework to modality degradation, confirming that differential attention enables reliable OSA detection using minimal biosignals.