A Lightweight CNN–Transformer Model for Obstructive Sleep ApneaDetection Using Single-Channel SpO2 Signals

Introduction/Objective: Obstructive Sleep Apnea (OSA) is a common sleep disorder with significant health risks, yet current diagnostic methods, such as polysomnography (PSG) are complex and costly. This study aims to develop an accurate and computationally efficient OSA detection framework using single-channel SpO2 signals. Methods: We propose a two-tier framework is proposed, integrating a CNN–Transformer baseline with structured lightweight optimization. Multi-scale CNN modules extract local temporal features, while a Transformer captures long-range dependencies across respiratory events. Efficiency is further improved through dynamic weight pruning and an L1 sparsity constraint. Results: Experiments on the UCD database demonstrate that the baseline model achieves an Accuracy of 95.77%, Sensitivity of 95.02%, and Specificity of 96.18%. After lightweight optimization, model parameters were decreased by 87.1% and computational complexity was decreased by 81.5%, while maintaining model accuracy of 88.34%. Discussion: The results suggest that the proposed model achieves a practical trade-off between detection performance and computational efficiency. Using single-channel SpO₂ signals simplifies data acquisition while retaining acceptable detection capability. Although lightweight optimization results in a modest performance reduction, it significantly reduces model complexity, thereby improving its feasibility for deployment in resource-constrained environments such as wearable devices and home-based monitoring. Conclusion: The proposed framework enables accurate and efficient OSA detection using single- channel SpO₂ signals. Its lightweight and modular design facilitates potential deployment in clinical and resource-limited environments and shows promise for scalable OSA screening.