Abstract
Sleep apnea-hypopnea syndrome (SAHS) is a common sleep-related breathing disorder associated with substantial cardiovascular and neurocognitive risks. Although polysomnography (PSG) remains the clinical gold standard for diagnosis, its cost, operational burden, and limited accessibility hinder scalable and longitudinal home monitoring. Frequency-modulated continuous-wave (FMCW) radar provides unobtrusive, non-contact respiration sensing, yet radar-based event detection is often constrained by scarce annotations and pronounced domain shifts relative to PSG signals. In this work, we propose a deep learning framework for apnea-hypopnea event detection from FMCW radar that combines a 1D U-Net segmentation backbone with multi-head self-attention (MHSA) and cross-modality transfer learning. The model was first pre-trained on a large public PSG dataset to learn transferable respiratory-event representations and then fine-tuned on a smaller clinically annotated radar respiration dataset using synchronized PSG labels. It produced per-sample event probabilities, which were further refined via temporal post-processing to generate event-level detections and apnea-hypopnea index (AHI) estimates. Experimental results demonstrate strong performance in the radar domain, achieving precision of 0.8137±0.0332, recall of 0.8369±0.0470, and an F1-score of 0.8167±0.0052. Overall, these results indicate that PSG-to-radar transfer learning enables accurate, low-cost, and non-contact sleep apnea screening, supporting scalable longitudinal monitoring in home-like settings.