Transition-aware sparse attention for long-range single-lead ECG sleep apnea detection

Sleep apnea (SA) is a prevalent sleep disorder associated with substantial long-term health risks. Single-lead electrocardiogram (ECG)-based approaches offer a promising avenue for low-burden monitoring and large-scale screening, but many existing methods mainly rely on segment-independent predictions or limited local context, and therefore may insufficiently capture overnight temporal continuity and boundary ambiguity around apnea transitions. In this work, we develop a many-to-many framework for SA detection from RR interval (RRI) and R-wave peak amplitude (RPA) time series derived from single-lead ECG. A shared segment-level encoder extracts local representations from the RRI/RPA inputs, while a sequence-level encoder with Top-K sparse self-attention models long-range temporal dependencies across consecutive one-minute segments. We further introduce a transition-aware auxiliary task to emphasize segments around label changes and support segment-wise classification. Experiments were conducted on two public datasets: Apnea-ECG, containing 70 recordings evaluated using the official 35/35 train-test split with a validation subset from the training split, and UCDDB, containing 25 subjects evaluated using subject-independent five-fold cross-validation. The proposed method achieved segment-level accuracies of 93.1% on Apnea-ECG and 82.3% on UCDDB, with corresponding sensitivities of 91.9% and 64.9%, F1-scores of 91.1% and 61.4%, and record-level AHI correlations of 0.991 and 0.926. These results suggest that jointly modeling long-range sequential context and transition-related cues can be useful for physiological time-series analysis and ECG-based SA screening. The code is publicly available at [https://github.com/DejunZhang-Online/Sleep-apnea-detection](https://github.com/DejunZhang-Online/Sleep-apnea-detection).