Enhanced OSA Detection via Multi-Scale Modeling of EEG and ECG Signals.

Obstructive Sleep Apnea (OSA) is a prevalent sleep disorder that has been linked to various serious health conditions, including cardiovascular diseases, metabolic disorders, and cognitive impairments. Early detection of OSA is crucial for preventing these associated health risks. Although significant progress has been made in the automatic detection of OSA, existing methods predominantly rely on single signals, such as electrocardiogram (ECG), blood oxygen saturation, Airflow, and mainly use one-dimensional networks for feature extraction, which fail to fully leverage the time-frequency characteristics of the signals. To address these limitations, this study proposes TASPPNet, a novel approach that combines time-frequency representations of electroencephalogram and ECG signals. By utilizing multimodal feature fusion and Atrous Spatial Pyramid Pooling modules for multi-scale contextual information extraction, TASPPNet significantly enhances detection performance. Experiments conducted on the UCDDB and the subset of SHHS2 datasets demonstrated that TASPPNet achieved an accuracy of 0.783 and F1 of 0.799 on UCDDB and an accuracy of 0.866 and F1 of 0.717 on SHHS2-200. These results showcase the ability of TASPPNet to deliver competitive performance even with fewer training and testing samples, providing an innovative solution for accurate OSA detection.