Non-contact Detection of Sleep Apnea Using Multi-Timescale Cardiopulmonary Features

Sleep Apnea Syndrome (SAS) is a sleep-related respiratory disorder with a high incidence rate but low awareness, particularly among the elderly population. This study concerns with the non-contact detection of sleep apnea events (SA) by utilizing piezoelectric sensors placed under the pillow to record ballistocardiogram (BCG) and BCG-derived respiratory (BDR) signals from subjects during their nocturnal sleep. Based on the recorded vital signs, we propose to extract multi-timescale features in terms of BCG, BDR and the coordinated cardiopulmonary features, and investigate the detection performance of sleep apnea by using Dendrite Net (DD). In experimental study, 50 SAS patients with different severity were involved and simultaneously examined by Polysomnography (PSG). Numerical results demonstrate that the multi-timescale feature extraction can better characterize SA in comparison with that extracted in a fixed timescale, and can achieve sensitivity, specificity, accuracy, and F1 score of 76.49%, 82.45%, 80.96%, and 66.80%, respectively, for the detection of SA. It is expected that the non-contact sensor-based SA detection have the potential for initial screening of the severity of SAS in home settings.