A Deep Neural Network Based Wake-After-Sleep-Onset Time Aware Sleep Apnea Severity Estimation Scheme Using Single-Lead ECG Data

Obstructive sleep apnea (OSA) is a prevalent yet potentially severe sleep disorder. Polysomnography (PSG) is most commonly used to assess the severity of OSA. However, there have been numerous studies to find OSA patients more effectively since running a PSG test is expensive and time-consuming. The existing studies, however, raise four major concerns, such as (i) the use of inaccurate sleep time data to calculate the apnea-hypopnea index, (ii) the use of poor preprocessing techniques for real patient clinical datasets, (iii) the lack of multi-stage classification capability, and (iv) the absence of experiments on sufficiently large data sets. To address these concerns, we propose a novel OSA severity classification scheme based on single-lead electrocardiogram (ECG) data, as well as a novel deep learning model, CLNet, to perform apnea/hypopnea and sleep stage classification. By identifying apnea/hypopnea events from a patient's ECG data and computing AHI using "pure" sleep duration via CLNet, our method improves patient OSA severity degree estimation. CLNet was trained and evaluated using two different real-world datasets containing 286 OSA patient records and a total of 2,155 hours of ECG data. In our experiments, the proposed scheme outperforms existing approaches by up to 10% in total accuracy and AUC on the public PhysioNet dataset. In terms of apnea classification sensitivity, we show that the proposed CLNet model outperforms the state-of-the-art model by up to 41.8% for our clinical dataset. Our scheme can be used as a successful, high-quality pre-screening tool by more effectively prioritizing prospective OSA patients. We will be able to perform PSG on only the most severe patients, saving both time and money. Our algorithms are publicly available on GitHub.