ECG-Based Automated Detection of Sleep Apnea Using Deep Neural Networks and Hidden Markov Models

Sleep disorders constitute a substantial global health burden, with more than eighty distinct conditions currently recognized. Among these, obstructive sleep apnea (OSA) represents the most prevalent sleep-related respiratory disorder, characterized by recurrent episodes of complete or partial upper airway obstruction during sleep. Although often asymptomatic, OSA exerts profound detrimental effects on cardiovascular, neurological, and pulmonary systems, thereby necessitating timely and accurate diagnosis. Conventional diagnostic approaches rely on polysomnography (PSG), which, despite its high diagnostic accuracy, remains costly, time-intensive, and dependent on specialized equipment and expert clinical supervision. These limitations in accessibility and operational complexity have prompted the development of more practical alternatives. Electrocardiogram (ECG)-based approaches have consequently attracted considerable attention owing to their capacity for continuous cardiac monitoring and sensitivity to subtle physiological alterations associated with sleep-disordered breathing. However, the inherent nonstationarity of ECG signals and substantial inter-subject variability continue to constrain model generalizability, thereby underscoring the critical need for robust and computationally efficient deep learning architectures. In this study, we propose a deep learning framework that integrates multiple surface-level ECG-derived features—including R-R intervals (RRI), ECG-derived respiration (EDR), and respiratory amplitude (RAMP)—extracted from the PhysioNet Apnea-ECG database. A record-wise data partitioning strategy was implemented to rigorously prevent data leakage across training, validation, and test sets. A hidden Markov model (HMM) was further incorporated as a post-processing module to refine apnea episode detection. Across five repeated hold-out validation experiments with varying training and validation partitions, the CNN-Transformer-LSTM architecture achieved an accuracy of 89.16 ± 0.94%, sensitivity of 81.42 ± 3.27%, and specificity of 94.00 ± 1.16%. Five-fold cross-validation yielded enhanced performance with accuracy of 90.62 ± 1.54%, sensitivity of 84.15 ± 3.80%, and specificity of 94.44 ± 2.22%. Integration of the HMM module further improved classification performance by approximately 5.00–6.00%, demonstrating the efficacy of the proposed framework for reliable and efficient OSA screening in both clinical and home-based monitoring applications.