A Novel Approach for Atrial Fibrillation-related Obstructive Sleep Apnea Detection Using Enhanced Single-Lead ECG Features with Customized Deep Learning Algorithm

Abstract Study Objectives Atrial fibrillation (AF) and obstructive sleep apnea (OSA) are interrelated conditions that substantially increase the risk of cardiovascular complications. However, concurrent detection of these conditions remains a critical unmet need in clinical practice. Current home sleep apnea test (HSAT) devices often fail to detect arrhythmias essential for diagnosing OSA-associated AF due to limited ECG monitoring capabilities, and their integration with continuous positive airway pressure (CPAP) data for treatment optimization remains underutilized. Methods This study introduces SHHDeepNet, an advanced deep learning-based framework designed for the detection of OSA in patients with AF, leveraging enhanced features extracted from single-lead electrocardiogram (ECG) signals. The ECG signals were preprocessed and refined using reconstruction independent component analysis (RICA), which isolates statistically independent features for improved data representation. These features were subsequently classified using the customized SHHDeepNet architecture. SHHDeepNet utilizes advanced signal processing and deep learning techniques to enhance ECG-based detection of AF-associated OSA. Results The framework was validated using overnight ECG recordings from 101 subjects derived from the Sleep Heart Health Study Visit 1 (SHHS1) database, encompassing 36 prevalent AF (PAF) cases, 25 incident AF (IAF) cases, and 40 OSA cases. Detection performance was evaluated through binary classification (AF AH vs. AF non-AH) and multi-class classification (AF AH, AF non-AH, non-AF AH, and non-AF non-AH). During 5-fold cross-validation (5fold-CV), the framework achieved a binary classification accuracy of 98.22%, sensitivity of 96.8%, specificity of 99%, and an area under the curve (AUC) of 0.9981. For multi-class classification, 5fold-CV yielded 98.36% accuracy, 97.14% sensitivity, 98.77% specificity, and an AUC of 0.9975. Validation using leave-one-subject-out cross-validation (LOSO-CV) achieved a binary classification accuracy of 86.42%, sensitivity of 79.4%, specificity of 90.2%, and an AUC of 0.9372. For multi-class classification under LOSO-CV, the average accuracy, sensitivity, and F1-score were 86.7%, 72.6%, and 0.7224, respectively. External validation was performed on a cohort of 123 subjects from the Osteoporotic Fractures in Men (MrOS) database, comprising 68 cases of PAF and 55 cases of OSA. The proposed method achieved a multi-class classification accuracy of 88.51%, sensitivity of 73.50%, specificity of 91.34%, and an AUC of 0.9363. Conclusions These findings underscore the significance of simultaneous detection of AF and OSA, providing a more comprehensive evaluation of cardiovascular health. The proposed SHHDeepNet framework offers a promising tool to support clinical decision-making, enhance management strategies, and improve patient outcomes by mitigating the risks associated with these conditions.