A multifaceted approach for obstructive sleep apnea classification from ECG signal using deep learning
Obstructive sleep apnea (OSA) is a common sleep disorder associated with increased cardiovascular and neurocognitive risks. While polysomnography remains the clinical gold standard for diagnosis, it is costly and unsuitable for large-scale or real-time screening. Electrocardiogram (ECG) signals offer a non-invasive, low-cost alternative for sleep apnea detection. We present a holistic new framework for OSA detection and forecasting using ECG data on two datasets: PhysioNet Apnea-ECG datasets (healthy patients with apnea), and OSASUD dataset (patients in a stroke unit). Our framework integrates feature engineering methods rooted in dynamical systems theory and statistical analysis. These features are used across a range of models, from conventional machine learning algorithms to novel deep learning architectures. To improve generalization and personalization, we incorporate transfer learning in two ways: across datasets to adapt models trained on large cohorts to smaller clinical datasets, and at the patient level to personalize models using limited individual data, hence demonstrating the use of precision medicine.