NeuralApnea Triage: Machine Learning Powered ECG Analysis System for Sleep Apnea Detection

Abstract

Sleep apnea is a respiratory disorder marked by recurrent airflow reductions during sleep, causing intermittent low oxygen levels and autonomic instability that can produce measurable changes in the electrocardiogram (ECG). Because polysomnography (PSG), the clinical gold standard for sleep apnea diagnosis, is resource-intensive and often associated with long wait times, there is strong interest in ECG-based triage tools that can screen large populations and flag higher-risk patients for confirmatory PSG. Using the PhysioNet Apnea-ECG dataset, we evaluate minute-level apnea detection by training a CNN–Transformer model with 5-fold cross-validation and selecting a checkpoint based on validation loss. On a held-out test cohort, the selected model achieves 89.21% accuracy, 84.24% sensitivity, 92.30% specificity, an 85.68% F1-score, and a 95.60% AUROC. We also provide Grad-CAM overlays and an explainability agent that summarizes high-confidence minutes for clinician review; performance is competitive with recent Apnea-ECG studies using the same protocol, though external validation is required before clinical deployment. Code repository: