Investigation of HR and QT Variability for Monitoring Sleep Apnea: An Interpretable Machine Learning Approach

Polysomnography, the gold standard technique for monitoring sleep apnea, is a costly, cumbersome, and time-consuming process that often causes disturbance to sleep and, therefore, is unsuitable for long-term monitoring. This paper investigates the single-channel electrocardiogram (ECG) derived heart rate variability (HRV) and QT variability (QTV) features, which are low-cost and suitable for long-term monitoring for automated sleep apnea monitoring. Using HRV alone and HRV combined with QTV features, different classifiers were trained to distinguish apneic events from healthy sleep events. The proposed model is trained and tested using 70 full-night ECG recordings acquired from the PhysioNet apnea ECG database. The extreme gradient boosting classifier outperformed a series of classifiers with sensitivity, specificity, and accuracy of 82.70%, 76.34%, and 79.38%, respectively, for HRV features. Adding QT features improved the sensitivity, specificity, and accuracy to 84.18%, 82.15%, and 83.16%, respectively. The performance suggests that HRV and QTV features have the potential to detect sleep apnea. Moreover, its non-invasive nature and cost-efficiency make it more suitable for wearable-based sleep apnea monitoring.