tsta

فارسی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Predicting Sleep Efficiency and Apnea Index Using ECG-Derived and Sleep Quality Features: A Machine Learning Approach

Authors :

Mahla Khodaverdi¹ Raheleh Davoodi²

1- دانشگاه شهید بهشتی تهران 2- دانشگاه شهید بهشتی تهران

Keywords :

Sleep efficiency،Apnea index،ECG،Machine learning،Feature selection

Abstract :

Sleep quality and obstructive sleep apnea profoundly influence cardiovascular function, cognition, and overall well-being, yet conventional monitoring approaches remain largely invasive or cumbersome, underscoring the imperative for streamlined, non-invasive alternatives. Herein, we present a machine learning framework that synergistically integrates electrocardiogram (ECG)-derived features with sleep quality metrics to forecast sleep efficiency and apnea-hypopnea index (AHI). Drawing upon the ECSMP (A Dataset on Emotion, Cognition, Sleep, and Multi-Modal Physiological Signals) dataset—encompassing recordings from 89 healthy participants—we curated a subset of 33 subjects whose data exhibited complete and unimpaired capture across all ECG-sleep modalities, thereby ensuring analytical fidelity; incomplete records from the remaining participants, attributable to recording artifacts or procedural inconsistencies, were judiciously excluded to uphold data integrity. From these selected recordings, 22 ECG-derived and sleep quality features were extracted and subsequently refined through recursive feature elimination (RFE) to mitigate redundancy and enhance predictive salience. We evaluated three regression models—Ridge Regression, Random Forest, and Gradient Boosting—employing subject-based 5-fold cross-validation to foster generalizability across individuals. For sleep efficiency, Ridge Regression attained a mean R² of 0.8734, indicating a high degree of explained variance; by comparison, Random Forest registered an R² of 0.2756 for AHI, which underscores the formidable obstacles in modeling sporadic apnea episodes amid constrained empirical resources. Feature importance scrutiny further illuminated wake hours and deep sleep ratio as preeminent correlates for sleep efficiency, complemented by deep sleep ratio and QRS amplitude for AHI. Collectively, this framework lays a promising foundation for non-invasive, individualized sleep monitoring, offering reliable estimates of sleep efficiency and preliminary insights into apnea patterns, albeit within the constraints of a modest sample size.