Interpretable feature-based machine learning for automatic sleep detection using photoplethysmography

Sleep disorders affect millions globally, leading to serious health issues. Accurate sleep-wake classification is essential for diagnosis and management. While polysomnography is the gold standard, it is costly and invasive; photoplethysmography (PPG) offers a viable alternative. Using the Cyclic Alternating Pattern Sleep Database (84 participants, 85,542 epochs), we extracted 330 features and reduced dimensionality via statistical tests and the SelectFromModel method. To address class imbalance, we applied Adaptive Synthetic (ADASYN) sampling. A Random Forest model, validated with 20-fold cross-validation on the unbalanced dataset (75 features), achieved an F1 score of 89.05% but struggled with wake detection. With ADASYN balancing and 35 features, it achieved 88.57% sensitivity (sleep) and 71.31% specificity (wake), with an F1 score of 81.40%. This feature-based approach improves PPG-based sleep classification, supporting clinical adoption and integration into wearable devices for remote sleep monitoring.