Development and Explainable Machine Learning Validation of a Novel Sleep Disturbance Ratio for Obstructive Sleep...

Background/Objectives: Obstructive sleep apnea (OSA) is traditionally classified according to the apnea-hypopnea index (AHI), although AHI alone does not fully capture the heterogeneity of disease severity. This study introduces a novel polysomnography-derived biomarker, the Sleep Disturbance Ratio (SDR), and evaluates its contribution to OSA severity classification using explainable machine learning approaches. Methods: A retrospective study was conducted using polysomnographic data from 767 adults who underwent overnight sleep studies. SDR was calculated as the logarithmic ratio between light sleep (N1 + N2) and restorative sleep (N3 + REM). Predictive models were trained and evaluated using four established machine learning algorithms: Decision Tree, Random Forest, Extreme Gradient Boosting (XGBoost), and Artificial Neural Network (ANN). Model performance was assessed using accuracy, Cohen's kappa, F1-score, multiclass AUC, ROC analysis, feature importance ranking, partial dependence plots, and clinical risk mapping. Results: Although SDR did not differ significantly across conventional OSA severity groups in univariate analysis (p = 0.77), explainable machine learning analyses consistently demonstrated that increasing SDR was associated with a higher probability of severe OSA, particularly in combination with lower mean oxygen saturation. SDR also showed strong physiological relevance by correlating positively with N2 sleep (r = 0.85) and negatively with N3 sleep (r = -0.83), supporting its role as a biomarker of sleep fragmentation. Among the predictive models, Random Forest achieved the highest classification accuracy (75.0%), whereas XGBoost demonstrated the best multiclass discrimination (AUC = 0.895) and the highest ROC performance for severe OSA (AUC = 0.962). ESS remained the most influential predictor across all models. Conclusions: This study introduces SDR as a novel polysomnography-derived biomarker that captures sleep architecture disruption beyond conventional AHI-based assessment. Although SDR is not an independent diagnostic marker, explainable machine learning analyses demonstrated that it provides complementary physiological information for OSA severity classification, particularly when integrated with oxygenation parameters.