Machine learning–based estimation of AHI from polysomnography data without direct AHI calculation: Development of a physiological-parameter–based regression model

Machine Learning–Based Estimation of AHI from Polysomnography Data Without Direct AHI Calculation: Development of a Physiological-Parameter–Based Regression Model Objective: The aim of this study was to predict the Apnea–Hypopnea Index (AHI) using physiological parameters measured during polysomnography (PSG) that are not directly included in the standard AHI calculation formula. The study introduces a novel machine learning–based approach that integrates oxygenation, arousal, and sleep architecture variables into multidimensional predictive models. By focusing on both regression and classification tasks, it seeks to develop clinically applicable decision-support models capable of accurately identifying patients with severe Obstructive Sleep Apnea (OSA) at an early stage. Materials and Methods: This retrospective cross-sectional study included 922 individuals who had completed overnight PSG testing. Demographic and PSG parameters were recorded. Three different machine learning algorithms — Random Forest, XGBoost, and CatBoost — were employed for both regression (AHI prediction) and classification (severe OSA detection) tasks. Nested cross-validation and hyperparameter optimization were applied to prevent overfitting. Correlation, Variance Inflation Factor (VIF), and SHAP analyses were used to assess variable relationships, multicollinearity, and model explainability. Model performance was evaluated using ROC-AUC, PR-AUC, F1 score, MAE, RMSE, and Brier score metrics. Results: Oxygenation metrics (≥3% desaturation event count, mean desaturation percentage, and minimum SaO₂) and arousal index were identified as the most significant predictors of AHI. Among all algorithms, the CatBoost model achieved the best performance across both regression and classification analyses (ROC-AUC > 0.90, high PR-AUC, low MAE/RMSE, and good calibration). All VIF values were below 10, confirming the absence of critical multicollinearity. SHAP analysis revealed that oxygenation-related variables and arousal index had the greatest impact on model predictions. Decision Curve Analysis (DCA) demonstrated that CatBoost provided the highest net clinical benefit across all probability thresholds. Conclusion: This study shows that AHI can be accurately predicted using PSG-derived physiological parameters not included in its conventional formula. The developed machine learning models, particularly CatBoost, exhibit strong clinical potential in supporting early and accurate diagnosis of severe OSA. Oxygen desaturation and sleep fragmentation were found to play a decisive role in disease burden, independent of AHI. The models' high predictive accuracy and calibration indicate their potential for integration into automated scoring systems, improving diagnostic efficiency and cost-effectiveness in clinical sleep medicine