Predicting Moderate to Severe Obstructive Sleep Apnea Using Machine Learning Models and Respiratory Indices

This study assesses the predictive performance of four machine learning models-Support Vector Regression (SVR), k-Nearest Neighbors (kNN), Random Forest, and Gradient Boosting-in evaluating obstructive sleep apnea (OSA) severity. The primary objective was to enhance model generalization, minimize overfitting, and optimize feature engineering for improved clinical applicability. The dataset comprised polysomnographic (PSG) data from both community members and OSA patients, with key variables such as REM fragmentation, HADSD scores, neck circumference, and respiratory indices. A 5fold cross-validation strategy was employed to ensure robustness across datasets, while polynomial feature transformations and Recursive Feature Elimination (RFE) were used to refine feature selection. Random Forest incorporated out-of-bag (OOB) evaluation, achieving an OOB score of 0.9741, indicating reliable model stability. The results showed consistent high performance across models, with Random Forest and SVR achieving average AUCs of 1.00 and 0.99, respectively. Random Forest demonstrated balanced sensitivity (0.89-1.00) and specificity (0.90-1.00) across folds, while Gradient Boosting provided competitive sensitivity (0.89-1.00) and specificity (0.70-1.00) with reduced risk of overfitting. These findings highlight the potential of combining machine learning models with complementary strengths for robust OSA severity prediction. The study suggests that such models can support clinical decision-making and intervention strategies. Future work will involve expanding the dataset to include participants from diverse backgrounds and further optimizing feature engineering for enhanced model generalizability and reliability.