Machine Learning to Predict Upper Airway Stimulation Outcomes

Introduction: Upper Airway Stimulation (UAS) provides significant therapeutic and quality of life benefits for patients with obstructive sleep apnea (OSA), however optimal patient selection is still unclear. In this study, we investigate the use of machine learning algorithms in predicting UAS treatment success.

Methods: A total of 537 patients who underwent UAS from 2014-2025 were included. Preoperative features including demographics, comorbidities, sleep study data, and drug induced sleep endoscopy data were retrospectively collected. Missing values were imputed using k-nearest neighbors. Model development employed 10-fold cross-validation with hyperparameter tuning within each training fold using nested resampling. Recursive feature elimination was completed to avoid data leakage. Algorithms tested included logistic regression, k-nearest neighbors, support vector machines, random forests, naive Bayes, AdaBoost, convolutional neural network, and XGBoost. Model performance was assessed with accuracy and area under the curve (AUC). Analysis was completed in R version 4.5.1.

Results: Random Forest (accuracy 0.69±0.051; AUC 0.71±0.075) and XGBoost (accuracy 0.69±0.055; AUC 0.71±0.085) demonstrated the highest classification performance. Other algorithms tested ranged from 0.44-0.65 accuracy. The five features most predictive of treatment success based on Random Forest included apnea hypopnea index, oxygen nadir, time below 90% oxygen, supine apnea hypopnea index, and national area deprivation index.

Conclusions: Machine learning provides increased accuracy in treatment success prediction as compared to standard predictive modeling. Additional sample size and features may improve model performance.