Early prediction of longitudinal treatment adherence in obstructive sleep apnea using machine learning approaches

Continuous Positive Airway Pressure (CPAP) is the most prescribed treatment for Obstructive Sleep Apnea (OSA), but adherence remains a critical challenge, especially long term. This study aims to predict CPAP adherence at 3, 6, and 12 months based on baseline patient status and usage data from initial 30 days of treatment. A retrospective cohort of 2180 patients was analyzed. Feature selection was performed using the maximum relevance minimum redundancy (mRMR) algorithm with bootstrapping. Three machine learning (ML) models were trained and evaluated: support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP). A core subset of early CPAP usage variables (median nightly use, days of use, and 30-day adherence) emerged as the most relevant features across all timepoints. Each prediction window also revealed exclusive features, suggesting that adherence at different stages may be driven by distinct factors. At 3 months, MLP exhibited the strongest predictive capacity (kappa = 0.823, AUC = 0.910). At 6 months, RF and SVM models yielded the highest results (kappa = 0.727, AUC = 0.863), whereas at 12 months, RF consistently outperformed the other algorithms (kappa = 0.698, AUC = 0.849), underscoring its robustness in forecasting long-term adherence. Our results suggest that ML algorithms can effectively predict CPAP adherence using very short-term usage data. Accordingly, our models are able to provide efficient early identification of patients at risk of non-adherence in order to support personalized interventions, improve outcomes, and reduce the healthcare burden in OSA management.