Personalized comfort settings in positive airway pressure machines using a causal inference approach: a retrospective observational analysis.

Introduction: To evaluate whether data-driven personalization of positive airway pressure (PAP) comfort settings improves adherence by developing and evaluating a causal machine learning (ML) model for individualized recommendations. Methods: The model was developed using AirSense 10 data and independently validated in temporally separated AirSense 10 and AirSense 11 cohorts, including 90-day and 1-year follow-up. The primary outcome was Centers for Medicare & Medicaid Services (CMS)-defined PAP adherence (device usage ≥ 4 h/night on ≥ 70% of nights during the first 90 days). Group-level average treatment effect (ATE) was estimated using backdoor adjustment methods. A causal forest model was applied to estimate the conditional average treatment effect for recommending personalized settings. A propensity score-matched analysis compared outcomes between patients whose actual settings matched model recommendations vs. those remaining on device default settings. Results: The model estimated ATE at the group level was 2.9 percentage points (p < 0.001), indicating that personalized settings were associated with improved CMS adherence. Covariate balance was achieved (standardized mean difference < 0.1), and model calibration was strong (expected calibration error 0.91%). Patients whose actual PAP settings matched the model-recommended configuration had higher device usage than matched controls who remained on default settings. Subgroup analyses confirmed consistent benefits across age groups, gender, apnea-hypopnea index, and mask type. SHapley Additive exPlanations analysis identified minimum pressure, start pressure and age as key drivers of personalization. Independent validation showed sustained usage benefits across AirSense 10 and AirSense 11 cohorts, with residual AHI remaining below the clinical reference threshold and no clinically meaningful deterioration in mask leak. Discussion: Causal ML-based personalization of PAP comfort settings was associated with improved CMS adherence/device usage. Integration of comfort setting personalization into setup workflows may enhance PAP therapy usage. As a retrospective observational analysis, these findings are associative and hypothesis-generating; the causal-inference framework, refutation testing, and large-scale independent validation provide robustness beyond that of conventional observational studies, although prospective randomized evaluation remains the definitive next step.