Prototyping OSA: A New Framework for Predicting Cardiovascular Outcomes and CPAP Benefit

Rationale: Obstructive sleep apnea (OSA) is associated with major adverse cardiac events (MACE); continuous positive airway pressure (CPAP) has not reduced MACE in randomized trials. We hypothesized that multi-domain machine learning would identify OSA subgroups differing in cardiovascular risk and CPAP response. Methods: We conducted a secondary analysis of SAVE trial data. After preprocessing, Gaussian mixture models were fitted; Bayesian information criterion and C-index selected five prototypes. Cox models estimated MACE risk by prototype. Baseline quality of life (EQoL), Epworth Sleepiness Scale (ESS), and Framingham Risk Score (FRS) were compared; an existing causal survival forest model estimated prototype-specific CPAP effects on cardiovascular risk. Measurements and Main Results: After preprocessing, five OSA prototypes were identified among 2,676 participants using 88 baseline variables. Prototypes differed in selected demographic, cardiometabolic, medication, and sleep variables, including smoking status, diabetes, prior myocardial infarction/stroke, antiplatelet/lipid-lowering therapy, BMI, oxygenation metrics, apnea-hypopnea index, and ESS (overall p<0.001). They also differed in baseline EQoL, FRS, MACE rate, and predicted CPAP response. Prototype 1 had the highest ESS, FRS, and MACE rate, whereas Prototype 5 had the lowest FRS and MACE rate. Using the causal survival forest model to assess prototype-specific CPAP effects on cardiovascular risk, Prototype 5 was predicted to benefit from CPAP (p<0.05), whereas Prototypes 1, 3, and 4 showed no benefit or possible harm. Conclusion: These findings demonstrate heterogeneity in cardiovascular risk and predicted CPAP response, potentially explaining neutral trial results. After external validation, these prototypes could guide trial enrichment and targeted CPAP use to improve cardiovascular outcomes.