C100-07 Phenotype-based Thresholding of Hypoxemia and Autonomic Markers in Continuous Positive Airway Pressure (CPAP) Treatment for Cardiovascular Outcomes in Obstructive Sleep Apnea (OSA)

Rationale Current heterogeneity of treatment effect (HTE) estimation approaches in obstructive sleep apnea (OSA) either rely on physiologic metrics requiring secondary processing of raw polysomnography or use non-data-driven selection of clinical variables to determine who benefits or is harmed by continuous positive airway pressure (CPAP). Our group first demonstrated HTE in the ISAACC trial, but one predictor required derived physiologic inputs. Here, we overcome this limitation by applying robust HTE-focused machine learning methods in the larger SAVE trial to develop an easily implementable CPAP treatment-decision tool based solely on standard, clinically available patient characteristics. Methods Two independent clinicians screened 195 baseline variables in SAVE for routine clinical relevance, with adjudication by a third reviewer, yielding 61 non-collinear predictors. The dataset was randomly split into training (70%) and testing (30%) subsets. Model-based recursive partitioning with a Cox framework was used to identify subgroups with differential CPAP treatment effects on a composite cardiovascular outcome. Hyperparameters were tuned by minimizing the integrated Brier score, and model stability was assessed via bootstrap resampling. Results Model-based recursive partitioning identified the strongest modifiers of CPAP treatment effect as: time spent below 90% oxygen saturation (T90), body mass index (BMI), minimum heart rate, Epworth Sleepiness Scale score, baseline diastolic blood pressure, history of heart disease or stroke, and use of oral antidiabetic or nitrate-based medications. CPAP benefit was observed for T90≤20 min with BMI>26.8 (training hazard ratio [HR] 0.42 [0.22-0.82], p=0.01, n = 304; testing HR 0.94 [0.41-2.18], p=0.89, n = 147). CPAP harm emerged for 20<T90 ≤ 36 min without baseline organic nitrates (training HR 5.28 [2.22-12.59], p <0.001, n = 262; testing HR 1.22 [0.48-3.09], p=0.69, n = 91). Additional subgroups (e.g., T90 >36 min with 55<pulse≤60) showed borderline harm (training HR 2.49 [0.96-6.49], p=0.061; n = 148) and are considered exploratory. Conclusions Using robust HTE-focused machine learning applied to routine clinical variables, we identified clinically recognizable phenotypes with differential cardiovascular response to CPAP. Although test-set estimates did not reach statistical significance, the train-to-test directional consistency and biological plausibility of these patterns align with prior literature implicating nocturnal hypoxemia, autonomic state, and ischemic coronary physiology as key modifiers of CPAP effect. Importantly, these variables were not uniformly predictive across their full range; hypoxemia and autonomic tone influenced treatment response only within specific physiologic zones, indicating that commonly used metrics may require phenotype-based thresholding rather than linear interpretation. Taken together, these findings support prospective validation of phenotype-based thresholding approaches for assessing heterogeneity of CPAP treatment effect.