B80-3-02 Parsimonious Prediction Models in Mesa to Identify Non-sleepy Individuals With Obstructive Sleep Apnea at Risk for Cardiovascular Events

Abstract Rationale Untreated obstructive sleep apnea (OSA) increases the risk of cardiovascular (CV) disease, underscoring the need to identify at-risk patients. However, there is a lack of well-validated and easily accessible tools for this purpose. We hypothesized that using existing multimodal data from well-characterized datasets and applying machine learning (ML) methods, we could identify patient characteristics that accurately predict CV event risk. Methods We analyzed 2,237 participants from Multi-Ethnic Study of Atherosclerosis (MESA). Two endpoints were defined: atherosclerosis progression, defined as an increase in coronary artery calcium on ungated lung computed tomography scans; and incident CV events, including myocardial infarction, stroke, revascularization, heart failure, or CV death. After a rigorous preprocessing, 228 clinical, 316 sleep, and 15 CAC predictors were retained with multiple (100) imputation for missing data. For atherosclerosis progression, we trained seven ML classifiers: ridge regression, decision tree, K-nearest neighbors 1, multilayer perceptron 2, random forest 3, Gaussian Naive Bayes 4, support vector machine5, and XGBoost 6. For time to CV events, elastic net Cox Proportional Hazards (CPH)7,8, Random Survival Forests (RSF)9, Deep Survival Machines (DSM)10, and Deep CPH11,12 were used. Models were trained on 70% of the cohort and tested on 30% preserving event prevalence. Results Atherosclerosis progression models: Models using only clinical or sleep features achieved moderate discrimination (area under the precision-recall curve AUPRC = 0.62±0.04 for clinical vs 0.55±0.05 for sleep). Combining clinical and sleep data modestly improved performance (AUPRC≈0.63). Adding baseline CAC measures to clinical and sleep data markedly enhanced accuracy, with random forest and XGBoost achieving AUPRC=0.79 in five-fold cross-validation. Both models were validated in the holdout testing set with AUC=0.79 and 0.87, respectively. Top predictors included baseline CAC, age, systolic blood pressure, low-density lipoprotein cholesterol (LDL-C), oxygen desaturation index, and time under 90% oxygen saturation. CV event models: Across imputed datasets, the elastic net CPH model provided the best balance of performance (C-index=0.71; Integrated Brier score IBS=0.023-0.034). Multimodal models consistently outperformed clinical-only predictors. Feature importance analyses revealed that conventional risk factors (age, hypertension, LDL-C, diabetes) and OSA-specific markers (oxygen-desaturation index, average oxygen saturation, hypoxic burden, sleep fragmentation) were major contributors to CV event prediction. The addition of ungated CT-derived CAC features further enhanced performance, reflecting the complementary value of vascular calcification burden in multimodal risk prediction. Conclusions Developing parsimonious CV event risk prediction tools for OSA patients will significantly improve risk assessment and implementation of CV interventions in this group. This abstract is funded by: NIH