Abstract TH847: Explainable Machine Learning–Identified Protective Social, Behavioral, and Clinical Factors for Cardiovascular Diseases

Abstract

Background: Cardiovascular diseases (CVDs) remain the leading cause of death globally, yet traditional risk tools underweight protective social and behavioral contexts that shape cardiovascular resilience. Explainable machine learning (ML) offers a path to quantify such protective factors while maintaining transparency for population health deployment. Hypothesis: We hypothesize that explainable ML applied to a nationally representative survey can accurately distinguish adults without CVDs and identify a reproducible profile of protective social, behavioral, and clinical factors associated with lower disease prevalence. Methods: We analyzed adults from the 2021 Behavioral Risk Factor Surveillance System (n=116,608); 11 candidate predictors spanned demographics/socioeconomics (age, sex, race/ethnicity, income, insurance), behaviors (smoking, alcohol use, fruit/vegetable intake), and clinical/mental health (diabetes, depressive disorder). Baselines included logistic regression, random forests, support vector machines, and XGBoost that underwent Optuna hyperparameter tuning with nested cross-validation. Discrimination (area under the receiver-operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC)), accuracy, and predictive values, were reported; interpretability used SHAP to quantify direction and magnitude of feature effects. Results: CVD prevalence rose with age (1.3% at 18–24 years to 24.6% at ≥65 years) and was higher in men than women (17.2% vs 12.2%); prevalence was greater in Black non-Hispanic than Hispanic adults (16.7% vs 9.6%) and in lower- versus higher-income groups (23.6% for <$15,000 vs 7.5% for ≥$200,000). On held-out testing, the best XGBoost model achieved AUROC of 0.76 and AUPRC of 0.95, with high negative predictive value (90.49%) and substantially low positive predictive value (32.36%); cross-validation confirmed robustness (mean AUROC=0.76, high AUPRC=0.95) and competitive overall accuracy of 76.98±0.45 across five folds. SHAP analyses consistently highlighted protective profiles characterized by younger age, higher income, health insurance coverage, and absence of diabetes and depressive disorder; dietary variables contributed modestly. Conclusions: Our explainable ML model identified protective profiles against CVDs, while achieving reliable performance. These results support prevention strategies that improve access to care, reduce metabolic risk, and motivate prospective validation.

Ähnliche Arbeiten