Abstract A137: Toward a Deeper Understanding of PREVENT for 10 Year Atherosclerotic Cardiovascular Risk: Subgroup Fairness and Predictive Value of Social Determinants of Health

The American Heart Association’s Predicting Risk of Cardiovascular Disease Events (PREVENT) model offers a modern, race-free approach to risk prediction from a large contemporary cohort, but its fairness across subgroups and the added value of social determinants of health (SDOH) predictors remain underexplored in large real-world settings. We conducted a retrospective cohort study of 554,675 adults aged 30–79 years without baseline CVD, using de-identified electronic health records from Truveta, a multi-system U.S. clinical data platform (Figure 1). We evaluated three models, including the original PREVENT equations, and retrained PREVENT models (coefficients re-estimated in the Truveta cohort) with and without social determinants of health (SDOH) predictors. The primary outcome was 10-year ASCVD events. Model evaluation focused on two aspects: (1) Fairness across subgroups assessed using (a) percentile calibration plots, comparing Kaplan–Meier (KM)-estimated event rates across predicted risk percentiles within demographic and SDOH subgroups, and (b) the cross concordance index (xCI), quantifying how consistently models ranked earlier events across and within subgroups. (2) Incremental value of SDOH assessed by comparing discrimination, calibration, and fairness between retrained PREVENT models with and without SDOH predictors. The 10-year ASCVD event rate was 1.8%. Most subgroups exhibited consistent calibration (Figure 2) and xCI values (Figure 3). The most pronounced disparities were observed between Whites and Asians (event rate: 10.3% vs. 6.6% at the 95th percentile; xCI = 0.849 vs. 0.679), college-educated and non–college-educated individuals (event rate: 14.2% vs. 10.7% at the 95th percentile; xCI = 0.645 vs. 0.935), and private and public insurance groups (event rate: 0.7% vs. 1.5% at the 25th percentile; xCI = 0.578 vs. 0.859). Across all subgroups, the inclusion of SDOH predictors in the PREVENT model had minimal impact on discrimination and calibration performance, with most changes being small and directionally inconsistent. PREVENT showed fairness across demographic and SDOH subgroups, supporting its practical use to predict ASCVD risk. Adding SDOH predictors or recalibrating offered minimal incremental benefit, reinforcing the original PREVENT equations’ utility as a reliable, fair tool for diverse populations.

Duke Scholars

Author Matthew M. Engelhard Biostatistics & Bioinformatics, Division of Translational Bi ...