Abstract Body (Do not enter title and authors here): Introduction
The American Heart Association’s PREVENT equation now includes zip-code level Social Deprivation Index (SDI), highlighting the growing role of socio-environmental (SE) factors in risk prediction. However, concerns remain about fully capturing the breadth of exposures, particularly those prevalent in disadvantaged populations, within existing tools. This study sought to develop and evaluate a novel SE Risk Score using machine learning, and to compare its utility against SDI in the context of Coronary Artery Calcium (CAC) screening.
Research Questions
Whether a novel machine learning-derived SE Risk Score incorporating more than 150 area level environmental pollution, social and economic variables, when added to CAC scores, improves Major Adverse Cardiovascular Events (MACE) prediction than SDI, in Blacks.
Methods
We analyzed CAC scores, MACE outcomes, demographics, and census-tract SE variables from the CLARIFY registry (NCT04075162), a large prospective study of no-charge CAC testing (84,233 White and 7,940 Black participants). A SE Score was derived using an XGBoost machine learning model. This SE Score was then compared to zip-code level SDI. Analyses included Cox proportional hazards models, mediation analysis, and model performance evaluation (Harrell’s C-index, area under the curve (AUC), calibration metrics, Net Reclassification Improvement [NRI]) for models including CAC alone, CAC+SE Risk Score, and CAC+SDI.
Results
Compared to Whites, Black participants had higher MACE (14.0% vs 6.4%), despite lower mean CAC (151.5 vs 175.5). Adding SE Score to CAC improved C-index from 0.681 to 0.712, while adding census-tract SDI yielded 0.705 and zip-code SDI 0.700 respectively. For Blacks, the AUC at Year 4 improved from 0.642 (CAC alone) to 0.669 (CAC+SE Risk Score), surpassing the 0.672 achieved with CAC+census-tract SDI (Figures 1 and 2). The improvement in risk reclassification was more pronounced for Black individuals (Net Reclassification Improvement: 0.153) than for White individuals (0.081). SE factors mediated 47.12% of the relationship between race and MACE.
Conclusion
The machine learning–derived SE Score outperformed SDI in predicting MACE, improving both discrimination and calibration. SE factors mediated the race–MACE link, and their inclusion with CAC scores significantly enhanced risk reclassification, particularly in Black individuals. More refined tools are needed to better assess and address risk in socially disadvantaged populations.