Abstract Body: Background: Premature cardiovascular disease (CVD) is a major contributor to lost life-years, yet widely used risk calculators (e.g., PCE, Framingham) frequently miss early signals in adults under 50. These tools also exhibit substantial racial and ethnic performance gaps, with C-statistics often 0.10 lower in Black and Hispanic populations compared with White populations. To address these limitations, this study applies a multi-domain, fairness-aware AI framework to capture the complex interplay of genomic, inflammatory, clinical, and social-environmental factors while ensuring equitable predictive performance.
Methods: Using the NIH All of Us Research Program (N=3,920 premature CVD cases; 80% from underrepresented groups), we integrated 30× whole-genome sequencing data (6.5M variants), serial inflammatory biomarkers (CRP, IL-6, TNF-α), electronic health records, and social determinants of health (SDOH). We developed a fairness-aware XGBoost survival model incorporating adversarial debiasing to minimize performance variation across racial and ethnic groups. Unsupervised clustering and graph neural networks were used to identify “immune–social interaction signatures” associated with early-onset events.
Results: Preliminary unconstrained models achieved an overall C-statistic of 0.81 but demonstrated marked disparities (White: 0.84 vs. Hispanic: 0.73). The fairness-constrained architecture harmonized performance across groups, achieving C-statistics >0.80 with <10% variation. Four distinct immune–social signatures emerged; individuals with high polygenic risk, chronic inflammation, and low food security had a 2.4-fold higher risk of premature myocardial infarction (95% CI: 1.9–3.0) compared with those assessed using clinical risk factors alone. Causal mediation analysis indicated that inflammatory pathways mediated approximately 40% of the association between structural discrimination and premature CVD.
Conclusions: Fairness-guided AI models can effectively integrate multidimensional data to uncover hidden predictors of premature CVD while avoiding the amplification of existing health inequities. By moving beyond linear risk assessments to equity-constrained, multi-domain modeling, clinicians can identify high-risk individuals decades before overt disease, offering a scalable blueprint for precision prevention and cardiovascular health equity.