Abstract 4368861: The Impact of Artificial Intelligence-Derived Assessment of Natural and Built Environments on Cardiovascular Risk: A Moderated Mediation and Geospatial Analysis from the UK Biobank

Background: Artificial Intelligence (AI) applied to high-resolution imagery offers a novel, scalable approach to comprehensively characterize nuanced environmental exposures. A growing body of evidence implicates the built and natural environment in atherosclerotic cardiovascular disease (ASCVD) risk. Hypothesis: We hypothesized that a proportion of observed ASCVD events attributable to AI-derived geospatial features from the natural and built environment is mediated by traditional risk factors and moderated by underlying genetic susceptibility. Methods: Individual-level data from the UK ( n > 502,000, UKB) and the US Mass General Brigham ( n > 144,000, MGB) biobanks served as training/test and validation cohorts, respectively. Using convolutional neural networks (CNNs), > 8,000 natural and built environment features extracted from Google Satellite Images (GSI) and Street Views (GSV) were used to derive cross-validated sparse partial least squares personalized geospatial GSI/GSV scores. Associations between these scores with MACE and key risk factors (LDL-C, BMI, SBP, T2DM) were modeled by multivariate Cox PH models, adjusted for demographics, clinical, and area-level socioeconomic status covariates. Polygenic Risk Score (PRS) moderation and causal mediation analyses were performed along with geospatial mapping. Results: Out of 7 GSI and 4 GSV derived scores, 10 were significantly associated with MACE (all p < 0.001), with hazard ratios ranging from 0.961 to 1.048, comparable in magnitude to social deprivation and air pollution measures in UKB. Adding GSI/GSV scores to traditional risk models combining demographics, socio-determinants of health, and risk factors significantly improved model fit (GSI: likelihood ratio test (LRT) = 279.3, p < 0.001; GSV: LRT = 99.3, p < 0.001). Findings were validated in the MGB cohort for 4 of 7 GSI and 2 of 4 GSV features. Significant interaction effects were observed between GSI/GSV and PRS for LDL-C (p = 0.007), BMI (p = 0.05), and T2DM (p = 0.013). Causal mediation analysis revealed significant indirect effects and mediated proportions of GSI/GSV on SBP (p < 0.0001; 0.7%), T2DM (p < 0.0001; 0.5%), and BMI (p < 0.0001; 9.7%). Conclusion: AI-derived environmental geospatial features are associated with ASCVD risk, partially mediated by traditional risk factors, while moderated by genetic susceptibility. These findings highlight the potential for scalable environmental risk modeling to advance precision cardiovascular prevention.