Abstract 6313: Integration of Genomic and Lab Data with Dynamic EHR Modeling

Abstract

Background: Polygenic risk scores (PRS) for cardiovascular disease are almost always used as static covariates in proportional-hazards (PH) models - a single number, fixed for life. However, genetic effects on coronary artery disease (CAD) and other CV outcomes are non-stationary, rising and falling with age; when genetic risk inflects reshapes prevention timing and within-phenotype heterogeneity. We reshape Aladynoulli, an interpretable Bayesian generative model that simultaneously discovers latent disease structure and predicts individual trajectories from the longitudinal electronic health record (EHR) and genome, by allowing age-varying genetic effects (gamma_slope) to accommodate non-proportional hazards, and jointly integrating cardiometabolic-lab integration. Methods: We trained Aladynoulli on longitudinal records of 400,000 UK Biobank (UKB) and 53,000 Mass General Brigham (MGB) participants, jointly modeling 348 disease phecodes and 40 routine labs across ages 30-80. For each of 21 latent disease signatures we learned a level (gamma_level) and age-velocity (gamma_slope) coefficient on each of 36 PRS, letting effects evolve with age. Prospective 1-year prediction used leave-one-batch-out cross-validation (40 batches; disjoint train/test), benchmarked against the no-slope model on identical folds. Results: gamma_slope captured non-proportional hazards in 9/10 CV disease-PRS pairs. For CAD, the observed Q4-vs-Q1 PRS hazard ratio fell from 7.1 at age 45 to 2.5 at age 70 (-0.18/yr); the slope model recovered this decline while the static model predicted a flat HR. The same age-attenuation held for atrial fibrillation (AF) and type 1 diabetes. The slope model improved 1-year prediction in 27/28 diseases (mean +0.10 AUC), with the largest CV gains (AUC, bootstrap 95% CI): ASCVD 0.925->0.980 (0.976-0.983), AF 0.907->0.985 (0.974-0.991), heart failure 0.886->0.980 (0.968-0.992), stroke 0.754->0.899 (0.858-0.936). Crucially, the same model is a discovery engine: it learned all 21 comorbidity signatures and their PRS drivers de novo, recovered established CAD, AF, and diabetes genetic associations, newly localized the age at which each genetic effect peaks, and the signatures replicated across UKB and MGB. Joint lab modeling in MGB further discovered cardiometabolic subtypes - LDL/cholesterol within the ischemic signature, glucose/HbA1c within the diabetes/hypertension signature. Conclusions: Modeling CV polygenic risk as time-varying rather than static lets a single interpretable model discover disease structure and predict individual trajectories at once - substantially improving 1-year prediction and recovering the age-attenuation that PH methods miss. Aladynoulli's interpretable structure allows it to uniquely capture the interaction between time and genomics unlike black-box EHR foundation models.