Predictive Analytics in Public Health Genomics

Abstract

Population-scale genomic data -- from national biobanks enrolling hundreds of thousands to millions of participants -- are transforming public health from reactive disease management to proactive, genomically informed prevention. Polygenic risk scores (PRS) aggregate the effects of thousands to millions of common genetic variants into individual-level disease risk predictions, enabling stratified screening and early intervention for cardiovascular disease, cancer, diabetes, and psychiatric disorders. However, PRS clinical utility is limited by modest discrimination (typical AUROC 0.60-0.70 for common diseases), poor transferability across ancestries (30-80% accuracy loss in non-European populations), and the absence of frameworks integrating genomic risk with clinical, environmental, and social determinants. This study developed and evaluated six predictive analytics frameworks -- standard PRS (PRS-CS), PRS with clinical risk factors (PRS+CRF), machine learning integration (XGBoost on PRS+CRF+lifestyle), federated multi-ancestry PRS (FL-PRS), deep learning on raw genotypes (DeepPRS), and a proposed Genomic-Environmental Risk Integration Network (GERIN) combining ancestry-adaptive PRS, longitudinal clinical trajectories, geospatial environmental exposures, and social determinants via a temporal graph neural network -- for predicting 10-year incidence of five major diseases (coronary artery disease, type 2 diabetes, breast cancer, colorectal cancer, major depressive disorder) across four population biobanks: UK Biobank (n = 408,624), FinnGen (n = 342,499), BioBank Japan (n = 179,416), and All of Us (n = 245,388, multi-ancestry). GERIN achieved the highest discrimination across all diseases and populations: mean AUROC 0.812 (vs PRS-CS 0.648, PRS+CRF 0.724, XGBoost 0.762, FL-PRS 0.698, DeepPRS 0.742; p < 0.001 for all). Critically, GERIN reduced the European-to-non-European performance gap from 0.082 AUROC (PRS-CS) to 0.024 (GERIN; p < 0.001), achieving near-equitable prediction across ancestries. Net reclassification improvement over PRS+CRF was +12.4% for CAD and +14.8% for T2D, translating to identification of 8,400 additional high-risk individuals per million screened who would benefit from preventive intervention. These results establish multi-modal genomic-environmental integration as the state-of-the-art for public health genomics prediction.