Fig. 1: Study design.

a We extracted data from the UK Biobank, including clinical, genomic, metabolomic, and proteomic predictors, as well as six cardiovascular diseases (CVDs) defined by self-reported diagnoses, hospital episode statistics, and death records. b A two-stage data partition strategy was applied: the Multiomics cohort, which remained untouched throughout model training and was used as an independent validation set to evaluate the added value of integrating multiple omics data types, and separate training, validation, and geographic testing sets within the Metabolomics-only and Proteomics-only cohorts for model development. Our CardiOmicScore framework utilized two multitask deep neural networks to predict the risk of six CVDs from 168 metabolites and 2920 proteins separately, generating MetScore and ProScore for each CVD. We trained Cox proportional hazards (CPH) models on various combinations of polygenic risk score (PRS), MetScore, ProScore, and three predefined clinical predictor sets (i.e., AgeSex, Clin, and PANEL). c Model performance was first evaluated for MetScore and ProScore in their respective geographic testing sets, and subsequently for all feature combinations in the Multiomics cohort. Performance metrics included Harrell’s C-index, calibration plots, and net benefit curves, with confidence intervals estimated using 1000 bootstrap resamples. Icons were designed by Freepik from Flaticon (www.flaticon.com).