Fig. 1: Study overview.

First, we extracted data from 52,006 UKB participants with a median follow-up time of 14.1 years, including 45 endpoints defined by three-character ICD10 codes, 1461 plasma proteomics, and 54 clinical predictors spanning demographic, lifestyles, physical measures, medical and medication history, family disease history, and serum assays. Next, we developed a proteomic neural network to generate proteomic risk scores (ProRS) for each endpoint. Downstream survival analysis was performed using Cox proportional hazard models to explore ProRS and clinical predictor sets individually or in combinations. The model establishment and evaluations were implemented through internal leave-one-region-out cross-validation. Our investigation not only focused on evaluating models’ efficacy in stratifying populations at risk but also aimed to inform their potential clinical utility. CPH model Cox proportional hazard model, ProRS proteomic risk score.