Multimodal fusion prediction model based on prospective plasma proteomics: preclinical risk prediction and key pathway analysis of rheumatoid arthritis

Rheumatoid arthritis (RA) management remains largely reactive rather than proactive. although most patients are now diagnosed before radiographic erosions develop, the systemic immunological alterations that precede clinical onset are still not captured by routine clinical and laboratory assessment. To address this gap, this study aimed to identify preclinical plasma proteomic signatures of RA and to develop an integrated multimodal predictive model for precise risk stratification among individuals at risk. The study included 47,537 participants from the population-based UK Biobank cohort, with a median follow-up of 14.27 years. High-throughput plasma proteomic profiling of 2,920 proteins was performed once at baseline using the Olink proximity extension assay. RA-associated proteins were identified using Cox proportional hazards regression. Time-to-diagnosis-aligned clustering was used to characterize differences in baseline protein levels according to time to clinical diagnosis (up to 15 years). Mendelian randomization (MR) was performed to infer causal associations between plasma proteins and RA. A dual-branch residual DeepSurv model integrating proteomic and clinical features was developed for RA risk screening and validated in an independent held-out cohort ( N = 9,446). A total of 213 plasma proteins were significantly associated with the onset of rheumatoid arthritis (RA). Trajectory analysis indicated persistent molecular regulatory dysregulation throughout the long preclinical phase. Cis-Mendelian randomization analysis suggested that ICAM3 (odds ratio [OR] = 1.283; 95% confidence interval [CI]: 1.185–1.390) and IL18BP (OR = 0.635; 95% CI: 0.445–0.885) play causal roles in the development of rheumatoid arthritis. Drug–target matching analysis identified 112 targetable genes, corresponding to 580 approved drugs. The integrated model had a mean C-index of 0.777 in the development cohort and 0.786 in the independent validation cohort, outperforming the clinical-only model (0.723/0.736) and the protein-only model (0.737/0.738). In the validation cohort, the time-dependent area under the curve (AUC) reached 0.915 at 3 years, 0.867 at 5 years, and 0.810 at 10 years; the Kaplan–Meier hazard ratio for the integrated model was 3.45 (95% CI: 2.86–4.16). Specificity validation showed that the C-index for osteoarthritis (0.627) was significantly lower than that for rheumatoid arthritis (0.738), confirming the model’s predictive performance for rheumatoid arthritis. This study delineates the pre-clinical proteomic landscape of RA and establishes a high-performance multi-modal predictive model, offering selectable options for RA risk stratification, early disease screening, and potential drug repurposing.