AI-Driven Mendelian Randomization for Biomarker Discovery in Rheumatoid Arthritis: A Systematic Review of Precision Medicine Approaches

Background: Rheumatoid arthritis (RA) still represents a heterogeneous autoimmune disease where the predictability of biomarkers is an important early step that is pivotal in personalized therapy. Mendelian randomization (MR) provides causal estimates of candidate biomarkers, despite the possible limitation of traditional MR with weak instruments and pleiotropy. Recent advances of artificial intelligence (AI) into MR processes hold the promise of better instrument selection and model stability, potentially revolutionizing precision medicine regarding RA. Objective: To perform a systematic review and synthesize of evidence about the application of AI-enhanced Mendelian randomization (AI-MR) to identify and confirm diagnoses and prognostic biomarkers in RA. Methods: PubMed, EMBASE, Web of science, Scopus, Cochrane library, IEE Xplorer and Google Scholar were searched systematically due to publications during 2012 to March 2025. Qualified articles used AI-based feature selection or pleiotropy-control of MR-based biomarker prioritization applications in RA. Two independent reviewers screened and extracted the data. The quality of the methodology was assessed with the help of ROBINS-I and QUADAS-2 and PRISMA 2020 guidelines were adhered to. Random-effects meta-analysis combined odds ratios (ORs), sensitivity, specificity and area under the curve (AUC). Correlation analysis was used to examine relationships between characteristics of study design and predictive accuracy. Results: Of the initial 4,615 records, 34 studies were included and they included 7,200 RA cases and 11,000 controls in multi-ethnic cohorts. Instruments were reinforced using AI techniques that included LASSO, elastic net, random forest, and deep neural pleiotropy filters ( median F-statistic 34). Combined diagnostic performance of AI-MR biomarker panels was sensitivity 0.81 (95% CI 0.770.85) specificity 0.79 (95% CI 0.740.83) and AUC 0.86. IL-6, TNF and IL-1b had the strongest causal information (average pooled OR of 1.5), and autoantibody panels (ACPA, RF) and metabolomic responses had moderate incremental value. Correlation was used to determine that sample size was positively correlated to AUC (r = 0.71), and the reproducibility was better with the use of AI-guided pleiotropy checks. External validation was only done in 41 percent of studies. Conclusions: AI-enhanced Mendelian randomization is a better causal inference than conventional MR and has higher diagnostic accuracy in RA biomarkers. Precision rheumatology can be improved by combining AI based optimization of instruments and multi-omics inputs, however, pipeline standardization, transparent reporting, and large multicenter validation are all pressing requirements before clinical culture adoption.