Background: Rheumatoid arthritis (RA) is a chronic autoimmune disease characterized by marked heterogeneity in clinical presentation, inflammatory burden, and therapeutic response. Although several Janus kinase inhibitors (JAKi) are available, treatment selection in real-world practice remains largely empirical. Artificial intelligence (AI)–based predictive models may support precision medicine by identifying drug–patient matches most likely to achieve meaningful clinical benefit.

Objectives: To develop and validate drug-specific AI models for predicting clinical response to individual JAK inhibitors in RA using real-world data, to characterize molecule-specific determinants of therapeutic effectiveness, and to estimate the potential clinical impact of AI-guided treatment selection compared with observed physician prescribing.

Methods: We conducted a retrospective observational study using a longitudinal real-world cohort of RA patients treated with baricitinib, tofacitinib, upadacitinib, or filgotinib between 2017 and 2025. Drug-specific supervised machine learning models were developed to predict clinical response to individual Janus kinase inhibitors. Separate gradient boosting models were trained for each molecule using routinely collected baseline demographic, clinical, and laboratory variables, including disease duration, baseline DAS28-CRP, CDAI, and SDAI, prior exposure to biologic or targeted synthetic DMARDs, concomitant therapies, and markers of systemic inflammation, including composite indices derived from blood counts. The primary outcome was a clinically meaningful improvement (≥1.2-point reduction in DAS28-CRP at 6 months), with achievement of low disease activity by CDAI and/or SDAI as a secondary outcome. Model performance was assessed using nested stratified cross-validation, discrimination (AUC), calibration (Brier score), and decision curve analysis. Model interpretability was supported by feature importance analyses. The potential impact of AI-guided prescribing was estimated through off-policy analyses comparing predicted response probabilities under observed physician prescribing versus model-recommended treatment, while traditional multivariable regression and inverse probability of treatment weighting were used in parallel to estimate average treatment effects.

Results: The study included 210 RA patients (baricitinib n=82, upadacitinib n=77, filgotinib n=22, tofacitinib n=29). Overall, 102 patients (48.6%) achieved LDA according to CDAI and/or SDAI at follow-up. Rates of LDA differed across molecules: upadacitinib showed the highest proportion of patients achieving LDA 33.9%, followed by baricitinib 22.4%, filgotinib 14.5%, and tofacitinib 13.4%, with consistent results across CDAI and SDAI definitions. Despite these numerical differences, causal analyses demonstrated comparable average treatment effects across JAK inhibitors, suggesting that variability in observed LDA rates reflects differences in patient-level response profiles rather than intrinsic population-level superiority. All drug-specific AI models demonstrated excellent discrimination and calibration, with area under the ROC curve (AUC) values ranging from 0.81 to 0.86 and Brier scores ≤0.05. The baricitinib model achieved an AUC of 0.84 (95% CI 0.80–0.88), with response probability primarily driven by baseline CDAI (OR 0.89; p<0.001), patient global assessment (OR 0.93; p=0.006), and systemic immune-inflammation index (OR 0.86; p=0.003). The tofacitinib model showed lower discrimination (AUC 0.81, 95% CI 0.77–0.85) and greater sensitivity to inflammatory burden and disease chronicity. Upadacitinib demonstrated the highest overall performance (AUC 0.86, 95% CI 0.82–0.89) with preserved efficacy across a wide spectrum of inflammatory activity. Filgotinib showed robust performance (AUC 0.83, 95% CI 0.78–0.87) but a narrower optimal response window. In off-policy analyses focused on LDA, the AI-recommended strategy yielded a positive probability uplift in approximately 74–82% of patients compared with observed physician prescribing, although the magnitude of improvement was generally modest. A clinically relevant uplift in predicted LDA probability (≥0.10–0.15) was observed in a small subset of patients (18%).

Conclusions: AI-based, drug-specific predictive models can accurately identify distinct patient-level response profiles among JAK inhibitors in RA using routinely collected baseline data. While population-level effectiveness and average treatment effects are broadly comparable across molecules, substantial heterogeneity exists at the individual level. AI-guided treatment selection appears to provide small but consistent gains at the population level while identifying a minority of patients with clinically meaningful benefit. Integrating AI-driven predictions with clinical judgment may therefore improve the efficiency of JAK inhibitor allocation, reduce trial-and-error prescribing, and support the implementation of precision medicine in real-world rheumatology.

REFERENCES: NIL.

Acknowledgments: NIL.

Disclosure of Interests: None declared.