Causal inference framework for personalised b/tsDMARD selection in rheumatoid arthritis: ANSWER cohort validation

Selecting the optimal biologic or targeted synthetic disease-modifying antirheumatic drug (b/tsDMARD) for patients with rheumatoid arthritis remains challenging. We developed and externally validated a causal inference framework combining double machine learning-based causal forests with guideline-derived safety constraints. Using multi-centre registry data from eight Japanese rheumatology facilities (4885 treatment courses), the model was trained on six facilities ( n = 2425) and externally validated on two facilities ( n = 2460). The framework estimates individualised treatment effects and applies a guideline-based safety rule that excludes Janus kinase inhibitors for patients with multiple cardiovascular risk factors. Patients receiving AI-concordant treatments had higher response rates than those receiving discordant treatments: 33.6% vs 27.2% in crude analysis (absolute risk difference [ARD] = 6.4%, P = 0.002; number needed to treat [NNT] = 16), with consistent effects after propensity score matching (ARD = 5.5%, P = 0.037) and inverse probability of treatment weighting (ARD = 4.3%, P = 0.044). The model identified treatment-effect heterogeneity (group average treatment effects Q4–Q1: Z = 4.64, P < 0.001) and 12 clinically interpretable effect modifiers confirmed by two analytical methods. Benefit was greatest in second-line patients (ARD = 16.0%, NNT = 6.3, P = 0.0002), supporting personalised b/tsDMARD selection.