A causal inference framework for personalised b/tsDMARD selection in rheumatoid arthritis: multi-centre development and external validation from the ANSWER cohort

A causal inference framework for personalised b/tsDMARD selection in rheumatoid arthritis: multi-centre development and external validation from the ANSWER cohort | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A causal inference framework for personalised b/tsDMARD selection in rheumatoid arthritis: multi-centre development and external validation from the ANSWER cohort Kosuke Kita, Kosuke Ebina, Yuki Suzuki, Yuki Etani, Yasutaka Okita, and 21 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9197723/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Selecting the optimal biologic or targeted synthetic disease-modifying antirheumatic drug (b/tsDMARD) for individual patients with rheumatoid arthritis (RA) remains an unresolved clinical challenge. We developed and externally validated a causal inference framework that integrates a double machine learning-based causal forest model with guideline-derived safety constraints for b/tsDMARD category selection. Using multi-centre registry data (4,885 treatment courses from eight Japanese rheumatology facilities), the model was trained on six facilities (n = 2,425) and externally validated on two independent facilities (n = 2,460). The framework first estimates individualised treatment effects via the causal forest with double machine learning, then applies a guideline-based safety rule that exclude JAK inhibitors for patients with multiple cardiovascular risk factors before finalising recommendations. Patients receiving treatments concordant with AI recommendations achieved significantly higher clinical disease activity index response rates than those with discordant treatments: crude analysis 33.6% vs 27.2% (absolute risk difference [ARD] = 6.4%, P = 0.002; number needed to treat = 16), propensity score-matched analysis 33.6% vs 28.2% (ARD = 5.5%, P = 0.037), and inverse probability of treatment weighting 31.6% vs 27.3% (ARD = 4.3%, P = 0.044). The system identified significant heterogeneity in conditional average treatment effects (group average treatment effects Q4–Q1: Z = 4.64, P < 0.001) and 12 clinically interpretable effect modifiers confirmed by two independent analytical methods. Treatment line-stratified analysis revealed that second-line patients derived the greatest benefit (ARD = 16.0%, number needed to treat = 6.3, P = 0.0002). These findings demonstrate that an AI framework combining data-driven causal inference with guideline-derived safety guardrails can identify patients who benefit from personalised b/tsDMARD selection in RA. Health sciences/Diseases Health sciences/Medical research Health sciences/Rheumatology Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Rheumatoid arthritis (RA) is a chronic autoimmune disease where methotrexate (MTX), a conventional synthetic disease-modifying antirheumatic drug (csDMARD), remains the first-line treatment. However, approximately 20–30% of patients exhibit an inadequate response to MTX, and some patients are intolerant or have contraindications to the drug 1 , 2 . For these patients, clinicians face a complex array of therapeutic options, including biological DMARDs (bDMARDs) and targeted synthetic DMARDs (tsDMARDs), each with distinct mechanisms of action 3 , 4 . Current clinical guidelines, such as those by the Japan College of Rheumatology (JCR) and the European Alliance of Associations for Rheumatology (EULAR), list these agents as parallel options but lack evidence-based stratification criteria to predict which agent yields the highest efficacy for an individual patient. Drug selection therefore relies on empirical trials and safety profiles rather than predicted efficacy. This trial-and-error approach results in only 60–70% of patients achieving low disease activity or remission 3 , and delays in effective treatment lead to irreversible joint damage and functional impairment. Previous research has explored various predictors of treatment response. Single biomarkers such as rheumatoid factor (RF) and anti-cyclic citrullinated peptide (ACPA) antibodies have limited predictive accuracy for guiding drug selection 5 – 7 . Machine learning (ML) approaches integrating routine clinical data have demonstrated the ability to predict response to bDMARDs with moderate to high accuracy 8 , 9 . Recent studies have identified drug class-specific predictive features: age for tumour necrosis factor (TNF) inhibitors, disease duration for cytotoxic T-lymphocyte antigen 4–immunoglobulin (CTLA4-Ig), and C-reactive protein (CRP) levels for interleukin-6 (IL-6) inhibitors 8 . These findings suggest that routine clinical data contain signals capable of forecasting treatment outcomes. A critical gap, however, separates prediction from recommendation. Most ML studies predict the probability of response to a single drug, rather than comparing multiple options to determine the optimal choice for a specific patient. Estimating the individualised treatment effect (ITE), defined as the difference in potential outcomes under different treatments, is essential for precision medicine but has been difficult to implement in RA 10 . Causal ML methods, particularly causal forests based on the generalised random forest framework 11 , offer a principled approach to estimating heterogeneous treatment effects from observational data. Beyond algorithmic accuracy, responsible clinical deployment of AI-based treatment recommendation systems requires human oversight 12 , 13 . AI models trained on observational data may learn patterns that conflict with established safety evidence. Current JCR guidelines recommend caution with Janus kinase (JAK) inhibitors in patients who are elderly, smokers, or who possess risk factors for malignancy, cardiovascular disease, or thromboembolism 14 , yet a model trained solely on efficacy outcomes cannot incorporate such safety considerations. A clinician-guided design, where safety constraints derived from clinical guidelines are encoded as formal rules that override data-driven recommendations, ensures that the system operates within accepted clinical boundaries while preserving the AI’s ability to optimise within those boundaries. We developed and validated a clinician-guided causal inference framework for b/tsDMARD category selection in RA, by using data from the K ans ai Consortium for We ll-being of R heumatic Disease Patients (ANSWER) cohort, which is a large multicenter, observational RA registry in Japan 15 – 18 . The framework integrates a causal forest 11 with double ML 19 (CausalForestDML) for data-driven ITE estimation with clinician-encoded safety constraints that exclude specific drug categories for patients meeting predefined risk criteria. We retrospectively evaluated clinical utility through concordant versus discordant analysis with multiple confounding adjustment methods, characterised heterogeneous treatment effects, and identified clinically interpretable effect modifiers. Results Study population The training cohort comprised 2,425 treatment courses from six facilities and the external validation cohort comprised 2,460 treatment courses from two independent facilities (Fig. 1 ). Baseline characteristics are presented in Table 1 . Several significant differences were observed between cohorts, including age, disease activity such as Clinical Disease Activity Index (CDAI) and Simplified Disease Activity Index (SDAI), inflammatory markers (ESR), RF, and renal function (estimated Glomerular Filtration Rate; eGFR), Steinbrocker Stage and Class distributions, and concomitant MTX use. Despite these differences, both cohorts had similar sex distribution (approximately 82% female), disease duration, treatment line (median 2), and CDAI response rates (26.5% vs 28.9%, P = 0.067). The distribution of prescribed drug categories differed between cohorts (P < 0.001), with TNFi being the most frequently prescribed in both (training 40.7%, validation 46.6%; Table 1 ). Of 258 candidate predictors, 83 (32.2%) had no missing values; missing rates for the remaining predictors ranged from 0.1% to 78.8% (Supplementary Fig. S1 ). Table 1 Baseline characteristics of the training and external validation cohorts Variable Training (n = 2,425) Validation (n = 2,460) P value Age, years 61.4 ± 15.1 60.2 ± 15.1 < 0.001 Female sex, n (%) 1,962 (80.9%) 2,033 (82.6%) 0.125 BMI, kg/m² 21.8 ± 4.3 22.2 ± 4.0 0.086 Disease duration, months 91.4 [27.5–194.8] 102.8 [33.8–206.0] 0.009 Treatment Line 2.0 [1.0–3.0] 2.0 [1.0–3.0] 0.008 Steinbrocker Stage, n (%) < 0.001 I 710 (33.3%) 433 (29.4%) II 479 (22.5%) 406 (27.5%) III 389 (18.3%) 232 (15.7%) IV 551 (25.9%) 404 (27.4%) Steinbrocker Class, n (%) < 0.001 I 735 (34.4%) 366 (24.0%) II 907 (42.5%) 853 (56.0%) III 437 (20.5%) 276 (18.1%) IV 57 (2.7%) 27 (1.8%) CDAI 17.9 ± 11.8 15.3 ± 9.9 < 0.001 SDAI 19.5 ± 12.9 17.0 ± 10.8 < 0.001 DAS28-CRP 3.7 ± 1.4 3.6 ± 1.2 0.465 Patient VAS, cm 4.9 ± 2.8 5.0 ± 2.6 0.294 CRP, mg/dL 1.6 ± 2.5 1.5 ± 2.4 0.060 ESR, mm/h 35.4 ± 28.0 39.8 ± 29.7 < 0.001 RF, U/mL 70.0 [20.0–191.0] 53.2 [14.0–156.0] 15 U/mL), n (%) 1,830 (75.5%) 1,781 (72.4%) 0.016 ACPA, U/mL 51.1 [6.7–225.2] 49.2 [2.7–280.0] 0.492 ACPA positive (> 4.5 U/mL), n (%) 1,663 (68.6%) 1,654 (67.2%) 0.330 MMP-3, ng/mL 125.2 [66.4–259.2] 117.8 [61.1–247.0] 0.038 Haemoglobin, g/dL 12.2 ± 1.5 12.1 ± 1.6 0.872 eGFR, mL/min/1.73m² 74.5 ± 24.0 76.4 ± 24.2 0.005 Concomitant MTX use, n (%) 1,216 (50.1%) 1,355 (55.1%) < 0.001 Dose among users, mg/week 10.1 ± 5.9 8.0 ± 3.3 < 0.001 Concomitant PSL use, n (%) 845 (34.8%) 836 (34.0%) 0.546 Dose among users, mg/day 6.3 ± 4.4 6.4 ± 10.4 < 0.001 Responder, n (%) 643 (26.5%) 711 (28.9%) 0.067 Prescribed drug category < 0.001 TNFi, n (%) 986 (40.7%) 1,147 (46.6%) IL-6i, n (%) 542 (22.4%) 528 (21.5%) JAKi, n (%) 520 (21.4%) 410 (16.7%) CTLA4-Ig, n (%) 377 (15.5%) 375 (15.2%) Data are mean ± SD, median [IQR], or n (%). CDAI = Clinical Disease Activity Index. SDAI = Simplified Disease Activity Index. DAS28-CRP = Disease Activity Score in 28 joints using CRP. VAS = Visual Analogue Scale. CRP = C-reactive protein. ESR = erythrocyte sedimentation rate. RF = rheumatoid factor. ACPA = anti-cyclic citrullinated peptide antibody. MMP-3 = matrix metalloproteinase-3. eGFR = estimated glomerular filtration rate. MTX = methotrexate. PSL = prednisolone. TNFi = tumour necrosis factor inhibitor. IL-6i = interleukin-6 inhibitor. JAKi = Janus kinase inhibitor. CTLA4-Ig = cytotoxic T-lymphocyte-associated protein 4 immunoglobulin (abatacept). P values were calculated using Mann–Whitney U test for continuous variables and chi-square test for categorical variables. Causal forest with double machine learning and nuisance model selection We developed a CausalForestDML model, which combines Causal Forests with Double Machine Learning. This framework estimates individualised treatment effects while adjusting for confounding through auxiliary propensity and outcome models. Among five candidate models compared on the training cohort (Supplementary Table S1 ), Random Forest achieved the narrowest 95% confidence interval width for the doubly robust estimator (0.072), indicating the most stable treatment effect estimation. Random Forest was therefore selected for both the propensity and outcome models. AI recommendation distribution and safety constraint effects The clinician-guided causal inference framework operates in three sequential stages (Fig. 2 ): (1) patient baseline data are input to the CausalForestDML model, which estimates ITEs for four drug categories (TNFi, JAKi, IL-6i, CTLA4-Ig); (2) clinician-defined safety constraint exclude JAKi for patients meeting two or more cardiovascular risk factors (age ≥ 65, smoking history, body mass index [BMI] ≥ 30), in accordance with the JCR guidelines 14 ; and (3) the category with the highest estimated ITE among eligible options is selected as the final recommendation. The trained model was frozen and applied to the external validation cohort. When JAKi was excluded, the category with the next-highest ITE was recommended. After applying the safety constraint, the model recommended JAKi for 53.6% (n = 1,318), TNFi for 28.9% (n = 711), IL-6i for 17.4% (n = 427), and CTLA4-Ig for 0.2% (n = 4) of external validation patients. Safety constraints excluded JAKi for 241 patients (9.8%), redirecting their recommendations to alternative categories. Treatment recommendations showed significant age-dependent heterogeneity (chi-square P < 0.001): elderly-onset RA (EORA) patients (age of onset ≥ 60, n = 542) received JAKi recommendations at 64.6% with TNFi at 14.9% and IL-6i at 20.5%, while young-onset RA patients (n = 1,477) had JAKi at 48.3%, TNFi at 34.7%, and IL-6i at 16.8% (Supplementary Fig. S2a). Among EORA patients, 122 (22.5%) had JAKi excluded by safety constraints, increasing IL-6i and TNFi recommendations in this group (Supplementary Fig. S2b). Concordant versus discordant analysis Patients receiving treatments concordant with AI recommendations had significantly higher CDAI response rates across all three analytical approaches (Table 2 ). In crude analysis (n = 664 concordant, n = 1,796 discordant), response rates were 33.6% vs 27.2% (ARD = 6.4%, relative risk [RR] = 1.24, odds ratio [OR] = 1.36, P = 0.002; NNT = 16). After propensity score matching (PSM; n = 660 matched pairs), response rates were 33.6% vs 28.2% (ARD = 5.5%, P = 0.037; NNT = 18). Inverse probability of treatment weighting (IPTW) yielded weighted response rates of 31.6% vs 27.3% (ARD = 4.3%, P = 0.044; NNT = 23). Table 2 Concordant versus discordant analysis in the external validation cohort Analysis Concordant Discordant ARD RR OR P value NNT E-value Crude (n = 664 vs 1,796) 33.6% 27.2% 6.4% 1.24 1.36 0.002 16 1.60 PSM (n = 660 matched pairs) 33.6% 28.2% 5.5% — — 0.037 18 1.53 IPTW 31.6% 27.3% 4.3% — — 0.044 23 1.46 ARD = absolute risk difference. RR = relative risk. OR = odds ratio. NNT = number needed to treat. PSM = propensity score matching. IPTW = inverse probability of treatment weighting. E-value = E-value for the point estimate. — indicates RR and OR not calculated; PSM used McNemar’s test for matched pairs and IPTW used bootstrap inference. The attenuation from crude to IPTW estimates (ARD 6.4% to 4.3%) reflects increasingly conservative confounding adjustment, with all three methods maintaining statistical significance. E-values for unmeasured confounding were 1.60 (crude), 1.53 (PSM), and 1.46 (IPTW) (Table 2 ). Subgroup and sensitivity analyses Disease duration subgroup analysis Concordant versus discordant analysis stratified by disease duration revealed stage-dependent benefit (Table 3 ). Established RA (2–10 years) demonstrated the strongest effect (ARD = 13.8%, NNT = 7.2, P = 0.0005), while long-standing RA (> 10 years) showed no benefit (ARD = − 0.2%, P = 1.000). Table 3 Subgroup analyses by disease duration and treatment line Subgroup n Concordant n Discordant Concordant (%) Discordant (%) ARD NNT P value Disease duration Early RA ( 10 years) 233 673 23.6% 23.8% −0.2% — 1.000 Treatment line 1st line (biologic-naive) 262 795 40.1% 34.0% 6.1% 16.4 0.086 2nd line 171 408 39.8% 23.8% 16.0% 6.3 0.0002 3rd+ line 231 593 21.6% 20.4% 1.2% — 0.765 ARD = absolute risk difference. NNT = number needed to treat. Crude analysis results shown. — indicates NNT not calculated due to non-significant or negligible ARD. Treatment line stratification Analysis stratified by treatment history demonstrated the strongest AI recommendation benefit in second-line treatment (ARD = 16.0%, NNT = 6.3, P = 0.0002), corresponding to a NNT approximately 2.5 times lower than the overall cohort (NNT = 16). This represented the strongest benefit observed across all subgroups (Table 3 ). Heterogeneous treatment effects Group average treatment effects Group average treatment effects (GATEs) analysis demonstrated an increasing pattern across conditional average treatment effect (CATE) quartiles: Q1 (GATE = − 0.116, SE = 0.042), Q2 (− 0.034, SE = 0.044), Q3 (0.002, SE = 0.044), and Q4 (0.163, SE = 0.043). The Q4–Q1 difference was significant (difference = 0.279, Z = 4.64, P < 0.001; Fig. 3 ), validating the AI system’s ability to rank patients by predicted treatment benefit. Category-specific GATEs analyses confirmed significant heterogeneity for JAKi (Q4–Q1: Z = − 2.59, P = 0.005) and CTLA4-Ig (Z = − 3.63, P < 0.001; Supplementary Table S2). Effect modifiers Twelve covariates were identified as robust effect modifiers by both Best Linear Projection (BLP) and Lasso methods (Supplementary Table S3). These grouped into clinically meaningful categories: disease activity markers (Disease Activity Score using 28 joints with C-Reactive Protein [DAS28-CRP], CRP, ESR, Patient Visual Analog Scale [VAS]), haematological parameters (White Blood Cell [WBC], haemoglobin, lymphocyte), demographics (age), structural damage (Matrix Metalloproteinase-3 [MMP-3], Steinbrocker Stage), and treatment history (past golimumab use, tacrolimus use). The convergence of a tree-based approach (BLP) and a linear regularised approach (Lasso) provide model-independent evidence for genuine treatment effect heterogeneity. CATE-SHAP analysis CATE-SHAP analysis identified CRP as the most influential driver of between-patient variation in estimated treatment effects, followed by Patient VAS, haemoglobin, MMP-3, and ESR. Renal function (eGFR), serological markers (RF, ACPA), and demographic factors (age) also contributed to treatment effect heterogeneity (Fig. 4 ). These features substantially overlap with the 12 robust effect modifiers identified independently by BLP and Lasso, providing convergent evidence from three distinct analytical approaches. Discussion This study developed and externally validated a clinician-guided causal inference framework for b/tsDMARD category selection in RA. The framework combines CausalForestDML for data-driven treatment effect estimation with clinician-encoded safety constraints that exclude JAKi for patients with multiple specific risk factors. Four principal findings emerged: (1) AI-concordant treatment was associated with significantly higher CDAI response rates across unadjusted, PSM, and IPTW analyses; (2) this effect was maintained despite significant between-cohort heterogeneity in baseline characteristics; (3) the system identified 12 clinically interpretable covariates driving patient-level differences in treatment benefit; and (4) the safety constraint layer appropriately modified 9.8% of recommendations, with the resulting age-dependent patterns consistent with current safety guidelines. The concordant versus discordant analysis provides direct evidence of clinical utility. The NNT of 16 indicates that for every 16 patients treated according to AI recommendations, one additional patient would achieve CDAI response, with the benefit remaining significant across all three confounding adjustment methods. E-value analysis indicated that an unmeasured confounder would need to be associated with both concordant treatment selection and CDAI response by a risk ratio of at least 1.60 to fully explain the observed effect. The monotonically increasing GATEs pattern validates the core premise of personalised treatment recommendation: the model meaningfully ranks patients by predicted treatment benefit. The 12 robust effect modifiers grouped into clinically coherent categories (disease activity, haematological parameters, demographics, and structural damage), and CATE-SHAP analysis confirmed substantial overlap with these modifiers. This convergence across three methodologically distinct approaches — a tree-based method (BLP), a linear regularised method (Lasso), and a model-explanation method (SHAP) — strengthens confidence that the model captures genuine biological heterogeneity rather than spurious correlations. Subgroup analyses further clarified the clinical contexts in which personalised recommendations confer the greatest benefit. Disease duration subgroup analysis revealed that AI recommendations provided the greatest benefit in established RA (2–10 years; ARD = 13.8%, NNT = 7.2, P = 0.0005), with a trend in early RA ( 10 years; ARD = − 0.2%, P = 1.000). This pattern likely reflects that in established RA, sufficient clinical heterogeneity has emerged to enable treatment differentiation, yet disease remains modifiable, whereas in long-standing RA, treatment-refractory inflammation and structural damage may limit model discriminative capacity. Treatment line-stratified analysis revealed the strongest benefit in second-line treatment (ARD = 16.0%, NNT = 6.3, P = 0.0002), with diminishing benefit in third-line and beyond (ARD = 1.2%, P = 0.765), suggesting that AI recommendations provide maximal value when treatment options remain relatively abundant. These findings extend previous ML studies in RA, which have focused on predicting response to specific drugs with AUCs of 0.60–0.91 20,21 but cannot directly compare expected outcomes across treatment options for individual patients. To our knowledge, this is the first study to apply a causal inference framework (CausalForestDML) combined with clinician-guided safety constraints to b/tsDMARD selection in RA with multi-centre external validation. The concordant versus discordant design provides a more rigorous validation framework than internal validation commonly reported in the literature 8 , 21 , and the system maintained significant ARD despite substantial between-cohort heterogeneity in disease activity, inflammatory markers, structural damage (Steinbrocker Stage and Class), and demographics. Notably, the proportion of JAKi recommendations (53.6%) contrasts with current prescribing patterns where TNFi remains the most commonly prescribed first-line biologic DMARD 22 , 23 , which reflects the model’s focus on optimising treatment outcomes rather than reproducing existing practice. The integration of clinician-defined safety constraints addresses this discrepancy by ensuring that efficacy-driven recommendations remain within accepted clinical boundaries. The JCR guidelines recommend caution with JAK inhibitors in patients with specific risk factors 14 , and encoding this guideline as formal rules within the recommendation pipeline ensures that safety considerations, which the CausalForestDML model cannot learn from efficacy-outcome training data alone, are incorporated into final recommendations. In this cohort, 241 patients (9.8%) had JAKi excluded by safety constraints, and 122 of 542 EORA patients (22.5%) had recommendations redirected to alternative categories. This constraint layer operates transparently: the clinician can review both the unconstrained and constrained recommendations, understand which rules were triggered, and retain full authority over the final prescribing decision. This design aligns with growing consensus that clinical AI systems should augment rather than replace physician judgement 12 , 13 , and that safety-critical domains require explicit human oversight mechanisms. This study has several limitations. First, the retrospective observational design precludes causal conclusions; although PSM, IPTW, and E-value sensitivity analyses were employed, residual unmeasured confounding cannot be excluded. Second, the study population was derived from Japanese rheumatology registries; treatment patterns, drug availability, and genetic backgrounds may differ in other populations. Therefore, evaluation in non-Japanese populations remains necessary to assess cross-ethnic applicability. Third, the safety constraints are currently limited to JAKi specific risk factors, and due to the nature of the cohort, the risk factors of malignancy could not be sufficiently evaluated.; expansion to other drug-specific safety considerations (such as infection risk stratification for bDMARDs) would improve the framework’s clinical applicability. Fourth, the near-absence of CTLA4-Ig recommendations (0.2%) likely reflects confounding by indication rather than true therapeutic inferiority. In our cohort, patients prescribed CTLA4-Ig (abatacept) had the highest mean age among all drug categories (67.7 years), suggesting that clinicians preferentially selected this agent for elderly or frail patients in whom safety considerations outweighed efficacy optimisation. These patients — characterised by advanced age, frailty, and elevated infection risk that precluded JAKi or TNFi — represent a population in whom CDAI improvement is inherently more difficult to achieve. While DML adjusts for observed confounders, unmeasured factors such as frailty severity, infection risk burden, and clinician safety concerns cannot be fully controlled. Consequently, the model may have attributed the poorer outcomes observed in CTLA4-Ig-treated patients to the drug category itself rather than to the underlying patient characteristics that drove its selection. This residual confounding by indication represents a fundamental limitation of observational treatment effect estimation and should be considered when interpreting the low CTLA4-Ig recommendation rate. To facilitate clinical access and independent evaluation, we have developed a publicly available web-based tool that enables clinicians to input patient characteristics and obtain individualised treatment recommendations with the safety constraint layer applied. Future directions include prospective randomised validation, expansion of safety constraints to additional drug classes, evaluation in non-Japanese populations. Methods Study design and data source This retrospective multi-centre cohort study analysed data from the ANSWER (Kansai Consortium for Well-being of Rheumatic Disease Patients) cohort, a large-scale registry of patients with RA in the Kansai region of Japan comprising nine university-related hospitals 15 – 18 . Data were available from eight of these hospitals; Nara Medical University did not contribute data during the study period and was excluded from the analysis. The study included 4,885 treatment courses from these facilities. Analysis was restricted to treatment courses with evaluable CDAI outcomes (responder or non-responder). The study was conducted in accordance with the Declaration of Helsinki and is reported in accordance with the TRIPOD + AI guidelines 24 . The two largest facilities were designated as the external validation cohort (combined n = 2,460), while the remaining six facilities served as the training cohort (n = 2,425), ensuring adequate statistical power for detecting clinically meaningful differences while maintaining sufficient training data for model development. Participants and eligibility criteria The study period encompassed treatment courses initiated between January 2000 and June 2024. Eligible participants were patients diagnosed with RA according to either the 1987 American College of Rheumatology (ACR) classification criteria 25 or the 2010 ACR/EULAR classification criteria 26 , who were registered in the ANSWER cohort and initiated treatment with a b/tsDMARD (TNFi, IL-6i, JAKi, or CTLA4-Ig). The administration of these agents was at the discretion of the attending rheumatologists, consistent with the JCR guidelines 14 , 27 . Treatment courses were excluded if CDAI outcome data were unavailable for determining the outcome. Outcome The outcome was CDAI-based treatment response, defined as a binary variable (1 = Responder, 0 = Non-Responder). The Responder group included patients meeting either: (1) achieving LDA (CDAI ≤ 10) or remission within six months after treatment initiation; or (2) maintaining treatment for more than six months with clinical benefit. The Non-Responder group consisted of patients who failed to meet these criteria. CDAI was calculated as the sum of tender joint count (28 joints), swollen joint count (28 joints), Patient Global Assessment (VAS, 0–10 cm), and Evaluator Global Assessment (VAS, 0–10 cm). Outcome assessment was performed by treating rheumatologists at each facility using standardised joint count and VAS procedures. Outcome assessors were not blinded to treatment allocation, consistent with routine clinical practice in observational registry data. Predictors All candidate predictors were extracted from the registry without pre-selection. These included demographic factors (age, sex, BMI), clinical characteristics (disease duration, smoking history, Steinbrocker stage/class), disease activity scores (CDAI, SDAI, DAS28-CRP, Patient VAS), laboratory values (RF, ACPA, CRP, ESR, MMP-3, WBC, haemoglobin, platelet count, eGFR, Krebs von den Lungen [KL]-6), treatment history (treatment line, past usage of individual b/tsDMARDs), and current medication details (MTX dose, prednisolone dose), yielding 222 candidate features. All predictors were measured at or before the time of b/tsDMARD initiation. Continuous predictors were used without transformation. Categorical variables were encoded as binary indicators. Missing data handling Missing values were handled using median imputation 28 , applied separately to training and validation datasets to prevent data leakage. Imputation procedures were applied identically across all subgroups. Causal forest with double machine learning CausalForestDML integrates Causal Forests 11 with Double Machine Learning (DML) 19 , a framework that uses ML models to estimate confounding components (nuisance functions) while preserving valid statistical inference for treatment effects. Nuisance model selection Five ML models (Random Forest, CatBoost, Logistic Regression, HistGradientBoosting, and LightGBM) were compared using default hyperparameters, prioritising estimation stability measured by the width of the 95% confidence interval for the doubly robust estimator. Selection was based on internal validation using training data only. Model configuration Random Forest was selected for both the propensity and outcome models. The CausalForestDML system (EconML: a python package) comprised a propensity model estimating drug category assignment probability given patient characteristics, an outcome model predicting treatment response, and a causal forest with 500 trees and 5-fold cross-fitting for nuisance estimation. Bootstrap with 5,000 iterations was used for confidence interval estimation. CausalForestDML estimates individual treatment effects by solving local linear moment equations at each tree node. The doubly robust estimator combines outcome model predictions with inverse propensity weighting for robust policy evaluation. Class imbalance The treatment outcome was imbalanced (responders ~ 27%). Class imbalance was addressed through inverse class frequency weighting in both the propensity and outcome Random Forest models. No resampling techniques were applied. Clinician-guided safety constraints Treatment recommendations were subject to clinician-defined safety constraints before concordant versus discordant evaluation. This clinician-guided layer was designed to integrate established post-marketing safety evidence that cannot be learned from efficacy-outcome training data alone. JAKi were excluded from the recommended category when patients met two or more of the following cardiovascular risk factors, in accordance with the JCR guidelines 14 : (1) age ≥ 65 years, (2) past or current smoking history, and (3) BMI ≥ 30 kg/m². When JAKi was excluded, the category with the next-highest estimated ITE was selected. The safety constraints operate as transparent, auditable rules: the treating clinician can review both the unconstrained and constrained recommendations, understand which rules were triggered, and exercise final prescribing authority. This framework is designed to be extensible; additional safety rules (for example, infection risk stratification for specific bDMARDs) can be incorporated as supporting evidence emerges. Validation protocol All model training and model selection were performed solely on the training dataset. The finalised model was frozen and applied to the external validation cohort for one-time evaluation. External data were not used to update model weights or any model development processes, ensuring complete independence of the external validation. Model output The CausalForestDML system outputs estimated ITEs for each drug category relative to a reference category. The recommended drug category is the one with the highest estimated ITE after applying safety constraints. The model does not output calibrated probability estimates for individual response; rather, it outputs the treatment category expected to yield the greatest relative benefit. No risk group thresholds were applied. Fairness assessment Model performance was evaluated across age subgroups (EORA vs young-onset RA). Given the predominantly Japanese study population, race-based fairness assessment was not applicable. Age-based fairness was addressed through the JAKi safety constraints and subgroup analyses. Concordant versus discordant analysis To evaluate clinical utility, we compared outcomes between patients whose prescribed drug category matched the AI-recommended category (Concordant group) and those with a mismatch (Discordant group). Three analytical approaches assessed robustness to confounding: (1) crude analysis using the chi-square test; (2) propensity score matching with logistic regression using 15 covariates (age, sex, BMI, CDAI, SDAI, CRP, ESR, RF, Treatment Line, Steinbrocker Stage, Steinbrocker Class, eGFR, disease duration, haemoglobin, Patient VAS), 1:1 nearest-neighbour matching with a caliper of 0.2 × SD of the propensity score; and (3) IPTW using stabilised inverse probability weights with bootstrap inference (1,000 resamples). Effect measures included ARD, RR, OR, and NNT (= 1/ARD). Heterogeneous treatment effect analysis Patients were stratified into four groups (Q1–Q4) based on estimated CATEs. GATEs within each quartile were estimated using the doubly robust estimator. SHAP analysis SHAP analysis was applied to the CausalForestDML model to identify patient characteristics driving heterogeneity in CATEs. SHAP values were computed separately for each treatment comparison (JAKi vs TNFi, IL-6i vs TNFi, CTLA4-Ig vs TNFi) using TreeExplainer applied to the causal forest’s internal tree structure. To obtain an aggregate measure of feature importance, SHAP values were averaged across the three pairwise comparisons and the mean absolute SHAP value per feature was used for ranking. Analysis was performed on the external validation cohort (n = 2,460). Subgroup analyses Disease duration subgroups (early RA [ 10 years]) and treatment line strata (first line, second line, third+ line) were analysed using the same three concordant versus discordant approaches. Treatment recommendation distributions were compared between EORA and young-onset RA using the chi-square test. E-values were calculated for all concordant versus discordant effect measures to quantify robustness to unmeasured confounding 29 . Statistics and reproducibility All statistical tests were two-tailed with α = 0.05 unless otherwise specified. The chi-square test was used for categorical comparisons. The Mann–Whitney U test was used for continuous variables that did not meet normality assumptions. PSM significance was assessed by McNemar’s test. IPTW significance was assessed by bootstrap inference (1,000 resamples). GATEs analyses used the doubly robust Z-test. Multiple testing was not adjusted for in subgroup analyses, which were pre-specified and exploratory. Sample sizes were determined by the available registry data; no formal power calculation was performed a priori. All analyses were conducted using Python 3.10 with econml, scikit-learn, and shap libraries. Results are reproducible from the code repository. Ethics statement The study protocol was reviewed and approved by the Ethics Committee of Osaka Metropolitan University, serving as the central institutional review board (approved on 9 June 2021; approval no. 2021-074). All participating institutions were included under this central approval and obtained site-specific authorization for study implementation in accordance with local regulatory requirements. The study was conducted in accordance with the Declaration of Helsinki. At each participating institution, informed consent was obtained either through comprehensive consent approved by the respective institutional ethics committee or through an opt-out procedure, as appropriate. Declarations Data availability The datasets collected and analysed during the current study are not publicly available due to patient privacy and ethical restrictions imposed by the institutional review boards but are available from the corresponding author on reasonable request. Code availability The source code developed for the analysis and the machine learning model in this study is publicly available on GitHub at https://github.com/kosukekita/RA_AI_npjDM . The web-based implementation of the recommendation system is accessible at https://ra-frontend-noauth-1099185790534.us-central1.run.app/predict . Competing Interests Y.E. is affiliated with, K.E. holds an additional post in, and K.N. supervises the Department of Sports Medical Biomechanics, The University of Osaka Graduate School of Medicine Faculty of Medicine, which is supported by Asahi Kasei. K.E. has received research grants fromEli Lilly, and speaker fees from AbbVie, Amgen, Argenx, Asahi Kasei, Astellas, Ayumi, Bristol-Myers Squibb, Chugai, Daiichi Sankyo, Eisai, Eli Lilly, Janssen, Mitsubishi Tanabe, Ono Pharmaceutical, Pfizer, Sanofi, Taisho, Teijin Pharma, and UCB Japan. Y.E. has received a speaker fee from Amgen, Asahi-Kasei, Astellas, Chugai, Daiichi-Sankyo, Eisai, Gilead Sciences Japan, Taisho, and UCB Japan. K.T. has received speaker fees from Ayumi, Mikasa, Eisai, and Asahi-Kasei Corporation. A.K. has received a research grant from Chugai and speaker fees from Chugai, AbbVie, and Ono Pharmaceuticals. Ko.M. is affiliated with a department that is financially supported by Asahi-Kasei Pharma Corp. and the city government (Nagahama City and Toyooka City), and has received speaking and/or consulting fees from Asahi Kasei Pharma Corp., Chugai Pharmaceutical Co., Ltd., UCB Japan Co., Ltd., Eli Lilly, AbbVie, Eisai Co., Ltd., and Daiichi Sankyo Co., Ltd. Mo.H. has received research grants and/or speaker fees from Asahi Kasei, Astellas, AstraZeneca, Ayumi Pharma, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences Japan, Janssen Pharma, Ono Pharma, Ohtsuka Pharma, Taisho Pharma, Tanabe Mitsubishi, and UCB Japan. T.O. has received speaker fees and/or research grants from AbbVie, Asahi Kasei, Astellas, Daiichi Sankyo, Eli Lilly, and UCB. H.Y. has received payments for lectures from AbbVie, Asahi Kasei, Astellas, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences, Ono, Otsuka, Pfizer, Taiho, and Taisho. To.T. has received research grants from Boehringer Ingelheim and Chugai, and payments for lectures from Astellas, Chugai, Eisai, Eli Lilly, MSD, Otsuka, and Taisho. Yu.N. has received speaker fees from Astellas Pharma, Asahi Kasei Pharma, Eisai, Eli Lilly, AstraZeneca, and Tanabe Pharma. K.K., Y.S., T.N., A.S., Y.O., Yu.M., T.F., Y.U., H.S., Y.So., Hi.M., D.T., W.Y., S.O., and Ma.H. declare no competing interests. Funding The ANSWER Cohort was supported by grants from 12 pharmaceutical companies (AbbVie GK, Asahi-Kasei, Ayumi, Chugai, Eisai, Eli Lilly, Janssen KK, Ono, Sanofi KK, Taisho, Teijin Healthcare, and UCB Japan) and an information technology service company (CAC). This study was conducted as an investigator-initiated study. This study was funded by JSPS KAKENHI Grant-in-Aid for Scientific Research (25K02761) and JCR Research Promotion Program for Optimal Medical Care in Elderly-Onset Rheumatoid Arthritis. The funders had no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. Author Contribution K.K. performed the formal analysis and wrote the original draft of the manuscript. K.E. conceptualised the study, administered the project, and reviewed and edited the manuscript. Y.S. provided methodology and guidance on software and code analysis. Y.E., T.N., A.S., Y.O., K.T., Yu.M., A.K., Ko.M., T.F., Mo.H., T.O., H.Y., Y.U., To.T., H.S., Y.So., Hi.M., Yu.N., D.T., and W.Y. contributed to data curation and reviewed and edited the manuscript. K.N. and S.O. supervised the study. Ma.H. supervised the study and provided critical revision of the manuscript. All authors read and approved the final manuscript. Data Availability The datasets collected and analysed during the current study are not publicly available due to patient privacy and ethical restrictions imposed by the institutional review boards but are available from the corresponding author on reasonable request. References D’Onofrio, B. et al. Timely escalation to second-line therapies after failure of methotrexate in patients with early rheumatoid arthritis does not reduce the risk of becoming difficult-to-treat. Arthritis Res Ther 26, 192 (2024). Huang, Y. et al. Factors influencing prescribing the first add-on disease-modifying antirheumatic drugs in patients initiating methotrexate for rheumatoid arthritis. Explor Res Clin Soc Pharm 11, 100296 (2023). Smolen, J. S. et al. EULAR recommendations for the management of rheumatoid arthritis with synthetic and biological disease-modifying antirheumatic drugs: 2022 update. Annals of the Rheumatic Diseases 82, 3–18 (2023). Harigai, M. et al. 2024 Update of the Japan College of Rheumatology Clinical Practice Guidelines for the Management of Rheumatoid Arthritis: Secondary publication. Modern Rheumatology 35, 387–401 (2025). Lend, K. et al. Association of rheumatoid factor, anti-citrullinated protein antibodies and shared epitope with clinical response to initial treatment in patients with early rheumatoid arthritis: data from a randomised controlled trial. Annals of the Rheumatic Diseases 83, 1657–1665 (2024). Wientjes, M. H. M., Den Broeder, A. A., Welsing, P. M. J., Verhoef, L. M. & Van Den Bemt, B. J. F. Prediction of response to anti-TNF treatment using laboratory biomarkers in patients with rheumatoid arthritis: a systematic review. RMD Open 8, e002570 (2022). Lv, Q. et al. The Status of Rheumatoid Factor and Anti-Cyclic Citrullinated Peptide Antibody Are Not Associated with the Effect of Anti-TNFα Agent Treatment in Patients with Rheumatoid Arthritis: A Meta-Analysis. PLoS ONE 9, e89442 (2014). Koo, B. S. et al. Machine learning model for identifying important clinical features for predicting remission in patients with rheumatoid arthritis treated with biologics. Arthritis Res Ther 23, 178 (2021). Alsaber, A. R. et al. Machine learning-based remission prediction in rheumatoid arthritis patients treated with biologic disease-modifying anti-rheumatic drugs: findings from the Kuwait rheumatic disease registry. Front. Big Data 7, 1406365 (2024). Curth, A., Peck, R. W., McKinney, E., Weatherall, J. & Van Der Schaar, M. Using Machine Learning to Individualize Treatment Effect Estimation: Challenges and Opportunities. Clin Pharma and Therapeutics 115, 710–719 (2024). Athey, S., Tibshirani, J. & Wager, S. Generalized random forests. The Annals of Statistics 47, (2019). Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 25, 44–56 (2019). Kelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G. & King, D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med 17, 195 (2019). Kawahito, Y. [Guidelines for the management of rheumatoid arthritis]. Nihon Rinsho 74, 939–943 (2016). Fujisawa, Y. et al. Baseline neutrophil-to-lymphocyte ratio predicts drug retention of IL-6 inhibitors and JAK inhibitors in RA: the ANSWER cohort study. Rheumatology 65, keaf602 (2026). Hiramatsu, Y. et al. Effect of treat-to-target strategies on maternal and neonatal outcomes of rheumatoid arthritis: a multicentre real-world ANSWER cohort study. Rheumatology 65, keaf558 (2026). Nozaki, Y. et al. Clinical efficacy of JAK inhibitors for RA patients with poor-prognosis factors: the ANSWER cohort study. Rheumatology 65, keaf552 (2026). Tsujimoto, K. et al. Sustained efficacy of second-line JAK inhibitors in patients with rheumatoid arthritis: insights from the ANSWER cohort. Rheumatology 64, 4207–4217 (2025). Chernozhukov, V. et al. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal 21, C1–C68 (2018). Benavent, D. et al. Artificial intelligence to predict treatment response in rheumatoid arthritis and spondyloarthritis: a scoping review. Rheumatol Int 45, 91 (2025). Salehi, F. et al. Machine Learning Prediction of Treatment Response to Biological Disease-Modifying Antirheumatic Drugs in Rheumatoid Arthritis. JCM 13, 3890 (2024). Pappas, D. A. et al. Comparative effectiveness of first-line tumour necrosis factor inhibitor versus non-tumour necrosis factor inhibitor biologics and targeted synthetic agents in patients with rheumatoid arthritis: results from a large US registry study. Ann Rheum Dis 80, 96–102 (2021). Edgerton, C. et al. Real-World Treatment and Care Patterns in Patients With Rheumatoid Arthritis Initiating First-Line Tumor Necrosis Factor Inhibitor Therapy in the United States. ACR Open Rheumatol 6, 179–188 (2024). Collins, G. S. et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ e078378 (2024) doi: 10.1136/bmj-2023-078378 . Arnett, F. C. et al. The American Rheumatism Association 1987 revised criteria for the classification of rheumatoid arthritis. Arthritis Rheum 31, 315–324 (1988). Aletaha, D. et al. 2010 rheumatoid arthritis classification criteria: an American College of Rheumatology/European League Against Rheumatism collaborative initiative. Ann Rheum Dis 69, 1580–1588 (2010). Koike, R. et al. Japan College of Rheumatology 2009 guidelines for the use of tocilizumab, a humanized anti-interleukin-6 receptor monoclonal antibody, in rheumatoid arthritis. Modern Rheumatology 19, 351–357 (2009). Berkelmans, G. F. N. et al. Population median imputation was noninferior to complex approaches for imputing missing values in cardiovascular prediction models in clinical practice. J Clin Epidemiol 145, 70–80 (2022). VanderWeele, T. J. & Ding, P. Sensitivity Analysis in Observational Research: Introducing the E-Value. Ann Intern Med 167, 268–274 (2017). Additional Declarations Competing interest reported. Y.E. is affiliated with, K.E. holds an additional post in, and K.N. supervises the Department of Sports Medical Biomechanics, The University of Osaka Graduate School of Medicine Faculty of Medicine, which is supported by Asahi Kasei. K.E. has received research grants fromEli Lilly, and speaker fees from AbbVie, Amgen, Argenx, Asahi Kasei, Astellas, Ayumi, Bristol-Myers Squibb, Chugai, Daiichi Sankyo, Eisai, Eli Lilly, Janssen, Mitsubishi Tanabe, Ono Pharmaceutical, Pfizer, Sanofi, Taisho, Teijin Pharma, and UCB Japan. Y.E. has received a speaker fee from Amgen, Asahi-Kasei, Astellas, Chugai, Daiichi-Sankyo, Eisai, Gilead Sciences Japan, Taisho, and UCB Japan. K.T. has received speaker fees from Ayumi, Mikasa, Eisai, and Asahi-Kasei Corporation. A.K. has received a research grant from Chugai and speaker fees from Chugai, AbbVie, and Ono Pharmaceuticals. Ko.M. is affiliated with a department that is financially supported by Asahi-Kasei Pharma Corp. and the city government (Nagahama City and Toyooka City), and has received speaking and/or consulting fees from Asahi Kasei Pharma Corp., Chugai Pharmaceutical Co., Ltd., UCB Japan Co., Ltd., Eli Lilly, AbbVie, Eisai Co., Ltd., and Daiichi Sankyo Co., Ltd. Mo.H. has received research grants and/or speaker fees from Asahi Kasei, Astellas, AstraZeneca, Ayumi Pharma, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences Japan, Janssen Pharma, Ono Pharma, Ohtsuka Pharma, Taisho Pharma, Tanabe Mitsubishi, and UCB Japan. T.O. has received speaker fees and/or research grants from AbbVie, Asahi Kasei, Astellas, Daiichi Sankyo, Eli Lilly, and UCB. H.Y. has received payments for lectures from AbbVie, Asahi Kasei, Astellas, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences, Ono, Otsuka, Pfizer, Taiho, and Taisho. To.T. has received research grants from Boehringer Ingelheim and Chugai, and payments for lectures from Astellas, Chugai, Eisai, Eli Lilly, MSD, Otsuka, and Taisho. Yu.N. has received speaker fees from Astellas Pharma, Asahi Kasei Pharma, Eisai, Eli Lilly, AstraZeneca, and Tanabe Pharma. K.K., Y.S., T.N., A.S., Y.O., Yu.M., T.F., Y.U., H.S., Y.So., Hi.M., D.T., W.Y., S.O., and Ma.H. declare no competing interests. Supplementary Files supplementaryinformation20260323.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 03 May, 2026 Reviewers invited by journal 09 Apr, 2026 Editor assigned by journal 24 Mar, 2026 Submission checks completed at journal 24 Mar, 2026 First submitted to journal 23 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9197723","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":624272360,"identity":"653c9513-115b-4340-86a5-3f62360fa7bd","order_by":0,"name":"Kosuke Kita","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kosuke","middleName":"","lastName":"Kita","suffix":""},{"id":624272361,"identity":"bf01b484-c0f8-4205-9f80-c67b6430f8d7","order_by":1,"name":"Kosuke Ebina","email":"data:image/png;base64,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","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Kosuke","middleName":"","lastName":"Ebina","suffix":""},{"id":624272362,"identity":"1c01d2aa-7407-4f2f-80bd-aea6ee1aec00","order_by":2,"name":"Yuki Suzuki","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yuki","middleName":"","lastName":"Suzuki","suffix":""},{"id":624272363,"identity":"08c771d2-8f27-46ed-968b-54aba9e1927d","order_by":3,"name":"Yuki Etani","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yuki","middleName":"","lastName":"Etani","suffix":""},{"id":624272364,"identity":"73c87788-3709-4b81-b38f-b5642c616132","order_by":4,"name":"Yasutaka Okita","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yasutaka","middleName":"","lastName":"Okita","suffix":""},{"id":624272365,"identity":"8571d5f6-0a34-4a35-b668-da872fb5be6d","order_by":5,"name":"Kohei Tsujimoto","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kohei","middleName":"","lastName":"Tsujimoto","suffix":""},{"id":624272366,"identity":"6287b6ec-a957-4dcd-acc8-d54d7349cd31","order_by":6,"name":"Yuichi Maeda","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yuichi","middleName":"","lastName":"Maeda","suffix":""},{"id":624272367,"identity":"bf4fc93f-14f6-4201-8901-798d1a82e876","order_by":7,"name":"Takaaki Noguchi","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Takaaki","middleName":"","lastName":"Noguchi","suffix":""},{"id":624272368,"identity":"cd198774-72dc-490a-a3cd-53a9b812043b","order_by":8,"name":"Atsushi Sugimoto","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Atsushi","middleName":"","lastName":"Sugimoto","suffix":""},{"id":624272369,"identity":"8ae7e88a-5967-445d-ba81-2d48b60ec821","order_by":9,"name":"Koichi Murata","email":"","orcid":"","institution":"Kyoto University Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Koichi","middleName":"","lastName":"Murata","suffix":""},{"id":624272370,"identity":"ec51a444-3f1f-4d57-8778-db50dac705dd","order_by":10,"name":"Takayuki Fujii","email":"","orcid":"","institution":"Kyoto University Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Takayuki","middleName":"","lastName":"Fujii","suffix":""},{"id":624272371,"identity":"4ce05087-56af-4fbf-8ce2-d578ee41d481","order_by":11,"name":"Motomu Hashimoto","email":"","orcid":"","institution":"Osaka Metropolitan University Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Motomu","middleName":"","lastName":"Hashimoto","suffix":""},{"id":624272372,"identity":"2a243251-14dd-49c6-a707-d5e91155cceb","order_by":12,"name":"Tadashi Okano","email":"","orcid":"","institution":"Osaka Metropolitan University Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Tadashi","middleName":"","lastName":"Okano","suffix":""},{"id":624272373,"identity":"c1023204-fcc1-4598-98dc-536b1d7b94a1","order_by":13,"name":"Hirotaka Yamada","email":"","orcid":"","institution":"Kobe University Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Hirotaka","middleName":"","lastName":"Yamada","suffix":""},{"id":624272374,"identity":"a86c6798-6d70-47dd-8026-fb73c39d01eb","order_by":14,"name":"Yo Ueda","email":"","orcid":"","institution":"Kobe University Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yo","middleName":"","lastName":"Ueda","suffix":""},{"id":624272375,"identity":"1f40798e-2430-47b3-9048-0aaa9930e1ea","order_by":15,"name":"Tohru Takeuchi","email":"","orcid":"","institution":"Osaka Medical and Pharmaceutical University","correspondingAuthor":false,"prefix":"","firstName":"Tohru","middleName":"","lastName":"Takeuchi","suffix":""},{"id":624272376,"identity":"0aea42bf-b614-4b85-92f3-d98348e5e9ff","order_by":16,"name":"Hideyuki Shiba","email":"","orcid":"","institution":"Osaka Medical and Pharmaceutical University","correspondingAuthor":false,"prefix":"","firstName":"Hideyuki","middleName":"","lastName":"Shiba","suffix":""},{"id":624272377,"identity":"e135809d-5d3f-4ba9-8032-602303375eea","order_by":17,"name":"Yonsu Son","email":"","orcid":"","institution":"Kansai Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yonsu","middleName":"","lastName":"Son","suffix":""},{"id":624272378,"identity":"1bd2f665-50c1-4b19-90c0-84b11f463a95","order_by":18,"name":"Hidehiko Makino","email":"","orcid":"","institution":"Kansai Medical University","correspondingAuthor":false,"prefix":"","firstName":"Hidehiko","middleName":"","lastName":"Makino","suffix":""},{"id":624272380,"identity":"0cdbfcbd-1099-4227-8e3a-c3d82c8f56c3","order_by":19,"name":"Yuji Nozaki","email":"","orcid":"","institution":"Kindai University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yuji","middleName":"","lastName":"Nozaki","suffix":""},{"id":624272381,"identity":"c85bdae6-f522-43a3-8c7e-18d5f98f29f3","order_by":20,"name":"Daisuke Tomita","email":"","orcid":"","institution":"Kindai University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Daisuke","middleName":"","lastName":"Tomita","suffix":""},{"id":624272382,"identity":"e66ecee9-5102-4175-a829-038a48037f5c","order_by":21,"name":"Wataru Yamamoto","email":"","orcid":"","institution":"Kurashiki Sweet Hospital","correspondingAuthor":false,"prefix":"","firstName":"Wataru","middleName":"","lastName":"Yamamoto","suffix":""},{"id":624272383,"identity":"e753c68a-29f0-4a0a-b1b0-28d92bac166a","order_by":22,"name":"Atsushi Kumanogoh","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Atsushi","middleName":"","lastName":"Kumanogoh","suffix":""},{"id":624272384,"identity":"c2ef64cf-cef7-4341-8fce-4792d2025a03","order_by":23,"name":"Ken Nakata","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Ken","middleName":"","lastName":"Nakata","suffix":""},{"id":624272385,"identity":"9ee0a363-5f49-49cd-9d7f-a418778c3855","order_by":24,"name":"Seiji Okada","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Seiji","middleName":"","lastName":"Okada","suffix":""},{"id":624272386,"identity":"0c565cf4-361f-4feb-9c65-be6be6b70ab0","order_by":25,"name":"Masatoshi Hori","email":"","orcid":"","institution":"The University of Osaka Graduate School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Masatoshi","middleName":"","lastName":"Hori","suffix":""}],"badges":[],"createdAt":"2026-03-23 08:39:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9197723/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9197723/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107485041,"identity":"889a5418-805b-4f78-b62c-3cb813c64445","added_by":"auto","created_at":"2026-04-22 02:33:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":137698,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy flow diagram. \u003c/strong\u003eSchematic overview of the study design showing the allocation of eight rheumatology facilities into training (six facilities, n = 2,425) and external validation (two facilities, n = 2,460) cohorts, causal forest model development with clinician-guided safety constraints, and concordant versus discordant evaluation. Treatment courses represent unique patient–treatment combinations; the same patient may contribute multiple courses across different treatment lines. Responder was defined as achieving Clinical Disease Activity Index ≤ 10 within 6 months, or treatment continuation \u0026gt;6 months, or discontinuation due to adverse events after \u0026gt;6 months. Non-responder was defined as failure to meet responder criteria. ANSWER, Kansai Consortium for Well-being of Rheumatic Disease Patients; RA, rheumatoid arthritis; bDMARDs, biological disease-modifying anti-rheumatic drugs; JAKi, Janus kinase inhibitor\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9197723/v1/87b2b8f861b324af38469855.png"},{"id":107256460,"identity":"8968a719-ab3f-4521-8f45-05fd5a07e6c5","added_by":"auto","created_at":"2026-04-19 12:17:06","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":173009,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchematic overview of the clinician-guided causal inference framework. \u003c/strong\u003eThe framework operates in three stages: (1) patient baseline data (demographics, disease activity, laboratory values, treatment history) are input to the Causal Forests with Double Machine Learning (CausalForestDML) model, which estimates individualised treatment effects for four biologic/targeted synthetic disease-modifying antirheumatic drug categories (tumour necrosis factor inhibitor (TNFi), Janus kinase inhibitor (JAKi), interleukin-6 inhibitor (IL-6i), cytotoxic T-lymphocyte-associated protein 4 immunoglobulin (CTLA4-Ig)); (2) clinician-defined safety constraints based on Japan College of Rheumatology guidelines exclude JAKi for patients with two or more cardiovascular risk factors (age ≥ 65 years, smoking history, body mass index (BMI) ≥ 30 kg/m²); and (3) the category with the highest estimated benefit among eligible options is selected as the final recommendation. ITE, individualised treatment effect; JCR, Japan College of Rheumatology; CDAI, Clinical Disease Activity Index; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9197723/v1/d739b93de6157285dea269cf.png"},{"id":107484407,"identity":"42a01e9e-7373-4d5f-a63a-3b2c5b6f16dc","added_by":"auto","created_at":"2026-04-22 02:31:54","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":111926,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGroup average treatment effects across conditional average treatment effect quartiles. \u003c/strong\u003eGroup average treatment effects (GATEs) with 95% confidence intervals for Q1-Q4 in the external validation cohort. The increasing pattern from Q1 to Q4 validates the AI system's ability to rank patients by predicted treatment benefit (Q4-Q1 difference = 0.279, Z = 4.64, P \u0026lt; 0.001). Error bars represent 95% confidence intervals estimated by the doubly robust estimator. CATE, conditional average treatment effect.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9197723/v1/24b59e6b494e5e912c339071.png"},{"id":107485147,"identity":"d0f45dea-4b44-4bcc-8fd2-eb83a3a26f04","added_by":"auto","created_at":"2026-04-22 02:33:40","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":352265,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConditional average treatment effect SHAP summary plot for treatment effect heterogeneity. \u003c/strong\u003eTop 15 features ranked by mean absolute SHAP (Shapley Additive Explanations) value derived from the causal forest model, showing which patient characteristics most influence between-patient variation in estimated treatment effects. SHAP values were averaged across three treatment comparisons (Janus kinase inhibitor vs tumour necrosis factor inhibitor [TNFi], interleukin-6 inhibitor vs TNFi, cytotoxic T-lymphocyte-associated protein 4 immunoglobulin vs TNFi). Each dot represents one patient; dot colour represents feature value (blue = low, red = high); horizontal position shows the SHAP value contribution to the conditional average treatment effect. Analysis performed on the external validation cohort (n = 2,460). CATE, conditional average treatment effect; SHAP, Shapley Additive Explanations; CRP, C-reactive protein; VAS, Visual Analogue Scale; MMP-3, matrix metalloproteinase-3; ESR, erythrocyte sedimentation rate; eGFR, estimated glomerular filtration rate; RF, rheumatoid factor; LDH, lactate dehydrogenase; CDAI, Clinical Disease Activity Index; WBC, white blood cell count; ACPA, anti-cyclic citrullinated peptide antibody; Plt, platelet count; SDAI, Simplified Disease Activity Index.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9197723/v1/1dd76870affb23c7abfbaedc.png"},{"id":107487202,"identity":"b715c6ba-5c67-4464-8ef7-5ca138eacf07","added_by":"auto","created_at":"2026-04-22 02:40:03","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1276527,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9197723/v1/7772ae75-76d8-4eb4-b2e5-fdb2364f9a7a.pdf"},{"id":107256459,"identity":"9cef57e4-48dd-4aa9-a404-f4b2b812d097","added_by":"auto","created_at":"2026-04-19 12:17:06","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":149192,"visible":true,"origin":"","legend":"","description":"","filename":"supplementaryinformation20260323.docx","url":"https://assets-eu.researchsquare.com/files/rs-9197723/v1/2d4f4dd8635b63939d3168d4.docx"}],"financialInterests":"Competing interest reported. Y.E. is affiliated with, K.E. holds an additional post in, and K.N. supervises the Department of Sports Medical Biomechanics, The University of Osaka Graduate School of Medicine Faculty of Medicine, which is supported by Asahi Kasei. K.E. has received research grants fromEli Lilly, and speaker fees from AbbVie, Amgen, Argenx, Asahi Kasei, Astellas, Ayumi, Bristol-Myers Squibb, Chugai, Daiichi Sankyo, Eisai, Eli Lilly, Janssen, Mitsubishi Tanabe, Ono Pharmaceutical, Pfizer, Sanofi, Taisho, Teijin Pharma, and UCB Japan. Y.E. has received a speaker fee from Amgen, Asahi-Kasei, Astellas, Chugai, Daiichi-Sankyo, Eisai, Gilead Sciences Japan, Taisho, and UCB Japan. K.T. has received speaker fees from Ayumi, Mikasa, Eisai, and Asahi-Kasei Corporation. A.K. has received a research grant from Chugai and speaker fees from Chugai, AbbVie, and Ono Pharmaceuticals. Ko.M. is affiliated with a department that is financially supported by Asahi-Kasei Pharma Corp. and the city government (Nagahama City and Toyooka City), and has received speaking and/or consulting fees from Asahi Kasei Pharma Corp., Chugai Pharmaceutical Co., Ltd., UCB Japan Co., Ltd., Eli Lilly, AbbVie, Eisai Co., Ltd., and Daiichi Sankyo Co., Ltd. Mo.H. has received research grants and/or speaker fees from Asahi Kasei, Astellas, AstraZeneca, Ayumi Pharma, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences Japan, Janssen Pharma, Ono Pharma, Ohtsuka Pharma, Taisho Pharma, Tanabe Mitsubishi, and UCB Japan. T.O. has received speaker fees and/or research grants from AbbVie, Asahi Kasei, Astellas, Daiichi Sankyo, Eli Lilly, and UCB. H.Y. has received payments for lectures from AbbVie, Asahi Kasei, Astellas, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences, Ono, Otsuka, Pfizer, Taiho, and Taisho. To.T. has received research grants from Boehringer Ingelheim and Chugai, and payments for lectures from Astellas, Chugai, Eisai, Eli Lilly, MSD, Otsuka, and Taisho. Yu.N. has received speaker fees from Astellas Pharma, Asahi Kasei Pharma, Eisai, Eli Lilly, AstraZeneca, and Tanabe Pharma. K.K., Y.S., T.N., A.S., Y.O., Yu.M., T.F., Y.U., H.S., Y.So., Hi.M., D.T., W.Y., S.O., and Ma.H. declare no competing interests.","formattedTitle":"A causal inference framework for personalised b/tsDMARD selection in rheumatoid arthritis: multi-centre development and external validation from the ANSWER cohort","fulltext":[{"header":"Introduction","content":"\u003cp\u003eRheumatoid arthritis (RA) is a chronic autoimmune disease where methotrexate (MTX), a conventional synthetic disease-modifying antirheumatic drug (csDMARD), remains the first-line treatment. However, approximately 20\u0026ndash;30% of patients exhibit an inadequate response to MTX, and some patients are intolerant or have contraindications to the drug \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. For these patients, clinicians face a complex array of therapeutic options, including biological DMARDs (bDMARDs) and targeted synthetic DMARDs (tsDMARDs), each with distinct mechanisms of action\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Current clinical guidelines, such as those by the Japan College of Rheumatology (JCR) and the European Alliance of Associations for Rheumatology (EULAR), list these agents as parallel options but lack evidence-based stratification criteria to predict which agent yields the highest efficacy for an individual patient. Drug selection therefore relies on empirical trials and safety profiles rather than predicted efficacy. This trial-and-error approach results in only 60\u0026ndash;70% of patients achieving low disease activity or remission\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, and delays in effective treatment lead to irreversible joint damage and functional impairment.\u003c/p\u003e \u003cp\u003ePrevious research has explored various predictors of treatment response. Single biomarkers such as rheumatoid factor (RF) and anti-cyclic citrullinated peptide (ACPA) antibodies have limited predictive accuracy for guiding drug selection\u003csup\u003e\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Machine learning (ML) approaches integrating routine clinical data have demonstrated the ability to predict response to bDMARDs with moderate to high accuracy\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Recent studies have identified drug class-specific predictive features: age for tumour necrosis factor (TNF) inhibitors, disease duration for cytotoxic T-lymphocyte antigen 4\u0026ndash;immunoglobulin (CTLA4-Ig), and C-reactive protein (CRP) levels for interleukin-6 (IL-6) inhibitors\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. These findings suggest that routine clinical data contain signals capable of forecasting treatment outcomes.\u003c/p\u003e \u003cp\u003eA critical gap, however, separates prediction from recommendation. Most ML studies predict the probability of response to a single drug, rather than comparing multiple options to determine the optimal choice for a specific patient. Estimating the individualised treatment effect (ITE), defined as the difference in potential outcomes under different treatments, is essential for precision medicine but has been difficult to implement in RA\u003csup\u003e10\u003c/sup\u003e. Causal ML methods, particularly causal forests based on the generalised random forest framework\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, offer a principled approach to estimating heterogeneous treatment effects from observational data.\u003c/p\u003e \u003cp\u003eBeyond algorithmic accuracy, responsible clinical deployment of AI-based treatment recommendation systems requires human oversight\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. AI models trained on observational data may learn patterns that conflict with established safety evidence. Current JCR guidelines recommend caution with Janus kinase (JAK) inhibitors in patients who are elderly, smokers, or who possess risk factors for malignancy, cardiovascular disease, or thromboembolism\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, yet a model trained solely on efficacy outcomes cannot incorporate such safety considerations. A clinician-guided design, where safety constraints derived from clinical guidelines are encoded as formal rules that override data-driven recommendations, ensures that the system operates within accepted clinical boundaries while preserving the AI\u0026rsquo;s ability to optimise within those boundaries.\u003c/p\u003e \u003cp\u003eWe developed and validated a clinician-guided causal inference framework for b/tsDMARD category selection in RA, by using data from the K\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eans\u003c/span\u003eai Consortium for \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eWe\u003c/span\u003ell-being of \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eR\u003c/span\u003eheumatic Disease Patients (ANSWER) cohort, which is a large multicenter, observational RA registry in Japan\u003csup\u003e\u003cspan additionalcitationids=\"CR16 CR17\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. The framework integrates a causal forest\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e with double ML\u003csup\u003e19\u003c/sup\u003e (CausalForestDML) for data-driven ITE estimation with clinician-encoded safety constraints that exclude specific drug categories for patients meeting predefined risk criteria. We retrospectively evaluated clinical utility through concordant versus discordant analysis with multiple confounding adjustment methods, characterised heterogeneous treatment effects, and identified clinically interpretable effect modifiers.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy population\u003c/h2\u003e \u003cp\u003eThe training cohort comprised 2,425 treatment courses from six facilities and the external validation cohort comprised 2,460 treatment courses from two independent facilities (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Baseline characteristics are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Several significant differences were observed between cohorts, including age, disease activity such as Clinical Disease Activity Index (CDAI) and Simplified Disease Activity Index (SDAI), inflammatory markers (ESR), RF, and renal function (estimated Glomerular Filtration Rate; eGFR), Steinbrocker Stage and Class distributions, and concomitant MTX use. Despite these differences, both cohorts had similar sex distribution (approximately 82% female), disease duration, treatment line (median 2), and CDAI response rates (26.5% vs 28.9%, P\u0026thinsp;=\u0026thinsp;0.067). The distribution of prescribed drug categories differed between cohorts (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001), with TNFi being the most frequently prescribed in both (training 40.7%, validation 46.6%; Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Of 258 candidate predictors, 83 (32.2%) had no missing values; missing rates for the remaining predictors ranged from 0.1% to 78.8% (Supplementary Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBaseline characteristics of the training and external validation cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining (n\u0026thinsp;=\u0026thinsp;2,425)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation (n\u0026thinsp;=\u0026thinsp;2,460)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e61.4\u0026thinsp;\u0026plusmn;\u0026thinsp;15.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e60.2\u0026thinsp;\u0026plusmn;\u0026thinsp;15.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale sex, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,962 (80.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,033 (82.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.125\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBMI, kg/m\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21.8\u0026thinsp;\u0026plusmn;\u0026thinsp;4.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22.2\u0026thinsp;\u0026plusmn;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.086\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDisease duration, months\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e91.4 [27.5\u0026ndash;194.8]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e102.8 [33.8\u0026ndash;206.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.009\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTreatment Line\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.0 [1.0\u0026ndash;3.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.0 [1.0\u0026ndash;3.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.008\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSteinbrocker Stage, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e710 (33.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e433 (29.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e479 (22.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e406 (27.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e389 (18.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e232 (15.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e551 (25.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e404 (27.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSteinbrocker Class, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e735 (34.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e366 (24.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e907 (42.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e853 (56.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e437 (20.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e276 (18.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e57 (2.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e27 (1.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCDAI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17.9\u0026thinsp;\u0026plusmn;\u0026thinsp;11.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15.3\u0026thinsp;\u0026plusmn;\u0026thinsp;9.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSDAI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e19.5\u0026thinsp;\u0026plusmn;\u0026thinsp;12.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17.0\u0026thinsp;\u0026plusmn;\u0026thinsp;10.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDAS28-CRP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.7\u0026thinsp;\u0026plusmn;\u0026thinsp;1.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.6\u0026thinsp;\u0026plusmn;\u0026thinsp;1.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.465\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatient VAS, cm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.9\u0026thinsp;\u0026plusmn;\u0026thinsp;2.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.0\u0026thinsp;\u0026plusmn;\u0026thinsp;2.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.294\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCRP, mg/dL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.6\u0026thinsp;\u0026plusmn;\u0026thinsp;2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.5\u0026thinsp;\u0026plusmn;\u0026thinsp;2.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.060\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eESR, mm/h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e35.4\u0026thinsp;\u0026plusmn;\u0026thinsp;28.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e39.8\u0026thinsp;\u0026plusmn;\u0026thinsp;29.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF, U/mL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e70.0 [20.0\u0026ndash;191.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e53.2 [14.0\u0026ndash;156.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF positive (\u0026gt;\u0026thinsp;15 U/mL), n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,830 (75.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,781 (72.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.016\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eACPA, U/mL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e51.1 [6.7\u0026ndash;225.2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49.2 [2.7\u0026ndash;280.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.492\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eACPA positive (\u0026gt;\u0026thinsp;4.5 U/mL), n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,663 (68.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,654 (67.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.330\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMMP-3, ng/mL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e125.2 [66.4\u0026ndash;259.2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e117.8 [61.1\u0026ndash;247.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.038\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHaemoglobin, g/dL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.2\u0026thinsp;\u0026plusmn;\u0026thinsp;1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.1\u0026thinsp;\u0026plusmn;\u0026thinsp;1.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.872\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eeGFR, mL/min/1.73m\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e74.5\u0026thinsp;\u0026plusmn;\u0026thinsp;24.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76.4\u0026thinsp;\u0026plusmn;\u0026thinsp;24.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConcomitant MTX use, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1,216 (50.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,355 (55.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDose among users, mg/week\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10.1\u0026thinsp;\u0026plusmn;\u0026thinsp;5.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.0\u0026thinsp;\u0026plusmn;\u0026thinsp;3.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConcomitant PSL use, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e845 (34.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e836 (34.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.546\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDose among users, mg/day\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.3\u0026thinsp;\u0026plusmn;\u0026thinsp;4.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.4\u0026thinsp;\u0026plusmn;\u0026thinsp;10.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponder, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e643 (26.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e711 (28.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.067\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrescribed drug category\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTNFi, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e986 (40.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1,147 (46.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIL-6i, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e542 (22.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e528 (21.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJAKi, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e520 (21.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e410 (16.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCTLA4-Ig, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e377 (15.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e375 (15.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003e\u003cem\u003eData are mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD, median [IQR], or n (%). CDAI\u0026thinsp;=\u0026thinsp;Clinical Disease Activity Index. SDAI\u0026thinsp;=\u0026thinsp;Simplified Disease Activity Index. DAS28-CRP\u0026thinsp;=\u0026thinsp;Disease Activity Score in 28 joints using CRP. VAS\u0026thinsp;=\u0026thinsp;Visual Analogue Scale. CRP\u0026thinsp;=\u0026thinsp;C-reactive protein. ESR\u0026thinsp;=\u0026thinsp;erythrocyte sedimentation rate. RF\u0026thinsp;=\u0026thinsp;rheumatoid factor. ACPA\u0026thinsp;=\u0026thinsp;anti-cyclic citrullinated peptide antibody. MMP-3\u0026thinsp;=\u0026thinsp;matrix metalloproteinase-3. eGFR\u0026thinsp;=\u0026thinsp;estimated glomerular filtration rate. MTX\u0026thinsp;=\u0026thinsp;methotrexate. PSL\u0026thinsp;=\u0026thinsp;prednisolone. TNFi\u0026thinsp;=\u0026thinsp;tumour necrosis factor inhibitor. IL-6i\u0026thinsp;=\u0026thinsp;interleukin-6 inhibitor. JAKi\u0026thinsp;=\u0026thinsp;Janus kinase inhibitor. CTLA4-Ig\u0026thinsp;=\u0026thinsp;cytotoxic T-lymphocyte-associated protein 4 immunoglobulin (abatacept). P values were calculated using Mann\u0026ndash;Whitney U test for continuous variables and chi-square test for categorical variables.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCausal forest with double machine learning and nuisance model selection\u003c/h3\u003e\n\u003cp\u003eWe developed a CausalForestDML model, which combines Causal Forests with Double Machine Learning. This framework estimates individualised treatment effects while adjusting for confounding through auxiliary propensity and outcome models. Among five candidate models compared on the training cohort (Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e), Random Forest achieved the narrowest 95% confidence interval width for the doubly robust estimator (0.072), indicating the most stable treatment effect estimation. Random Forest was therefore selected for both the propensity and outcome models.\u003c/p\u003e\n\u003ch3\u003eAI recommendation distribution and safety constraint effects\u003c/h3\u003e\n\u003cp\u003eThe clinician-guided causal inference framework operates in three sequential stages (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e): (1) patient baseline data are input to the CausalForestDML model, which estimates ITEs for four drug categories (TNFi, JAKi, IL-6i, CTLA4-Ig); (2) clinician-defined safety constraint exclude JAKi for patients meeting two or more cardiovascular risk factors (age\u0026thinsp;\u0026ge;\u0026thinsp;65, smoking history, body mass index [BMI]\u0026thinsp;\u0026ge;\u0026thinsp;30), in accordance with the JCR guidelines\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e; and (3) the category with the highest estimated ITE among eligible options is selected as the final recommendation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe trained model was frozen and applied to the external validation cohort. When JAKi was excluded, the category with the next-highest ITE was recommended.\u003c/p\u003e \u003cp\u003eAfter applying the safety constraint, the model recommended JAKi for 53.6% (n\u0026thinsp;=\u0026thinsp;1,318), TNFi for 28.9% (n\u0026thinsp;=\u0026thinsp;711), IL-6i for 17.4% (n\u0026thinsp;=\u0026thinsp;427), and CTLA4-Ig for 0.2% (n\u0026thinsp;=\u0026thinsp;4) of external validation patients. Safety constraints excluded JAKi for 241 patients (9.8%), redirecting their recommendations to alternative categories.\u003c/p\u003e \u003cp\u003eTreatment recommendations showed significant age-dependent heterogeneity (chi-square P\u0026thinsp;\u0026lt;\u0026thinsp;0.001): elderly-onset RA (EORA) patients (age of onset\u0026thinsp;\u0026ge;\u0026thinsp;60, n\u0026thinsp;=\u0026thinsp;542) received JAKi recommendations at 64.6% with TNFi at 14.9% and IL-6i at 20.5%, while young-onset RA patients (n\u0026thinsp;=\u0026thinsp;1,477) had JAKi at 48.3%, TNFi at 34.7%, and IL-6i at 16.8% (Supplementary Fig. S2a). Among EORA patients, 122 (22.5%) had JAKi excluded by safety constraints, increasing IL-6i and TNFi recommendations in this group (Supplementary Fig. S2b).\u003c/p\u003e\n\u003ch3\u003eConcordant versus discordant analysis\u003c/h3\u003e\n\u003cp\u003ePatients receiving treatments concordant with AI recommendations had significantly higher CDAI response rates across all three analytical approaches (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). In crude analysis (n\u0026thinsp;=\u0026thinsp;664 concordant, n\u0026thinsp;=\u0026thinsp;1,796 discordant), response rates were 33.6% vs 27.2% (ARD\u0026thinsp;=\u0026thinsp;6.4%, relative risk [RR]\u0026thinsp;=\u0026thinsp;1.24, odds ratio [OR]\u0026thinsp;=\u0026thinsp;1.36, P\u0026thinsp;=\u0026thinsp;0.002; NNT\u0026thinsp;=\u0026thinsp;16). After propensity score matching (PSM; n\u0026thinsp;=\u0026thinsp;660 matched pairs), response rates were 33.6% vs 28.2% (ARD\u0026thinsp;=\u0026thinsp;5.5%, P\u0026thinsp;=\u0026thinsp;0.037; NNT\u0026thinsp;=\u0026thinsp;18). Inverse probability of treatment weighting (IPTW) yielded weighted response rates of 31.6% vs 27.3% (ARD\u0026thinsp;=\u0026thinsp;4.3%, P\u0026thinsp;=\u0026thinsp;0.044; NNT\u0026thinsp;=\u0026thinsp;23).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eConcordant versus discordant analysis in the external validation cohort\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnalysis\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConcordant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDiscordant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eARD\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNNT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eE-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCrude (n\u0026thinsp;=\u0026thinsp;664 vs 1,796)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e33.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e27.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e1.60\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePSM (n\u0026thinsp;=\u0026thinsp;660 matched pairs)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e33.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e28.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.037\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e1.53\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIPTW\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e31.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e27.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.044\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e1.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"9\"\u003e\u003cem\u003eARD\u0026thinsp;=\u0026thinsp;absolute risk difference. RR\u0026thinsp;=\u0026thinsp;relative risk. OR\u0026thinsp;=\u0026thinsp;odds ratio. NNT\u0026thinsp;=\u0026thinsp;number needed to treat. PSM\u0026thinsp;=\u0026thinsp;propensity score matching. IPTW\u0026thinsp;=\u0026thinsp;inverse probability of treatment weighting. E-value\u0026thinsp;=\u0026thinsp;E-value for the point estimate. \u0026mdash; indicates RR and OR not calculated; PSM used McNemar\u0026rsquo;s test for matched pairs and IPTW used bootstrap inference.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe attenuation from crude to IPTW estimates (ARD 6.4% to 4.3%) reflects increasingly conservative confounding adjustment, with all three methods maintaining statistical significance.\u003c/p\u003e \u003cp\u003eE-values for unmeasured confounding were 1.60 (crude), 1.53 (PSM), and 1.46 (IPTW) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eSubgroup and sensitivity analyses\u003c/h3\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eDisease duration subgroup analysis\u003c/h2\u003e \u003cp\u003eConcordant versus discordant analysis stratified by disease duration revealed stage-dependent benefit (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Established RA (2\u0026ndash;10 years) demonstrated the strongest effect (ARD\u0026thinsp;=\u0026thinsp;13.8%, NNT\u0026thinsp;=\u0026thinsp;7.2, P\u0026thinsp;=\u0026thinsp;0.0005), while long-standing RA (\u0026gt;\u0026thinsp;10 years) showed no benefit (ARD\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.2%, P\u0026thinsp;=\u0026thinsp;1.000).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSubgroup analyses by disease duration and treatment line\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSubgroup\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003en Concordant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003en Discordant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eConcordant (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDiscordant (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eARD\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNNT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDisease duration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEarly RA (\u0026lt;\u0026thinsp;2 years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e110\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e286\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e49.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e39.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e9.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e10.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.093\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEstablished RA (2\u0026ndash;10 years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e203\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e510\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e41.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e28.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e13.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e7.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.0005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLong-standing RA (\u0026gt;\u0026thinsp;10 years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e233\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e673\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e23.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e23.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026minus;0.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTreatment line\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1st line (biologic-naive)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e262\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e40.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e34.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e16.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.086\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2nd line\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e171\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e408\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e39.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e23.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e16.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e6.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.0002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3rd+ line\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e231\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e593\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e21.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e20.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.765\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cem\u003eARD\u0026thinsp;=\u0026thinsp;absolute risk difference. NNT\u0026thinsp;=\u0026thinsp;number needed to treat. Crude analysis results shown. \u0026mdash; indicates NNT not calculated due to non-significant or negligible ARD.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eTreatment line stratification\u003c/h3\u003e\n\u003cp\u003eAnalysis stratified by treatment history demonstrated the strongest AI recommendation benefit in second-line treatment (ARD\u0026thinsp;=\u0026thinsp;16.0%, NNT\u0026thinsp;=\u0026thinsp;6.3, P\u0026thinsp;=\u0026thinsp;0.0002), corresponding to a NNT approximately 2.5 times lower than the overall cohort (NNT\u0026thinsp;=\u0026thinsp;16). This represented the strongest benefit observed across all subgroups (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eHeterogeneous treatment effects\u003c/h3\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eGroup average treatment effects\u003c/h2\u003e \u003cp\u003eGroup average treatment effects (GATEs) analysis demonstrated an increasing pattern across conditional average treatment effect (CATE) quartiles: Q1 (GATE\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.116, SE\u0026thinsp;=\u0026thinsp;0.042), Q2 (\u0026minus;\u0026thinsp;0.034, SE\u0026thinsp;=\u0026thinsp;0.044), Q3 (0.002, SE\u0026thinsp;=\u0026thinsp;0.044), and Q4 (0.163, SE\u0026thinsp;=\u0026thinsp;0.043). The Q4\u0026ndash;Q1 difference was significant (difference\u0026thinsp;=\u0026thinsp;0.279, Z\u0026thinsp;=\u0026thinsp;4.64, P\u0026thinsp;\u0026lt;\u0026thinsp;0.001; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), validating the AI system\u0026rsquo;s ability to rank patients by predicted treatment benefit. Category-specific GATEs analyses confirmed significant heterogeneity for JAKi (Q4\u0026ndash;Q1: Z\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;2.59, P\u0026thinsp;=\u0026thinsp;0.005) and CTLA4-Ig (Z\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;3.63, P\u0026thinsp;\u0026lt;\u0026thinsp;0.001; Supplementary Table S2).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eEffect modifiers\u003c/h2\u003e \u003cp\u003eTwelve covariates were identified as robust effect modifiers by both Best Linear Projection (BLP) and Lasso methods (Supplementary Table S3). These grouped into clinically meaningful categories: disease activity markers (Disease Activity Score using 28 joints with C-Reactive Protein [DAS28-CRP], CRP, ESR, Patient Visual Analog Scale [VAS]), haematological parameters (White Blood Cell [WBC], haemoglobin, lymphocyte), demographics (age), structural damage (Matrix Metalloproteinase-3 [MMP-3], Steinbrocker Stage), and treatment history (past golimumab use, tacrolimus use). The convergence of a tree-based approach (BLP) and a linear regularised approach (Lasso) provide model-independent evidence for genuine treatment effect heterogeneity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eCATE-SHAP analysis\u003c/h2\u003e \u003cp\u003eCATE-SHAP analysis identified CRP as the most influential driver of between-patient variation in estimated treatment effects, followed by Patient VAS, haemoglobin, MMP-3, and ESR. Renal function (eGFR), serological markers (RF, ACPA), and demographic factors (age) also contributed to treatment effect heterogeneity (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). These features substantially overlap with the 12 robust effect modifiers identified independently by BLP and Lasso, providing convergent evidence from three distinct analytical approaches.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study developed and externally validated a clinician-guided causal inference framework for b/tsDMARD category selection in RA. The framework combines CausalForestDML for data-driven treatment effect estimation with clinician-encoded safety constraints that exclude JAKi for patients with multiple specific risk factors. Four principal findings emerged: (1) AI-concordant treatment was associated with significantly higher CDAI response rates across unadjusted, PSM, and IPTW analyses; (2) this effect was maintained despite significant between-cohort heterogeneity in baseline characteristics; (3) the system identified 12 clinically interpretable covariates driving patient-level differences in treatment benefit; and (4) the safety constraint layer appropriately modified 9.8% of recommendations, with the resulting age-dependent patterns consistent with current safety guidelines.\u003c/p\u003e\n\u003cp\u003eThe concordant versus discordant analysis provides direct evidence of clinical utility. The NNT of 16 indicates that for every 16 patients treated according to AI recommendations, one additional patient would achieve CDAI response, with the benefit remaining significant across all three confounding adjustment methods. E-value analysis indicated that an unmeasured confounder would need to be associated with both concordant treatment selection and CDAI response by a risk ratio of at least 1.60 to fully explain the observed effect.\u003c/p\u003e\n\u003cp\u003eThe monotonically increasing GATEs pattern validates the core premise of personalised treatment recommendation: the model meaningfully ranks patients by predicted treatment benefit. The 12 robust effect modifiers grouped into clinically coherent categories (disease activity, haematological parameters, demographics, and structural damage), and CATE-SHAP analysis confirmed substantial overlap with these modifiers. This convergence across three methodologically distinct approaches \u0026mdash; a tree-based method (BLP), a linear regularised method (Lasso), and a model-explanation method (SHAP) \u0026mdash; strengthens confidence that the model captures genuine biological heterogeneity rather than spurious correlations.\u003c/p\u003e\n\u003cp\u003eSubgroup analyses further clarified the clinical contexts in which personalised recommendations confer the greatest benefit. Disease duration subgroup analysis revealed that AI recommendations provided the greatest benefit in established RA (2\u0026ndash;10 years; ARD\u0026thinsp;=\u0026thinsp;13.8%, NNT\u0026thinsp;=\u0026thinsp;7.2, P\u0026thinsp;=\u0026thinsp;0.0005), with a trend in early RA (\u0026lt;\u0026thinsp;2 years; ARD\u0026thinsp;=\u0026thinsp;9.9%, NNT\u0026thinsp;=\u0026thinsp;10, P\u0026thinsp;=\u0026thinsp;0.093) but no benefit in long-standing RA (\u0026gt;\u0026thinsp;10 years; ARD\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.2%, P\u0026thinsp;=\u0026thinsp;1.000). This pattern likely reflects that in established RA, sufficient clinical heterogeneity has emerged to enable treatment differentiation, yet disease remains modifiable, whereas in long-standing RA, treatment-refractory inflammation and structural damage may limit model discriminative capacity. Treatment line-stratified analysis revealed the strongest benefit in second-line treatment (ARD\u0026thinsp;=\u0026thinsp;16.0%, NNT\u0026thinsp;=\u0026thinsp;6.3, P\u0026thinsp;=\u0026thinsp;0.0002), with diminishing benefit in third-line and beyond (ARD\u0026thinsp;=\u0026thinsp;1.2%, P\u0026thinsp;=\u0026thinsp;0.765), suggesting that AI recommendations provide maximal value when treatment options remain relatively abundant.\u003c/p\u003e\n\u003cp\u003eThese findings extend previous ML studies in RA, which have focused on predicting response to specific drugs with AUCs of 0.60\u0026ndash;0.91\u003csup\u003e20,21\u003c/sup\u003e but cannot directly compare expected outcomes across treatment options for individual patients. To our knowledge, this is the first study to apply a causal inference framework (CausalForestDML) combined with clinician-guided safety constraints to b/tsDMARD selection in RA with multi-centre external validation. The concordant versus discordant design provides a more rigorous validation framework than internal validation commonly reported in the literature\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, and the system maintained significant ARD despite substantial between-cohort heterogeneity in disease activity, inflammatory markers, structural damage (Steinbrocker Stage and Class), and demographics. Notably, the proportion of JAKi recommendations (53.6%) contrasts with current prescribing patterns where TNFi remains the most commonly prescribed first-line biologic DMARD\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, which reflects the model\u0026rsquo;s focus on optimising treatment outcomes rather than reproducing existing practice.\u003c/p\u003e\n\u003cp\u003eThe integration of clinician-defined safety constraints addresses this discrepancy by ensuring that efficacy-driven recommendations remain within accepted clinical boundaries. The JCR guidelines recommend caution with JAK inhibitors in patients with specific risk factors\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, and encoding this guideline as formal rules within the recommendation pipeline ensures that safety considerations, which the CausalForestDML model cannot learn from efficacy-outcome training data alone, are incorporated into final recommendations. In this cohort, 241 patients (9.8%) had JAKi excluded by safety constraints, and 122 of 542 EORA patients (22.5%) had recommendations redirected to alternative categories. This constraint layer operates transparently: the clinician can review both the unconstrained and constrained recommendations, understand which rules were triggered, and retain full authority over the final prescribing decision. This design aligns with growing consensus that clinical AI systems should augment rather than replace physician judgement\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, and that safety-critical domains require explicit human oversight mechanisms.\u003c/p\u003e\n\u003cp\u003eThis study has several limitations. First, the retrospective observational design precludes causal conclusions; although PSM, IPTW, and E-value sensitivity analyses were employed, residual unmeasured confounding cannot be excluded. Second, the study population was derived from Japanese rheumatology registries; treatment patterns, drug availability, and genetic backgrounds may differ in other populations. Therefore, evaluation in non-Japanese populations remains necessary to assess cross-ethnic applicability. Third, the safety constraints are currently limited to JAKi specific risk factors, and due to the nature of the cohort, the risk factors of malignancy could not be sufficiently evaluated.; expansion to other drug-specific safety considerations (such as infection risk stratification for bDMARDs) would improve the framework\u0026rsquo;s clinical applicability. Fourth, the near-absence of CTLA4-Ig recommendations (0.2%) likely reflects confounding by indication rather than true therapeutic inferiority. In our cohort, patients prescribed CTLA4-Ig (abatacept) had the highest mean age among all drug categories (67.7 years), suggesting that clinicians preferentially selected this agent for elderly or frail patients in whom safety considerations outweighed efficacy optimisation. These patients \u0026mdash; characterised by advanced age, frailty, and elevated infection risk that precluded JAKi or TNFi \u0026mdash; represent a population in whom CDAI improvement is inherently more difficult to achieve. While DML adjusts for observed confounders, unmeasured factors such as frailty severity, infection risk burden, and clinician safety concerns cannot be fully controlled. Consequently, the model may have attributed the poorer outcomes observed in CTLA4-Ig-treated patients to the drug category itself rather than to the underlying patient characteristics that drove its selection. This residual confounding by indication represents a fundamental limitation of observational treatment effect estimation and should be considered when interpreting the low CTLA4-Ig recommendation rate.\u003c/p\u003e\n\u003cp\u003eTo facilitate clinical access and independent evaluation, we have developed a publicly available web-based tool that enables clinicians to input patient characteristics and obtain individualised treatment recommendations with the safety constraint layer applied. Future directions include prospective randomised validation, expansion of safety constraints to additional drug classes, evaluation in non-Japanese populations.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec16\" class=\"Section3\"\u003e\n \u003ch2\u003eStudy design and data source\u003c/h2\u003e\n \u003cp\u003eThis retrospective multi-centre cohort study analysed data from the ANSWER (Kansai Consortium for Well-being of Rheumatic Disease Patients) cohort, a large-scale registry of patients with RA in the Kansai region of Japan comprising nine university-related hospitals\u003csup\u003e\u003cspan additionalcitationids=\"CR16 CR17\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Data were available from eight of these hospitals; Nara Medical University did not contribute data during the study period and was excluded from the analysis. The study included 4,885 treatment courses from these facilities. Analysis was restricted to treatment courses with evaluable CDAI outcomes (responder or non-responder). The study was conducted in accordance with the Declaration of Helsinki and is reported in accordance with the TRIPOD\u0026thinsp;+\u0026thinsp;AI guidelines\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n \u003cp\u003eThe two largest facilities were designated as the external validation cohort (combined n\u0026thinsp;=\u0026thinsp;2,460), while the remaining six facilities served as the training cohort (n\u0026thinsp;=\u0026thinsp;2,425), ensuring adequate statistical power for detecting clinically meaningful differences while maintaining sufficient training data for model development.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n \u003ch2\u003eParticipants and eligibility criteria\u003c/h2\u003e\n \u003cp\u003eThe study period encompassed treatment courses initiated between January 2000 and June 2024. Eligible participants were patients diagnosed with RA according to either the 1987 American College of Rheumatology (ACR) classification criteria\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e or the 2010 ACR/EULAR classification criteria\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, who were registered in the ANSWER cohort and initiated treatment with a b/tsDMARD (TNFi, IL-6i, JAKi, or CTLA4-Ig). The administration of these agents was at the discretion of the attending rheumatologists, consistent with the JCR guidelines\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. Treatment courses were excluded if CDAI outcome data were unavailable for determining the outcome.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003eOutcome\u003c/h2\u003e\n \u003cp\u003eThe outcome was CDAI-based treatment response, defined as a binary variable (1\u0026thinsp;=\u0026thinsp;Responder, 0\u0026thinsp;=\u0026thinsp;Non-Responder). The Responder group included patients meeting either: (1) achieving LDA (CDAI\u0026thinsp;\u0026le;\u0026thinsp;10) or remission within six months after treatment initiation; or (2) maintaining treatment for more than six months with clinical benefit. The Non-Responder group consisted of patients who failed to meet these criteria. CDAI was calculated as the sum of tender joint count (28 joints), swollen joint count (28 joints), Patient Global Assessment (VAS, 0\u0026ndash;10 cm), and Evaluator Global Assessment (VAS, 0\u0026ndash;10 cm). Outcome assessment was performed by treating rheumatologists at each facility using standardised joint count and VAS procedures. Outcome assessors were not blinded to treatment allocation, consistent with routine clinical practice in observational registry data.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\n \u003ch2\u003ePredictors\u003c/h2\u003e\n \u003cp\u003eAll candidate predictors were extracted from the registry without pre-selection. These included demographic factors (age, sex, BMI), clinical characteristics (disease duration, smoking history, Steinbrocker stage/class), disease activity scores (CDAI, SDAI, DAS28-CRP, Patient VAS), laboratory values (RF, ACPA, CRP, ESR, MMP-3, WBC, haemoglobin, platelet count, eGFR, Krebs von den Lungen [KL]-6), treatment history (treatment line, past usage of individual b/tsDMARDs), and current medication details (MTX dose, prednisolone dose), yielding 222 candidate features. All predictors were measured at or before the time of b/tsDMARD initiation. Continuous predictors were used without transformation. Categorical variables were encoded as binary indicators.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\n \u003ch2\u003eMissing data handling\u003c/h2\u003e\n \u003cp\u003eMissing values were handled using median imputation\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e, applied separately to training and validation datasets to prevent data leakage. Imputation procedures were applied identically across all subgroups.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\n \u003ch2\u003eCausal forest with double machine learning\u003c/h2\u003e\n \u003cp\u003eCausalForestDML integrates Causal Forests\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e with Double Machine Learning (DML)\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, a framework that uses ML models to estimate confounding components (nuisance functions) while preserving valid statistical inference for treatment effects.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\n \u003ch2\u003eNuisance model selection\u003c/h2\u003e\n \u003cp\u003eFive ML models (Random Forest, CatBoost, Logistic Regression, HistGradientBoosting, and LightGBM) were compared using default hyperparameters, prioritising estimation stability measured by the width of the 95% confidence interval for the doubly robust estimator. Selection was based on internal validation using training data only.\u003c/p\u003e\n \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\n \u003ch2\u003eModel configuration\u003c/h2\u003e\n \u003cp\u003eRandom Forest was selected for both the propensity and outcome models. The CausalForestDML system (EconML: a python package) comprised a propensity model estimating drug category assignment probability given patient characteristics, an outcome model predicting treatment response, and a causal forest with 500 trees and 5-fold cross-fitting for nuisance estimation. Bootstrap with 5,000 iterations was used for confidence interval estimation.\u003c/p\u003e\n \u003cp\u003eCausalForestDML estimates individual treatment effects by solving local linear moment equations at each tree node. The doubly robust estimator combines outcome model predictions with inverse propensity weighting for robust policy evaluation.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\n \u003ch2\u003eClass imbalance\u003c/h2\u003e\n \u003cp\u003eThe treatment outcome was imbalanced (responders\u0026thinsp;~\u0026thinsp;27%). Class imbalance was addressed through inverse class frequency weighting in both the propensity and outcome Random Forest models. No resampling techniques were applied.\u003c/p\u003e\n \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\n \u003ch2\u003eClinician-guided safety constraints\u003c/h2\u003e\n \u003cp\u003eTreatment recommendations were subject to clinician-defined safety constraints before concordant versus discordant evaluation. This clinician-guided layer was designed to integrate established post-marketing safety evidence that cannot be learned from efficacy-outcome training data alone.\u003c/p\u003e\n \u003cp\u003eJAKi were excluded from the recommended category when patients met two or more of the following cardiovascular risk factors, in accordance with the JCR guidelines\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e: (1) age\u0026thinsp;\u0026ge;\u0026thinsp;65 years, (2) past or current smoking history, and (3) BMI\u0026thinsp;\u0026ge;\u0026thinsp;30 kg/m\u0026sup2;. When JAKi was excluded, the category with the next-highest estimated ITE was selected.\u003c/p\u003e\n \u003cp\u003eThe safety constraints operate as transparent, auditable rules: the treating clinician can review both the unconstrained and constrained recommendations, understand which rules were triggered, and exercise final prescribing authority. This framework is designed to be extensible; additional safety rules (for example, infection risk stratification for specific bDMARDs) can be incorporated as supporting evidence emerges.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\n \u003ch2\u003eValidation protocol\u003c/h2\u003e\n \u003cp\u003eAll model training and model selection were performed solely on the training dataset. The finalised model was frozen and applied to the external validation cohort for one-time evaluation. External data were not used to update model weights or any model development processes, ensuring complete independence of the external validation.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\n \u003ch2\u003eModel output\u003c/h2\u003e\n \u003cp\u003eThe CausalForestDML system outputs estimated ITEs for each drug category relative to a reference category. The recommended drug category is the one with the highest estimated ITE after applying safety constraints. The model does not output calibrated probability estimates for individual response; rather, it outputs the treatment category expected to yield the greatest relative benefit. No risk group thresholds were applied.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e\n \u003ch2\u003eFairness assessment\u003c/h2\u003e\n \u003cp\u003eModel performance was evaluated across age subgroups (EORA vs young-onset RA). Given the predominantly Japanese study population, race-based fairness assessment was not applicable. Age-based fairness was addressed through the JAKi safety constraints and subgroup analyses.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e\n \u003ch2\u003eConcordant versus discordant analysis\u003c/h2\u003e\n \u003cp\u003eTo evaluate clinical utility, we compared outcomes between patients whose prescribed drug category matched the AI-recommended category (Concordant group) and those with a mismatch (Discordant group). Three analytical approaches assessed robustness to confounding: (1) crude analysis using the chi-square test; (2) propensity score matching with logistic regression using 15 covariates (age, sex, BMI, CDAI, SDAI, CRP, ESR, RF, Treatment Line, Steinbrocker Stage, Steinbrocker Class, eGFR, disease duration, haemoglobin, Patient VAS), 1:1 nearest-neighbour matching with a caliper of 0.2 \u0026times; SD of the propensity score; and (3) IPTW using stabilised inverse probability weights with bootstrap inference (1,000 resamples).\u003c/p\u003e\n \u003cp\u003eEffect measures included ARD, RR, OR, and NNT (=\u0026thinsp;1/ARD).\u003c/p\u003e\n \u003ch2\u003eHeterogeneous treatment effect analysis\u003c/h2\u003e\n \u003cp\u003ePatients were stratified into four groups (Q1\u0026ndash;Q4) based on estimated CATEs. GATEs within each quartile were estimated using the doubly robust estimator.\u003c/p\u003e\n \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e\n \u003ch2\u003eSHAP analysis\u003c/h2\u003e\n \u003cp\u003eSHAP analysis was applied to the CausalForestDML model to identify patient characteristics driving heterogeneity in CATEs. SHAP values were computed separately for each treatment comparison (JAKi vs TNFi, IL-6i vs TNFi, CTLA4-Ig vs TNFi) using TreeExplainer applied to the causal forest\u0026rsquo;s internal tree structure. To obtain an aggregate measure of feature importance, SHAP values were averaged across the three pairwise comparisons and the mean absolute SHAP value per feature was used for ranking. Analysis was performed on the external validation cohort (n\u0026thinsp;=\u0026thinsp;2,460).\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e\n \u003ch2\u003eSubgroup analyses\u003c/h2\u003e\n \u003cp\u003eDisease duration subgroups (early RA [\u0026lt;\u0026thinsp;2 years], established RA [2\u0026ndash;10 years], and long-standing RA [\u0026gt;\u0026thinsp;10 years]) and treatment line strata (first line, second line, third+ line) were analysed using the same three concordant versus discordant approaches. Treatment recommendation distributions were compared between EORA and young-onset RA using the chi-square test.\u003c/p\u003e\n \u003cp\u003eE-values were calculated for all concordant versus discordant effect measures to quantify robustness to unmeasured confounding\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n \u003cdiv id=\"Sec33\" class=\"Section3\"\u003e\n \u003ch2\u003eStatistics and reproducibility\u003c/h2\u003e\n \u003cp\u003eAll statistical tests were two-tailed with \u0026alpha;\u0026thinsp;=\u0026thinsp;0.05 unless otherwise specified. The chi-square test was used for categorical comparisons. The Mann\u0026ndash;Whitney U test was used for continuous variables that did not meet normality assumptions. PSM significance was assessed by McNemar\u0026rsquo;s test. IPTW significance was assessed by bootstrap inference (1,000 resamples). GATEs analyses used the doubly robust Z-test. Multiple testing was not adjusted for in subgroup analyses, which were pre-specified and exploratory. Sample sizes were determined by the available registry data; no formal power calculation was performed a priori. All analyses were conducted using Python 3.10 with econml, scikit-learn, and shap libraries. Results are reproducible from the code repository.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec34\" class=\"Section3\"\u003e\n \u003ch2\u003eEthics statement\u003c/h2\u003e\n \u003cp\u003eThe study protocol was reviewed and approved by the Ethics Committee of Osaka Metropolitan University, serving as the central institutional review board (approved on 9 June 2021; approval no. 2021-074). All participating institutions were included under this central approval and obtained site-specific authorization for study implementation in accordance with local regulatory requirements. The study was conducted in accordance with the Declaration of Helsinki. At each participating institution, informed consent was obtained either through comprehensive consent approved by the respective institutional ethics committee or through an opt-out procedure, as appropriate.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"Declarations","content":"\u003ch3\u003eData availability\u003c/h3\u003e\n\u003cp\u003eThe datasets collected and analysed during the current study are not publicly available due to patient privacy and ethical restrictions imposed by the institutional review boards but are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003ch3\u003eCode availability\u003c/h3\u003e\n\u003cp\u003eThe source code developed for the analysis and the machine learning model in this study is publicly available on GitHub at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/kosukekita/RA_AI_npjDM\u003c/span\u003e\u003c/span\u003e. The web-based implementation of the recommendation system is accessible at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ra-frontend-noauth-1099185790534.us-central1.run.app/predict\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003ch2\u003eCompeting Interests\u003c/h2\u003e\u003cp\u003eY.E. is affiliated with, K.E. holds an additional post in, and K.N. supervises the Department of Sports Medical Biomechanics, The University of Osaka Graduate School of Medicine Faculty of Medicine, which is supported by Asahi Kasei. K.E. has received research grants fromEli Lilly, and speaker fees from AbbVie, Amgen, Argenx, Asahi Kasei, Astellas, Ayumi, Bristol-Myers Squibb, Chugai, Daiichi Sankyo, Eisai, Eli Lilly, Janssen, Mitsubishi Tanabe, Ono Pharmaceutical, Pfizer, Sanofi, Taisho, Teijin Pharma, and UCB Japan. Y.E. has received a speaker fee from Amgen, Asahi-Kasei, Astellas, Chugai, Daiichi-Sankyo, Eisai, Gilead Sciences Japan, Taisho, and UCB Japan. K.T. has received speaker fees from Ayumi, Mikasa, Eisai, and Asahi-Kasei Corporation. A.K. has received a research grant from Chugai and speaker fees from Chugai, AbbVie, and Ono Pharmaceuticals. Ko.M. is affiliated with a department that is financially supported by Asahi-Kasei Pharma Corp. and the city government (Nagahama City and Toyooka City), and has received speaking and/or consulting fees from Asahi Kasei Pharma Corp., Chugai Pharmaceutical Co., Ltd., UCB Japan Co., Ltd., Eli Lilly, AbbVie, Eisai Co., Ltd., and Daiichi Sankyo Co., Ltd. Mo.H. has received research grants and/or speaker fees from Asahi Kasei, Astellas, AstraZeneca, Ayumi Pharma, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences Japan, Janssen Pharma, Ono Pharma, Ohtsuka Pharma, Taisho Pharma, Tanabe Mitsubishi, and UCB Japan. T.O. has received speaker fees and/or research grants from AbbVie, Asahi Kasei, Astellas, Daiichi Sankyo, Eli Lilly, and UCB. H.Y. has received payments for lectures from AbbVie, Asahi Kasei, Astellas, Bristol-Myers Squibb, Chugai, Eisai, Eli Lilly, Gilead Sciences, Ono, Otsuka, Pfizer, Taiho, and Taisho. To.T. has received research grants from Boehringer Ingelheim and Chugai, and payments for lectures from Astellas, Chugai, Eisai, Eli Lilly, MSD, Otsuka, and Taisho. Yu.N. has received speaker fees from Astellas Pharma, Asahi Kasei Pharma, Eisai, Eli Lilly, AstraZeneca, and Tanabe Pharma. K.K., Y.S., T.N., A.S., Y.O., Yu.M., T.F., Y.U., H.S., Y.So., Hi.M., D.T., W.Y., S.O., and Ma.H. declare no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe ANSWER Cohort was supported by grants from 12 pharmaceutical companies (AbbVie GK, Asahi-Kasei, Ayumi, Chugai, Eisai, Eli Lilly, Janssen KK, Ono, Sanofi KK, Taisho, Teijin Healthcare, and UCB Japan) and an information technology service company (CAC). This study was conducted as an investigator-initiated study. This study was funded by JSPS KAKENHI Grant-in-Aid for Scientific Research (25K02761) and JCR Research Promotion Program for Optimal Medical Care in Elderly-Onset Rheumatoid Arthritis. The funders had no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eK.K. performed the formal analysis and wrote the original draft of the manuscript. K.E. conceptualised the study, administered the project, and reviewed and edited the manuscript. Y.S. provided methodology and guidance on software and code analysis. Y.E., T.N., A.S., Y.O., K.T., Yu.M., A.K., Ko.M., T.F., Mo.H., T.O., H.Y., Y.U., To.T., H.S., Y.So., Hi.M., Yu.N., D.T., and W.Y. contributed to data curation and reviewed and edited the manuscript. K.N. and S.O. supervised the study. Ma.H. supervised the study and provided critical revision of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets collected and analysed during the current study are not publicly available due to patient privacy and ethical restrictions imposed by the institutional review boards but are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eD\u0026rsquo;Onofrio, B. \u003cem\u003eet al.\u003c/em\u003e Timely escalation to second-line therapies after failure of methotrexate in patients with early rheumatoid arthritis does not reduce the risk of becoming difficult-to-treat. \u003cem\u003eArthritis Res Ther\u003c/em\u003e 26, 192 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, Y. \u003cem\u003eet al.\u003c/em\u003e Factors influencing prescribing the first add-on disease-modifying antirheumatic drugs in patients initiating methotrexate for rheumatoid arthritis. \u003cem\u003eExplor Res Clin Soc Pharm\u003c/em\u003e 11, 100296 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmolen, J. S. \u003cem\u003eet al.\u003c/em\u003e EULAR recommendations for the management of rheumatoid arthritis with synthetic and biological disease-modifying antirheumatic drugs: 2022 update. \u003cem\u003eAnnals of the Rheumatic Diseases\u003c/em\u003e 82, 3\u0026ndash;18 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarigai, M. \u003cem\u003eet al.\u003c/em\u003e 2024 Update of the Japan College of Rheumatology Clinical Practice Guidelines for the Management of Rheumatoid Arthritis: Secondary publication. \u003cem\u003eModern Rheumatology\u003c/em\u003e 35, 387\u0026ndash;401 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLend, K. \u003cem\u003eet al.\u003c/em\u003e Association of rheumatoid factor, anti-citrullinated protein antibodies and shared epitope with clinical response to initial treatment in patients with early rheumatoid arthritis: data from a randomised controlled trial. \u003cem\u003eAnnals of the Rheumatic Diseases\u003c/em\u003e 83, 1657\u0026ndash;1665 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWientjes, M. H. M., Den Broeder, A. A., Welsing, P. M. J., Verhoef, L. M. \u0026amp; Van Den Bemt, B. J. F. Prediction of response to anti-TNF treatment using laboratory biomarkers in patients with rheumatoid arthritis: a systematic review. \u003cem\u003eRMD Open\u003c/em\u003e 8, e002570 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLv, Q. \u003cem\u003eet al.\u003c/em\u003e The Status of Rheumatoid Factor and Anti-Cyclic Citrullinated Peptide Antibody Are Not Associated with the Effect of Anti-TNFα Agent Treatment in Patients with Rheumatoid Arthritis: A Meta-Analysis. \u003cem\u003ePLoS ONE\u003c/em\u003e 9, e89442 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoo, B. S. \u003cem\u003eet al.\u003c/em\u003e Machine learning model for identifying important clinical features for predicting remission in patients with rheumatoid arthritis treated with biologics. \u003cem\u003eArthritis Res Ther\u003c/em\u003e 23, 178 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlsaber, A. R. \u003cem\u003eet al.\u003c/em\u003e Machine learning-based remission prediction in rheumatoid arthritis patients treated with biologic disease-modifying anti-rheumatic drugs: findings from the Kuwait rheumatic disease registry. \u003cem\u003eFront. Big Data\u003c/em\u003e 7, 1406365 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCurth, A., Peck, R. W., McKinney, E., Weatherall, J. \u0026amp; Van Der Schaar, M. Using Machine Learning to Individualize Treatment Effect Estimation: Challenges and Opportunities. \u003cem\u003eClin Pharma and Therapeutics\u003c/em\u003e 115, 710\u0026ndash;719 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAthey, S., Tibshirani, J. \u0026amp; Wager, S. Generalized random forests. \u003cem\u003eThe Annals of Statistics\u003c/em\u003e 47, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTopol, E. J. High-performance medicine: the convergence of human and artificial intelligence. \u003cem\u003eNat Med\u003c/em\u003e 25, 44\u0026ndash;56 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G. \u0026amp; King, D. Key challenges for delivering clinical impact with artificial intelligence. \u003cem\u003eBMC Med\u003c/em\u003e 17, 195 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKawahito, Y. [Guidelines for the management of rheumatoid arthritis]. \u003cem\u003eNihon Rinsho\u003c/em\u003e 74, 939\u0026ndash;943 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFujisawa, Y. \u003cem\u003eet al.\u003c/em\u003e Baseline neutrophil-to-lymphocyte ratio predicts drug retention of IL-6 inhibitors and JAK inhibitors in RA: the ANSWER cohort study. \u003cem\u003eRheumatology\u003c/em\u003e 65, keaf602 (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHiramatsu, Y. \u003cem\u003eet al.\u003c/em\u003e Effect of treat-to-target strategies on maternal and neonatal outcomes of rheumatoid arthritis: a multicentre real-world ANSWER cohort study. \u003cem\u003eRheumatology\u003c/em\u003e 65, keaf558 (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNozaki, Y. \u003cem\u003eet al.\u003c/em\u003e Clinical efficacy of JAK inhibitors for RA patients with poor-prognosis factors: the ANSWER cohort study. \u003cem\u003eRheumatology\u003c/em\u003e 65, keaf552 (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTsujimoto, K. \u003cem\u003eet al.\u003c/em\u003e Sustained efficacy of second-line JAK inhibitors in patients with rheumatoid arthritis: insights from the ANSWER cohort. \u003cem\u003eRheumatology\u003c/em\u003e 64, 4207\u0026ndash;4217 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChernozhukov, V. \u003cem\u003eet al.\u003c/em\u003e Double/debiased machine learning for treatment and structural parameters. \u003cem\u003eThe Econometrics Journal\u003c/em\u003e 21, C1\u0026ndash;C68 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBenavent, D. \u003cem\u003eet al.\u003c/em\u003e Artificial intelligence to predict treatment response in rheumatoid arthritis and spondyloarthritis: a scoping review. \u003cem\u003eRheumatol Int\u003c/em\u003e 45, 91 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalehi, F. \u003cem\u003eet al.\u003c/em\u003e Machine Learning Prediction of Treatment Response to Biological Disease-Modifying Antirheumatic Drugs in Rheumatoid Arthritis. \u003cem\u003eJCM\u003c/em\u003e 13, 3890 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePappas, D. A. \u003cem\u003eet al.\u003c/em\u003e Comparative effectiveness of first-line tumour necrosis factor inhibitor versus non-tumour necrosis factor inhibitor biologics and targeted synthetic agents in patients with rheumatoid arthritis: results from a large US registry study. \u003cem\u003eAnn Rheum Dis\u003c/em\u003e 80, 96\u0026ndash;102 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEdgerton, C. \u003cem\u003eet al.\u003c/em\u003e Real-World Treatment and Care Patterns in Patients With Rheumatoid Arthritis Initiating First-Line Tumor Necrosis Factor Inhibitor Therapy in the United States. \u003cem\u003eACR Open Rheumatol\u003c/em\u003e 6, 179\u0026ndash;188 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins, G. S. \u003cem\u003eet al.\u003c/em\u003e TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. \u003cem\u003eBMJ\u003c/em\u003e e078378 (2024) doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmj-2023-078378\u003c/span\u003e\u003cspan address=\"10.1136/bmj-2023-078378\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArnett, F. C. \u003cem\u003eet al.\u003c/em\u003e The American Rheumatism Association 1987 revised criteria for the classification of rheumatoid arthritis. \u003cem\u003eArthritis Rheum\u003c/em\u003e 31, 315\u0026ndash;324 (1988).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAletaha, D. \u003cem\u003eet al.\u003c/em\u003e 2010 rheumatoid arthritis classification criteria: an American College of Rheumatology/European League Against Rheumatism collaborative initiative. \u003cem\u003eAnn Rheum Dis\u003c/em\u003e 69, 1580\u0026ndash;1588 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoike, R. \u003cem\u003eet al.\u003c/em\u003e Japan College of Rheumatology 2009 guidelines for the use of tocilizumab, a humanized anti-interleukin-6 receptor monoclonal antibody, in rheumatoid arthritis. \u003cem\u003eModern Rheumatology\u003c/em\u003e 19, 351\u0026ndash;357 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerkelmans, G. F. N. \u003cem\u003eet al.\u003c/em\u003e Population median imputation was noninferior to complex approaches for imputing missing values in cardiovascular prediction models in clinical practice. \u003cem\u003eJ Clin Epidemiol\u003c/em\u003e 145, 70\u0026ndash;80 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVanderWeele, T. J. \u0026amp; Ding, P. Sensitivity Analysis in Observational Research: Introducing the E-Value. \u003cem\u003eAnn Intern Med\u003c/em\u003e 167, 268\u0026ndash;274 (2017).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9197723/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9197723/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSelecting the optimal biologic or targeted synthetic disease-modifying antirheumatic drug (b/tsDMARD) for individual patients with rheumatoid arthritis (RA) remains an unresolved clinical challenge. We developed and externally validated a causal inference framework that integrates a double machine learning-based causal forest model with guideline-derived safety constraints for b/tsDMARD category selection. Using multi-centre registry data (4,885 treatment courses from eight Japanese rheumatology facilities), the model was trained on six facilities (n\u0026thinsp;=\u0026thinsp;2,425) and externally validated on two independent facilities (n\u0026thinsp;=\u0026thinsp;2,460). The framework first estimates individualised treatment effects via the causal forest with double machine learning, then applies a guideline-based safety rule that exclude JAK inhibitors for patients with multiple cardiovascular risk factors before finalising recommendations. Patients receiving treatments concordant with AI recommendations achieved significantly higher clinical disease activity index response rates than those with discordant treatments: crude analysis 33.6% vs 27.2% (absolute risk difference [ARD]\u0026thinsp;=\u0026thinsp;6.4%, P\u0026thinsp;=\u0026thinsp;0.002; number needed to treat\u0026thinsp;=\u0026thinsp;16), propensity score-matched analysis 33.6% vs 28.2% (ARD\u0026thinsp;=\u0026thinsp;5.5%, P\u0026thinsp;=\u0026thinsp;0.037), and inverse probability of treatment weighting 31.6% vs 27.3% (ARD\u0026thinsp;=\u0026thinsp;4.3%, P\u0026thinsp;=\u0026thinsp;0.044). The system identified significant heterogeneity in conditional average treatment effects (group average treatment effects Q4\u0026ndash;Q1: Z\u0026thinsp;=\u0026thinsp;4.64, P\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and 12 clinically interpretable effect modifiers confirmed by two independent analytical methods. Treatment line-stratified analysis revealed that second-line patients derived the greatest benefit (ARD\u0026thinsp;=\u0026thinsp;16.0%, number needed to treat\u0026thinsp;=\u0026thinsp;6.3, P\u0026thinsp;=\u0026thinsp;0.0002). These findings demonstrate that an AI framework combining data-driven causal inference with guideline-derived safety guardrails can identify patients who benefit from personalised b/tsDMARD selection in RA.\u003c/p\u003e","manuscriptTitle":"A causal inference framework for personalised b/tsDMARD selection in rheumatoid arthritis: multi-centre development and external validation from the ANSWER cohort","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-19 12:17:01","doi":"10.21203/rs.3.rs-9197723/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"235404994915259551868824772609988108664","date":"2026-05-03T14:20:17+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-09T18:36:20+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-24T19:52:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-24T11:07:12+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2026-03-23T08:22:15+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"be777090-6c66-4de1-a668-2d734d42b1c2","owner":[],"postedDate":"April 19th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"235404994915259551868824772609988108664","date":"2026-05-03T14:20:17+00:00","index":61,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":66445212,"name":"Health sciences/Diseases"},{"id":66445213,"name":"Health sciences/Medical research"},{"id":66445214,"name":"Health sciences/Rheumatology"}],"tags":[],"updatedAt":"2026-04-19T12:17:01+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-19 12:17:01","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9197723","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9197723","identity":"rs-9197723","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}