Skip to main content
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2026 May 7;78(11):2232–2245. doi: 10.1002/art.70165

Machine Learning to Predict Remission Between 6 and 24 Months in Rheumatoid Arthritis: Insights From JAK, an International Registry Collaboration

Zubeyir Salis 1,2,✉, Denis Mongin 2,3, Denis Choquette 4, Louis Coupal 4, Catalin Codreanu 5, Florenzo Iannone 6, Roberto Caporali 7, Tore K Kvien 8, Sella Provan 8, Ruth Fritsch‐Stork 9,10,11, Dan Nordström 12, Nina Trokovic 12, Karel Pavelka 13, Jakub Závada 13, Ana Rodrigues 14, Ziga Rotar 15,16, Prodromos Sidiropoulos 17,18, Irini Flouri 18, Céline Lamacchia 2, Michele Iudici 2,3, Delphine Courvoisier 2,3, Kim Lauper 2,3, Axel Finckh 2,3
PMCID: PMC13712546  PMID: 41940462

Abstract

Objective

To develop, externally validate, and simplify a machine learning model to predict remission between 6 and 24 months in patients with rheumatoid arthritis (RA) initiating tumor necrosis factor inhibitors, JAK inhibitors, interleukin‐6 inhibitors, abatacept, or rituximab using data from 11 international registries in the JAK‐pot collaboration.

Methods

We analyzed 21,675 treatment courses for model training, 5,418 courses for internal validation, and 1,807 courses from Switzerland for external validation. Remission was defined as a Clinical Disease Activity Index score ≤2.8 within 6 to 24 months after treatment initiation or switching. An XGBoost model was trained on 63 baseline demographic and clinical variables. Model performance was evaluated using the area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Predictor contributions were quantified using Shapley Additive Explanations values. A simplified model using the top 10 predictors and a logistic regression benchmark were also evaluated.

Results

In external validation, the full model achieved an AUC of 0.797 (95% confidence interval 0.774–0.821), sensitivity of 0.804, specificity of 0.650, PPV of 0.454, and NPV of 0.902. The simplified model performed comparably (AUC 0.802). Logistic regression using the same 10 predictors achieved an AUC of 0.809. Key predictors included patient global assessment, 28‐tender joint count, previous biologic/targeted synthetic disease‐modifying antirheumatic drug exposure, and Health Assessment Questionnaire Disability Index score.

Conclusion

Using routinely collected baseline data across 11 registries, prediction of remission showed limited discrimination and was best suited to ruling out remission. Performance was similar for a 10‐variable model and logistic regression, indicating limited incremental value of model complexity without richer predictors.

graphic file with name ART-78-2232-g001.webp


graphic file with name ART-78-2232-g004.webp

INTRODUCTION

Rheumatoid arthritis (RA) is a chronic autoimmune disease characterized by persistent inflammation, progressive joint damage, and reduced quality of life. 1 Despite advancements in disease‐modifying antirheumatic drugs (DMARDs), including biologic and targeted synthetic‐DMARDs (b/ts‐DMARDs), achieving remission, an established treatment goal, 2 remains difficult. 1 Approximately 30% to 40% of patients fail to respond to initial treatment, and only around 50% achieve sustained remission over time. 3 , 4 These outcomes highlight the need for tools to predict treatment response and support evidence‐based clinical decision‐making.

Numerous models have been developed to predict remission in RA, largely based on regression‐based statistical approaches. Although these models have provided valuable insights, they have typically focused on earlier b‐DMARDs, such as tumor necrosis factor inhibitors (TNFi), and seldom address newer therapies, such as JAK inhibitors (JAKi), or remission outcomes defined by long‐term clinical end points. Many also rely on small or narrowly defined patient cohorts and short follow‐up durations or incorporate specialized biomarkers that limit their applicability in routine care. 5 , 6

Machine learning (ML), a core tool in precision medicine, offers a fundamentally different approach. 7 It enables the discovery of complex, nonlinear interactions across diverse variables, interactions that traditional models may miss. Although several early ML studies in RA have reported encouraging accuracy, 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 they have mostly been limited to TNFi‐treated populations, lacked validation in external cohorts, or relied on omics‐level inputs not routinely collected in clinical settings. 16 , 17 Nonetheless, the scalability and adaptability of ML models make them an attractive framework for building more generalizable, clinically useful prediction tools.

To address these limitations, we developed an ML framework to predict remission between 6 and 24 months in patients with RA initiating treatment with b/ts‐DMARDs, including TNFi, JAKi, interleukin‐6 inhibitors (IL‐6i), abatacept (ABA), or rituximab (RTX) using only baseline demographic and clinical data. By focusing on widely available variables and using a large, diverse, multicountry cohort, we aimed to develop ML models that are externally validated, clinically relevant, and generalizable to real‐world RA populations.

MATERIALS AND METHODS

We employed a supervised ML model, eXtreme Gradient Boosting (XGBoost), to predict RA remission between 6 and 24 months after treatment initiation or switching. 18 XGBoost was chosen because gradient‐boosted tree ensembles consistently achieve state‐of‐the‐art performance on structured clinical data, outperforming or matching random forests, support vector machines, and deep learning models in large benchmark studies. 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 It also handles missing values natively, is computationally efficient, 18 and supports model interpretability. We applied Shapley Additive Explanations (SHAP) values 27 to quantify predictor contributions. The methodology for this study is summarized in Figure 1, which outlines the patient sample, data preprocessing, and model development and training.

Figure 1.

Figure 1

A summary of the methodology throughout the process of creation of trained models for the prediction of remission between 6 and 24 months after treatment initiation or switching. The figure shows the data source, selection of patients, and data preprocessing for the creation of the training data set, the internal validation data set, and the external validation data set. *Austria, Canada, Czech Republic, Finland, Greece, Italy, Norway, Portugal, Romania, Slovenia, and Switzerland. ABA, abatacept; CDAI, Clinical Disease Activity Index; IL‐6i, IL‐6 inhibitors; JAKi, JAK inhibitors; RA, rheumatoid arthritis; RITX, rituximab; TNFi, tumor necrosis factor inhibitors.

Data

We used data from the JAK‐pot registry collaboration, an investigator‐initiated international study aimed at evaluating the effectiveness and safety of JAKi and b‐DMARDs in RA. 28 , 29 , 30 Local ethics committees (when required according to local legislation) approved the collection of data in each participating register, and the Geneva Ethics Committee approved this specific study (CCER 2017‐02278, amendment 2). The analysis included individual treatment courses from 11 national registries: ATTRA (Czech Republic), BIOREG (Austria), BioRx.si (Slovenia), CATCH (Canada), University of Crete Rheumatology Clinic Registry (Greece), GISEA (Italy), NOR‐DMARD (Norway), Reuma.pt (Portugal), ROB‐FIN (Finland), RRBR (Romania), and SCQM (Switzerland). All data were analyzed at the level of individual treatment courses, with patients able to contribute multiple courses during the study period.

Outcome measures

Patients were classified as in remission if they had at least one Clinical Disease Activity Index (CDAI) score ≤2.8 recorded between 6 and 24 months after treatment initiation or switching. 1 This definition captured the achievement of remission at any visit within the follow‐up window and did not require sustained remission at subsequent visits. Remission before six months did not affect inclusion because such patients remained eligible, but early remission was not considered for classification. CDAI was chosen because it focuses on clinical assessments and is not affected by markers of inflammation, which can be strongly influenced by therapies such as IL‐6i and JAKi. 31 , 32

Patient sample

The data set included 28,900 treatment courses from 11 countries (Figure 1). Patients had a clinical diagnosis of RA and initiated treatment with TNFi, JAKi, IL‐6i, ABA, or RTX. To be included, treatment courses had to have at least one available CDAI measurement recorded between 6 and 24 months after treatment initiation or switching (Figure 1).

Data preprocessing: Overview and selection of predictors

To prepare the data set for analysis with the XGBoost model, we applied a structured approach to predictor selection and transformation. All candidate predictors were restricted to baseline variables, consistent with the study aim of prediction at the time of treatment initiation or switching. Predictors with no variability (eg, constant across all observations) or with administrative or identifier roles (eg, dates, country codes, specific IDs) were excluded. Numeric variables with more than 10 unique values were retained as continuous, and missing values in these were handled natively by the XGBoost algorithm. Numeric variables with 10 or fewer unique values (ie, low cardinality) were reclassified as categorical. Most of these were originally binary clinical or treatment indicators that included missing values and were therefore encoded as three‐level categorical variables (“yes,” “no,” and “missing”). Examples include rheumatoid factor, anti–cyclic citrullinated peptide (anti‐CCP) antibodies, concomitant methotrexate use, and comorbidity indicators such as diabetes, hyperlipidemia, and myocardial infarction, all listed in Table 1. All categorical variables, including those reclassified from low‐cardinality numeric variables, were then converted into binary indicator variables (ie, one‐hot encoding, in which each category is represented as a separate 0/1 column). Missing values in categorical variables were not imputed but treated as a distinct “missing” category before encoding. Predictors were also screened for near‐zero variance, but none met exclusion criteria. The final data set comprised 63 predictors categorized into four domains: demographic factors, clinical assessments, medical history, and treatment details. Data were randomly split into training and validation subsets, with stratification to ensure balanced representation of outcomes across the data sets. Specifically, 80% of the data (21,675 treatment courses) were used for training and 20% of the data (5,418 treatment courses) were for internal validation. External validation was conducted using data from Switzerland (1,807 treatment courses), which was excluded from the training and internal validation subsets (Figure 1). A priori, the SCQM (Switzerland) data set was selected as the external validation data set because it was not used for model development and provided an independent national registry for transportability testing. However, some predictors were not available in SCQM (notably baseline pain), and baseline CDAI score was frequently missing, reflecting structural differences in data capture across registries. Additionally, due to known inconsistencies in Health Assessment Questionnaire Disability Index (HAQ‐DI) scoring within the SCQM data set, corrected values were provided and incorporated. However, residual structural differences in the original questionnaire remain and are acknowledged as a limitation.

Table 1.

Data distribution of predictors, outcome, and countries in the training, internal validation, and external validation data sets*

Training data set (n = 21,675) Internal validation data set (n = 5,418) External validation data set (n = 1,807) Missing training/internal validation/external validation, %
Predictors (no.)
Demographics
1. Age, mean ± SD, years 56.02 ± 12.68 56.19 ± 12.52 56.59 ± 13.60 <0.1/<0.1/0.0
2. Sex, female, n (%) 16,275 (75.1) 4,112 (75.9) 1,398 (77.4) 0.0/0.0/0.0
3. Height, mean ± SD, cm 164.90 ± 8.43 165.02 ± 8.38 166.35 ± 9.03 0.0/0.0/0.0
4. Weight, mean ± SD, kg 71.97 ± 15.41 72.21 ± 15.78 73.44 ± 16.94 0.0/0.0/0.0
5. BMI, mean ± SD, kg/m2 26.46 ± 5.30 26.49 ± 5.33 26.52 ± 5.76 0.0/0.0/0.0
6. Educational status, years, n (%) 37.7/38.3/11.6
<10 4,607 (34.1) 1,187 (35.5) 318 (19.9) –
10–13 5,856 (43.4) 1,419 (42.5) 905 (56.6) –
>13 3,034 (22.5) 735 (22.0) 375 (23.5) –
Medical history
7. Previous b/ts‐DMARD, mean ± SD no. of treatments 1.37 ± 1.46 1.38 ± 1.49 1.51 ± 1.70 0.1/0.1/0.0
8. Past use of NSAIDs, n (%) 826 (24.8) 203 (24.5) 1,249 (98.4) 84.7/84.7/29.8
9. Past use of COX‐2 inhibitor, n (%) 363 (10.9) 94 (11.4) 231 (82.5) 84.7/84.7/84.5
10. Family history of cardiovascular event, n (%) 59 (12.9) 18 (15.7) NA 97.9/97.9/NA
Clinical assessments
11. Disease duration, mean ± SD, years 10.61 ± 8.84 10.58 ± 8.94 11.52 ± 10.24 5.0/4.8/0.5
12. CDAI, mean ± SD 25.80 ± 14.05 25.73 ± 14.03 19.00 ± 10.78 45.0/43.9/81.7
13. DAS28, mean ± SD 5.04 ± 1.53 5.05 ± 1.51 3.98 ± 1.39 41.5/40.3/73.5
14. ESR, mean ± SD, mm/hr 32.56 ± 24.54 32.20 ± 24.20 19.73 ± 18.31 41.9/40.2/73.8
15. HAQ‐DI, 0–3 scale, mean ± SD 1.25 ± 0.71 1.25 ± 0.70 0.82 ± 0.69 48.4/47.3/26.1
16. SJC28, mean ± SD 4.71 ± 5.25 4.74 ± 5.21 2.90 ± 3.90 9.6/8.7/4.0
17. TJC28, mean ± SD 6.70 ± 6.60 6.81 ± 6.52 3.23 ± 4.77 9.3/8.5/4.4
18. 66‐swollen joint count, mean ± SD 4.79 ± 6.10 5.29 ± 6.10 2.44 ± 3.95 97.4/97.5/93.0
19. 68‐tender joint count, mean ± SD 5.93 ± 7.69 5.76 ± 7.90 5.83 ± 7.62 97.4/97.5/92.9
20. RF, n (%) 12,203 (67.3) 2,990 (65.6) 1,197 (67.0) 16.3/15.9/1.1
21. anti‐CCP, n (%) 9,491 (59.4) 2,283 (57.7) 1,152 (65.6) 26.3/27.0/2.8
22. CRP concentration, mean ± SD, mg/L 13.20 ± 22.49 12.55 ± 21.33 6.95 ± 12.14 10.1/8.9/10.6
23. PhGA, 0–100 scale, mean ± SD 40.46 ± 25.53 40.41 ± 25.41 26.51 ± 20.53 12.7/11.8/11.5
24. PtGA, 0–100 scale, mean ± SD 48.43 ± 27.72 48.64 ± 27.71 34.84 ± 25.84 7.6/7.1/25.9
25. Patient‐assessed global health status, 0–100 scale, mean ± SD 47.76 ± 25.94 48.52 ± 25.49 27.54 ± 21.91 63.4/62.9/89.9
26. Patient assessment of pain, 0–100 scale, mean ± SD 53.42 ± 26.85 53.73 ± 26.65 NA 34.8/34.0/NA
27. Patient assessment of fatigue, 0–100 scale, mean ± SD 47.98 ± 29.00 50.26 ± 29.29 NA 79.1/79.6/NA
28. Ever smoker, n (%) 4,425 (25.0) 1,104 (24.8) 523 (53.4) 18.3/17.9/45.8
29. Myocardial infarction, n (%) 130 (1.3) 55 (2.2) 30 (1.7) 53.3/52.9/0.6
30. Stroke, n (%) 235 (1.5) 71 (1.8) NA 28.0/28.1/NA
31. Other cardiovascular problems, n (%) 2,933 (18.5) 796 (20.1) NA 26.7/26.8/NA
32. Lung, n (%) 1,226 (7.8) 335 (8.6) 217 (12.3) 27.8/27.8/2.6
33. Hypertension, n (%) 4,889 (30.5) 1,255 (31.3) 397 (22.4) 26.1/26.0/1.9
34. Hyperlipidemia, n (%) 1,776 (11.8) 454 (12.2) 241 (13.9) 30.8/31.1/4.0
35. With HIV, n (%) 0 (0.0) 0 (0.0) 0 (0.0) 71.4/71.6/0.6
36. Serious infections, n (%) 2,557 (17.0) 657 (17.6) 329 (19.0) 30.8/31.0/4.3
37. Malignancy, n (%) 568 (3.6) 119 (3.0) 106 (6.0) 27.7/27.9/1.5
38. Diabetes, n (%) 1,479 (9.4) 369 (9.4) 145 (8.1) 27.5/27.7/1.3
39. Past or current neuropsychiatric disorder, n (%) 1,382 (9.2) 366 (9.7) 299 (17.3) 30.5/30.7/4.3
40. PASI, n (%) <100.0/<100.0/NA
0 9 (90.0) 1 (0.0) NA –
1.3 1 (10.0) 0 (0.0) NA –
41. Current b/ts‐DMARD treatment type, n (%) 0.0/0.0/0.0
TNFi 12,343 (56.9) 3,128 (57.7) 739 (40.9) –
JAKi 3,047 (14.1) 749 (13.8) 394 (21.8) –
IL‐6i 2,501 (11.5) 608 (11.2) 295 (16.3) –
ABA 1,302 (6.0) 302 (5.6) 221 (12.2) –
RTX 2,482 (11.5) 631 (11.6) 158 (8.7) –
Concomitant medicine use
42. Other DMARD, n (%) 1,221 (7.5) 341 (8.5) 4 (0.2) 25.3/25.7/0.6
43. Antiaggregant, n (%) 369 (6.0) 96 (6.2) 164 (12.0) 71.5/71.4/24.7
44. Anticoagulant, n (%) 75 (1.2) 28 (1.8) NA 71.5/71.4/NA
45. NSAIDs, n (%) 3,022 (37.3) 760 (37.7) 696 (51.4) 62.6/62.8/25.1
46. COX‐2 inhibitor, n (%) 625 (7.8) 153 (7.6) 83 (6.1) 62.8/62.9/25.1
47. Current b‐DMARD dose, mean ± SD, mg 261.92 ± 395.56 261.94 ± 392.04 321.12 ± 331.33 39.7/39.1/21.8
48. Current b‐DMARD dosing interval, mean ± SD, days 19.20 ± 35.12 19.11 ± 34.83 38.02 ± 52.49 33.8/33.3/21.9
49. Current ts‐DMARD dose, mean ± SD, mg 40.43 ± 53.11 38.01 ± 50.34 6.80 ± 4.06 80.4/80.7/78.3
50. ts‐DMARD schedule, no. of intake per day, n (%) 77.7/78.1/78.4
0–1 21 (0.5) 5 (0.4) 5 (1.2) –
1 4,456 (92.0) 1,094 (92.2) 211 (54.0) –
2 365 (7.5) 88 (7.4) 175 (44.8) –
51. b‐DMARD route of administration, n (%) 42.1/42.2/21.9
Intravenous 3,092 (24.7) 757 (24.2) 558 (39.5) –
Oral 1,385 (11.0) 334 (10.7) 0 (0.0) –
Subcutaneous 8,062 (64.3) 2,041 (65.2) 854 (60.5) –
52. Methotrexate, n (%) 10,434 (58.2) 2,641 (59.0) 715 (39.8) 17.4/17.4/0.6
53. Methotrexate dose, mean ± SD, mg 14.09 ± 8.67 13.92 ± 8.79 16.01 ± 5.10 52.8/52.1/60.4
54. Methotrexate route of administration, n (%) 87.5/86.6/78.5
Intramuscular 3 (0.1) 0 (0.0) 0 (0.0) –
Intravenous 26 (1.0) 6 (0.8) 0 (0.0) –
Oral 1,731 (64.0) 466 (64.3) 115 (29.6) –
Subcutaneous 944 (34.9) 253 (34.9) 274 (70.4) –
55. Leflunomide, n (%) 3,809 (21.3) 969 (21.7) 309 (17.2) 17.4/17.4/0.6
56. Leflunomide dose, mean ± SD, mg 13.47 ± 9.13 13.35 ± 9.16 18.37 ± 3.82 76.8/76.5/82.9
57. Sulfasalazine, n (%) 2,500 (14.0) 647 (14.5) 92 (5.1) 17.4/17.4/0.6
58. Sulfasalazine dose, mean ± SD, mg 1,122.08 ± 1,104.90 1,117.86 ± 1,120.17 1,657.61 ± 601.87 82.1/81.3/94.9
59. Hydroxychloroquine, n (%) 2,625 (16.2) 671 (16.7) 121 (6.7) 25.3/25.7/0.6
60. Hydroxychloroquine dose, mean ± SD, mg 318.78 ± 86.45 325.22 ± 86.43 276.03 ± 96.85 89.1/89.0/93.3
61. Glucocorticoids, n (%) 10,145 (47.4) 2,520 (47.0) 529 (29.5) 1.3/1.1/0.6
62. Glucocorticoids dose, mean ± SD, mg 6.38 ± 12.34 6.37 ± 8.36 7.76 ± 5.58 53.8/54.2/71.1
63. Glucocorticoids route of administration, n (%) 84.2/83.8/83.4
Intravenous 18 (0.5) 2 (0.2) 0 (0.0) –
Local 18 (0.5) 8 (0.9) 0 (0.0) –
Oral 3,388 (98.8) 864 (98.4) 300 (100.0) –
Subcutaneous 4 (0.1) 4 (0.5) 0 (0.0) –
Outcome
Achieving RA remission, as defined by a CDAI ≤2.8, between 6 and 24 mos following treatment initiation or switching, n (%) 6,354 (29.3) 1,588 (29.3) 480 (26.6) 0.0/0.0/0.0
Country, n (%) 0.0/0.0/0.0
Austria 393 (1.8) 97 (1.8) NA –
Canada 473 (2.2) 119 (2.2) NA –
Czech Republic 1,531 (7.1) 381 (7.1) NA –
Finland 2,647 (12.2) 627 (11.6) NA –
Greece 250 (1.2) 83 (1.5) NA –
Italy 4,447 (20.5) 1,146 (21.2) NA –
Norway 1,932 (8.9) 481 (8.9) NA –
Portugal 2,856 (13.2) 711 (13.1) NA –
Romania 5,361 (24.7) 1,310 (24.2) NA –
Slovenia 1,785 (8.2) 463 (8.6) NA –
Switzerland NA NA 1,807 (100.0) –
*

ABA, abatacept; anti‐CCP, anti–cyclic citrullinated peptide; BMI, body mass index; b/ts‐DMARD, biologic/targeted synthetic disease‐modifying antirheumatic drug; CDAI, Clinical Disease Activity Index; COX‐2, cyclooxygenase 2; CRP, C‐reactive protein; DAS28, Disease Activity Score in 28 joints; ESR, erythrocyte sedimentation rate; HAQ‐DI, Health Assessment Questionnaire Disability Index; IL‐6i, IL‐6 inhibitors; JAKi, JAK inhibitors; NA, not applicable; NSAID, nonsteroidal anti‐inflammatory drug; PASI, Psoriasis Area Severity Index; PhGA, physician global assessment; PtGA, patient global assessment; RA, rheumatoid arthritis; RF, rheumatoid factor; RTX, rituximab; SJC28, 28‐swollen joint count; TJC28, 28‐tender joint count; TNFi, tumor necrosis factor inhibitors.

ML model development and training

Model configuration and optimization

Hyperparameter tuning was conducted using a systematic grid search approach to evaluate combinations of parameter values to identify the best‐performing configuration. The following parameters were systematically adjusted: maximum tree depth (3, 6, 9), the number of boosting rounds (50, 100, 150), learning rate (0.01, 0.05, 0.10), gamma (0, 0.1, 0.2), subsample ratio of columns per tree (0.5, 0.8), subsample fraction of training samples for boosting (0.6, 0.8, 1.0), and the minimum sum of instance weights in a child (1, 3, 5). Cross‐validation, a method of splitting the training data into subsets for iterative training and validation, was used to guide the grid search in identifying the optimal combination of parameters for maximizing predictive performance.

Evaluation metrics

Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) as the primary metric for discrimination. Additional metrics included sensitivity, specificity, balanced accuracy, positive predictive value (PPV), negative predictive value (NPV), and F1 score. The F1 score, computed for both remission (positive class) and nonremission (negative class), represents the harmonic mean of precision and sensitivity and summarizes predictive performance in imbalanced outcome settings.

Predictor importance

SHAP values, which quantify the contribution of each predictor to individual predictions, were used to identify and rank predictors based on their contribution to model output. Only main effect SHAP values were computed, and interaction effects between predictors were not included.

Model calibration

The classification threshold was calibrated using the training data set by selecting the value that maximized the Youden index (sensitivity + specificity − 1), thereby balancing sensitivity and specificity. This fixed threshold was then applied without re‐estimation to both the internal validation data set and the external validation data set (Switzerland) to ensure consistency across performance evaluation (Figure 1).

Model with top 10 most influential predictors

To improve interpretability and facilitate exploratory applications, we developed a simplified version of the model using the 10 most influential predictors, as identified by their mean absolute SHAP values in the comprehensive model.

Logistic regression model (benchmark comparison)

To benchmark performance against a traditional method, we fitted a logistic regression model using the same 10 baseline predictors selected for the simplified XGBoost model. Missing data in the training data set were handled using multiple imputation via chained equations (MICE), with 10 imputed data sets generated using predictive mean matching. A separate logistic regression model was fitted in each imputed training data set. For each iteration, model performance was evaluated on the imputed training data set, the internal validation data set, and the external validation data set. The internal and external validation data sets were singly imputed once per iteration using the same MICE method, with independent random seeds. Because the external validation data set lacked patient‐reported pain, a second logistic model excluding the “pain” predictor was refit in every imputation for the external validation data set. Predicted probabilities were converted to class labels using a fixed decision threshold calibrated on the training data set by maximizing the Youden index (sensitivity + specificity − 1). This threshold was then applied without re‐estimation to the internal validation and external validation data sets. Performance metrics were the same as those reported for XGBoost, including AUC, sensitivity, specificity, balanced accuracy, PPV, NPV, and F1 score. AUC and 95% confidence intervals were obtained via stratified bootstrap resampling (500 resamples) within each imputed data set and then averaged across imputations. For the other metrics, results were averaged across imputations and reported as point estimates.

Statistical analysis and software

The AUC and corresponding 95% CI were estimated using stratified bootstrap resampling, preserving outcome class proportions (eg, positive and negative cases) in each resampled data set. We used 1,000 bootstrap resamples for the XGBoost analyses and 500 resamples within each imputed data set for the logistic regression benchmark, with estimates averaged across imputations.

Sensitivity analyses

To assess the robustness of model performance to potential sources of bias related to baseline disease activity and patient‐level data structure, we performed the following three sensitivity analyses. First, we evaluated the final trained XGBoost model in the internal validation data set restricted to patients with active disease at baseline, defined as CDAI score > 10. Model discrimination was summarized using AUC, and SHAP values were recalculated to examine the stability of predictor importance across disease activity strata. Second, we developed a benchmark model using baseline CDAI score as the only predictor and evaluated the model discrimination using AUC in the internal validation data set. Third, internal training and validation splits were enforced at the patient level rather than the treatment course level, such that all treatment courses from a given patient were assigned to a single data split. This resulted in 27,093 treatment courses from 20,619 unique patients, with 21,630 courses from 16,496 patients allocated to training and 5,463 courses from 4,123 patients allocated to internal validation. All preprocessing steps, predictors, model configuration, and evaluation procedures were otherwise identical to those used in the primary analysis.

Subgroup analysis by baseline b/ts‐DMARD therapy class

To assess whether model performance differed according to baseline b/ts‐DMARD therapy class, we conducted a prespecified subgroup analysis stratified by baseline therapy class (TNFi, IL‐6i, JAKi, ABA, and RTX). The final trained XGBoost model was applied without retraining within each subgroup in the internal validation data set and the external validation data set. Discrimination and classification metrics were computed separately within each subgroup using the same probability threshold derived from the training ROC curve. For model interpretability, SHAP values derived from the training data set were summarized separately within each therapy class, and predictors were ranked by mean absolute SHAP value.

Software

All data preparation and ML analyses were performed using R version 4.5.0. The following R packages were used: XGBoost (version 1.7.5.1) for model training, pROC (version 1.18.4) for ROC curve analysis, corrplot (version 0.92) for creating correlation heatmaps, MICE (version 3.17.0) for imputation in the logistic regression, and caret (version 7.0‐1) for confusion matrix‐based metrics in the logistic regression. Key predictors were identified using SHAP values computed with the SHAPforxgboost package (version 0.1.3), with variables ranked based on their mean absolute contributions.

Data and code availability

The JAK‐pot collaboration is based on data contributions from multiple rheumatology registers. Although the data are hosted and analyzed centrally in Geneva, they remain the property of the individual contributing registers. As such, data cannot be shared by the collaboration. Any external data request must be addressed directly to each contributing register and requires approval from all data owners. The R codes for the models’ creation can be found at https://github.com/Zube-Geneve/RA_Remission_JAKPOT.git.

RESULTS

Data distribution across the training, internal validation, and external validation data sets

The training data set included 21,675 treatment courses, the internal validation data set comprised 5,418 treatment courses, and the external validation data set contained 1,807 treatment courses (Table 1). Across all data sets, participants were predominantly female (75.1%–77.4%) and middle‐aged (mean age 56.02–56.59 years) and had an average body mass index in the overweight range (mean 26.46–26.52) (Table 1).

Several clinical baseline characteristics varied across data sets (Table 1). Baseline patient global assessment (PtGA) and physician global assessment (PhGA) were markedly lower in the external validation data set (mean 34.84 and 26.51) compared with the training (mean 48.43 and 40.46) and internal validation (mean 48.64 and 40.41) data sets, respectively. Similarly, the 28‐tender joint count (TJC28) was substantially lower in the external validation data set (mean 3.23) than in the training and internal validation data sets (means 6.70–6.81, respectively), suggesting milder baseline disease. Functional status, as measured by HAQ‐DI, was also lower in the external validation data set (mean 0.82) than in the training and internal validation data sets (mean 1.25 in both). Consistent with this overall trend, baseline CDAI scores and C‐reactive protein (CRP) levels were lower in the external validation data set (mean CDAI score 19.00; mean CRP level 6.95 mg/L) compared with the training and internal validation data sets (mean CDAI scores 25.80 and 25.73, respectively; mean CRP levels 13.20 and 12.55 mg/L, respectively). Among treatment courses with available baseline CDAI scores, baseline remission (CDAI score ≤ 2.8) was uncommon: 296 of 11,932 (2.5%) in the training data set, 83 of 3,039 (2.7%) in the internal validation data set, and 10 of 331 (3.0%) in the external validation data set (baseline CDAI missingness: 45.0%, 43.9%, 81.7%). In contrast to these patterns, the proportion of patients testing positive for anti‐CCP antibodies was higher in the external validation data set (65.6%) compared with the training (59.4%) and internal validation data sets (57.7%). Notably, previous b/ts‐DMARD exposure was slightly higher in the external validation data set (mean 1.51 in the external validation data set vs 1.37–1.38 elsewhere). Patient‐reported pain levels at baseline were high in the training and internal validation data sets (mean visual analog scale >53/100), reflecting substantial symptom burden despite prior treatments. Pain was not recorded in the external validation data set (Table 1).

Patterns of b/ts‐DMARDs, including TNFi, JAKi, and IL‐6i, varied across data sets. TNFi was the most common treatment in all data sets, comprising 56.9% of courses in the training data set, 57.7% in the internal validation data set, and 40.9% in the external validation data set. Conversely, the external validation data set had a higher proportion of participants treated with JAKi (21.8%) and IL‐6i (16.3%) compared with the training (14.1% and 11.5%, respectively) and internal validation data sets (13.8% and 11.2%, respectively).

Remission rates between 6 and 24 months differed across data sets. In both the training and internal validation data sets, 29.3% of participants achieved remission (CDAI score ≤ 2.8), whereas the external validation data set showed a slightly lower remission rate of 26.6%.

Model metrics

Training data set

In the training data set, the model achieved an AUC of 0.781 with the initial training parameters. Following fine‐tuning, the optimized parameters were as follows: maximum tree depth, 6; the number of boosting rounds, 150; the learning rate, 0.05; gamma, 0; subsample ratio of columns by tree, 0.5; subsample ratio, 0.6; and the minimum child weight, 5. The optimal classification threshold was identified as 0.305. Under these optimized conditions, the model achieved an AUC of 0.863 (95% CI 0.858–0.868) with F1 scores of 0.679 for remission (positive) and 0.846 for nonremission (negative) classes. Sensitivity for predicting the remission was 0.752, specificity was 0.808, balanced accuracy was 0.780, PPV was 0.619, and NPV was 0.887.

Internal validation data set

Applying the optimized model to the internal validation data set, the model achieved an AUC of 0.759 (95% CI 0.744–0.773). At the same threshold of 0.305, the model had an F1 score of 0.564 for the positive class and 0.799 for the negative class (Figure 2A). Sensitivity was 0.608, specificity was 0.773, balanced accuracy was 0.691, PPV was 0.526, and NPV was 0.826.

Figure 2.

Figure 2

Receiver operating characteristic (ROC) curves showing the performance of models for predicting the achievement of RA remission between 6 and 24 months after treatment initiation or switching using 63 predictors. (A) Internal validation data set. (B) External validation data set. Red points show the corresponding true positive rate (TPR) and false positive rate (FPR) for the selected thresholds. AUC, area under the curve; RA, rheumatoid arthritis.

External validation data set

On the external validation data set, the model yielded an AUC of 0.797 (95% CI 0.774–0.821). At the prespecified threshold of 0.305, the model obtained an F1 score of 0.580 for the positive class and 0.755 for the negative class (Figure 2B). Sensitivity was 0.804, specificity was 0.650, balanced accuracy was 0.727, PPV was 0.454, and NPV was 0.902.

Key predictors

Figure 3 displays the 10 most influential baseline predictors, ranked by their mean absolute main effect SHAP values. The top‐ranked feature was PtGA, followed by TJC28, previous exposure to b/ts‐DMARDs, HAQ‐DI score, PhGA, age, CDAI score, CRP level, patient‐reported pain, and anti‐CCP antibodies. To assess the feasibility of a more parsimonious approach, we trained a simplified model using only these 10 predictors.

Figure 3.

Figure 3

Top 10 predictors of remission probability ranked by mean absolute Shapley Additive Explanations (SHAP) values from the fine‐tuned XGBoost model. Higher values indicate greater overall contribution of the predictor to the output of the model, irrespective of whether the effect increases or decreases remission probability. anti‐CCP, anti–cyclic citrullinated peptide antibodies; b/ts‐DMARD, biologic/targeted synthetic disease‐modifying antirheumatic drug; CDAI, Clinical Disease Activity Index; HAQ‐DI, Health Assessment Questionnaire Disability Index; VAS, visual analog scale.

Model metrics using the top 10 predictors

Training data set

The model trained with the top 10 predictors achieved an initial AUC of 0.758 during training. After fine‐tuning, the optimized parameters were as follows: maximum tree depth, 6; number of boosting rounds, 100; learning rate, 0.05; gamma, 0.2; subsample ratio of columns by tree, 0.5; subsample ratio, 1.0; and minimum child weight, 1. Using the optimal classification threshold of 0.295, the model achieved an AUC of 0.806 (95% CI 0.799–0.812). F1 scores for the positive and negative classes were 0.607 and 0.802, respectively. The model reached a sensitivity of 0.695, specificity of 0.754, and balanced accuracy of 0.724. Additionally, the PPV was 0.539, and the NPV was 0.856.

Internal validation data set

On the internal validation data set, the model achieved an AUC of 0.731 (95% CI 0.716–0.746). At the same optimal threshold of 0.295, the F1 score was 0.535 for the positive class and 0.775 for the negative class (Figure 4A). Sensitivity was 0.594, specificity was 0.739, and balanced accuracy was 0.667. The PPV was 0.486, and the NPV was 0.815.

Figure 4.

Figure 4

Receiver operating characteristic (ROC) curves showing the performance of models for predicting the achievement of RA remission between 6 and 24 months after treatment initiation or switching using 10 predictors. (A) Internal validation data set. (B) External validation data set. Red points show the corresponding true positive rate (TPR) and false positive rate (FPR) for the selected thresholds. AUC, area under curve; RA, rheumatoid arthritis.

External validation data set

The generalizability of the model was evaluated on the external validation data set, in which it achieved an AUC of 0.802 (95% CI 0.778–0.825). At the fixed threshold of 0.295, the model produced an F1 score of 0.572 for the positive class and 0.722 for the negative class (Figure 4B). Sensitivity was 0.848, specificity was 0.596, and balanced accuracy was 0.722. The PPV was 0.432, and the NPV was 0.916.

Logistic regression performance

As a benchmark model, logistic regression achieved pooled AUCs of 0.703 (95% CI 0.695–0.711) in the training data set, 0.695 (95% CI 0.679–0.711) in the internal validation data set, and 0.809 (95% CI 0.786–0.831) in the external validation data set. Using a fixed decision threshold calibrated on the training data set by maximizing the Youden index (mean threshold 0.340 across imputations), corresponding sensitivities, specificities, and balanced accuracies were 0.573, 0.725, and 0.649 (training); 0.564, 0.725, and 0.645 (internal validation); and 0.862, 0.578, and 0.720 (external validation), respectively. The PPV and NPV were 0.465 and 0.804 (training), 0.461 and 0.801 (internal validation), and 0.425 and 0.921 (external validation), respectively. Pooled F1 scores were 0.513 in the training data set, 0.506 in the internal validation data set, and 0.569 in the external validation data set.

Sensitivity analyses

In the first sensitivity analysis when the primary model was applied to the internal validation data set restricted to patients with CDAI scores >10 at baseline, the XGBoost model retained similar discrimination to the primary analysis, with an AUC of 0.720. The same 10 predictors composed the top SHAP features in both the primary and CDAI score >10 analyses, although their relative ranking varied (Supplementary Figure S1). In the second sensitivity analysis using baseline CDAI scores as the only predictor, discrimination was substantially lower (AUC 0.619 in internal validation) than both the full model and the reduced model including the top 10 predictors (AUC 0.759 and AUC 0.731, respectively). Finally, in the third sensitivity analysis enforcing patient‐level data splitting, discrimination in internal validation remained comparable to the primary treatment course level analysis (AUC 0.749).

Subgroup analysis by baseline b/ts‐DMARD therapy class

When stratified by baseline b/ts‐DMARD therapy class, model discrimination remained broadly consistent across treatment groups (Table 2). In the internal validation data set, AUC values ranged from 0.693 among patients initiating JAKi to 0.814 among those initiating RTX. In the external validation data set, AUC values ranged from 0.725 for RTX to 0.855 for ABA, with greater variability observed in smaller subgroups. Balanced accuracy and sensitivity–specificity profiles showed expected variability across therapy classes but remained within a similar range overall (Table 2).

Table 2.

Performance of the final XGBoost model in the internal and external validation data sets, stratified by baseline biologic/targeted synthetic disease‐modifying antirheumatic drug therapy class*

Treatment class Dataset AUC Sensitivity Specificity PPV NPV Balanced accuracy
ABA Internal 0.772 0.492 0.856 0.453 0.874 0.674
External 0.855 0.826 0.686 0.409 0.938 0.756
IL‐6i Internal 0.753 0.497 0.819 0.518 0.806 0.658
External 0.788 0.747 0.732 0.487 0.894 0.739
JAKi Internal 0.693 0.600 0.677 0.497 0.761 0.638
External 0.826 0.871 0.654 0.438 0.943 0.763
RTX Internal 0.814 0.617 0.862 0.590 0.874 0.739
External 0.725 0.429 0.785 0.300 0.864 0.607
TNFi Internal 0.758 0.636 0.757 0.531 0.828 0.696
External 0.776 0.836 0.563 0.476 0.879 0.700

ABA, abatacept; AUC, area under curve; IL‐6i, interleukin‐6 inhibitors; JAKi, JAK inhibitors; NPV, negative predictive value; PPV, positive predictive value; RTX, rituximab; TNFi, tumor necrosis factor inhibitors.

SHAP‐based subgroup analyses showed a highly consistent set of influential predictors across baseline b/ts‐DMARD therapy classes (computed within therapy class in the training data set). Eight predictors were present in the top 10 for all therapy classes, with only minor variation in relative ranking. These predictors were PtGA, TJC28, HAQ‐DI score, previous exposure to b/ts‐DMARDs, PhGA, age, baseline CDAI score, and CRP level (Supplementary Table S1).

DISCUSSION

Our study evaluated the capacity of an ML model trained on baseline demographic and clinical data from 11 international registries to predict RA remission between 6 and 24 months following initiation or switching of b/ts‐DMARD therapy. Using XGBoost and a large multinational data set from the JAK‐pot collaboration, we developed and externally validated a model that achieved an AUC of 0.797 and, when using a threshold‐optimized decision rule, yielded a high NPV >0.90. A simplified model using 10 baseline variables performed similarly, supporting practical use in research and exploratory settings.

Although the model showed robust performance in excluding remission as reflected in high NPV across internal and external validation, partly reflecting outcome prevalence, its precision in identifying true remitters (PPV 0.45–0.53) remained limited. These PPV values only modestly exceeded the baseline prevalence of remission (26%–29%). The observed sensitivity and specificity of the model highlight the difficulty of predicting complex treatment outcomes using only baseline clinical and demographic data.

In this context, the clinical value of the model lies in providing individualized estimates of remission probability at the time of treatment initiation or switching using routinely collected baseline clinical data as a research and decision‐support aid for probabilistic risk stratification. Prospective evaluation and calibration assessment would be required before clinical implementation. The model is intended to leverage information already available in routine care, rather than to maximize theoretical predictive performance. Within this defined scope, the model provides probabilistic information, particularly for ruling out future remission, while avoiding claims of treatment selection or causal inference.

These findings underscore a recurring pattern in clinical ML applications. When input features are limited to routine baseline demographic and clinical data, even advanced models may not substantially outperform simpler statistical approaches. 33 In our study, on the external validation data set, logistic regression achieved an AUC of 0.809 compared with 0.802 for the simplified XGBoost model, highlighting broadly similar discrimination despite differing trade‐offs in sensitivity and specificity when evaluated using the same training‐calibrated decision threshold to ensure fair comparison. This suggests that the predictive signal appeared largely linear and additive, with only modest opportunity for higher‐order or nonlinear interactions for the ML model to exploit. Rather than reflecting a limitation of the ML approach itself, these results highlight intrinsic constraints of the available input data. Accordingly, further gains in prediction are likely to depend less on model complexity and more on the identification of novel predictors of remission, such as molecular biomarkers, imaging‐derived features, or environmental exposures. 8 , 9 , 10 , 11 Recent work from the STRAP trial supports this, showing that synovial transcriptomic signatures can predict treatment response with good accuracy. 34

Several of the most influential predictors were measures of baseline disease activity, including components of CDAI, which is clinically expected given their central role in RA assessment. Importantly, several robustness analyses indicate that model discrimination was not driven by baseline disease activity alone. Baseline remission was uncommon and did not contribute to outcome definition, which relied exclusively on follow‐up visits between 6 and 24 months. Discrimination remained similar in patients with CDAI scores >10 at baseline (AUC 0.720), and the set of top predictors was consistent, supporting that performance was not driven by baseline remission or CDAI score alone.

In the primary analysis, the most influential predictors were PtGA and TJC28, previous exposure to b/ts‐DMARDs, HAQ‐DI score, PhGA, age, CDAI score, CRP level, patient‐reported pain, and anti‐CCP antibodies. These predictors are clinically plausible and consistent with prior remission literature, highlighting that the predictive signal of the model is largely driven by baseline disease activity and functional status. 35 Consistent with established evidence, patients with lower baseline joint counts, global assessments, disability scores, and markers of inflammation have a higher likelihood of remission, whereas previous biologic exposure and seropositivity are associated with reduced likelihood. 35 In contrast, baseline b/ts‐DMARD therapy class contributed minimally to remission prediction, which likely reflects heterogeneity of response across therapies and the nonrandomized assignment of treatments in real‐world cohorts. This observation was further supported by therapy class–stratified analyses, which showed broadly comparable model discrimination across baseline treatment groups and largely overlapping key predictors of remission. Although discrimination and predictor importance were broadly similar across baseline b/ts‐DMARD therapy classes, these analyses are descriptive and do not imply causal equivalence across mechanisms of action in this nonrandomized setting.

Our findings extend and refine prior work on ML‐based prediction models for RA. Norgeot et al 36 developed a longitudinal deep learning model using electronic health record data to predict short‐term disease activity across health care systems, addressing a clinical context distinct from baseline prediction of long‐term remission at treatment initiation or switching. Although Guan et al 8 and Tao et al 9 demonstrated good accuracy using genetic features, their reliance on genetic data, lack of external validation, and use of the Disease Activity Score in 28 joints (DAS28) as a remission outcome limit clinical applicability, particularly given the reduced reliability of DAS28 for this purpose. 1 , 37 Similarly, Rivellese et al 10 and Yoosuf et al 11 explored gene expression models for prediction of clinical response to TNFi, RTX, and tocilizumab, but these approaches are costly and impractical for routine use. In contrast, studies by Koo et al 12 and Alsaber et al 13 used routine clinical data to predict remission but lacked external validation and were confined to single registries in Korea and Kuwait, respectively, leaving generalizability unaddressed. These studies also relied on DAS28 as the outcome, further limiting their relevance for remission prediction. Although Sonomoto et al 14 developed a cost‐effective model for TNFi remission, its six‐month prediction horizon and lack of external validation reduce its broader relevance. Similarly, Venerito et al 15 developed models with small sample sizes, DAS28‐defined outcomes, and limited treatment scopes, such as single TNFi agents, reducing their clinical utility. Our study addresses these limitations methodologically by demonstrating the feasibility of ML implementation across international registry data sets. External validation in a distinct national cohort (Switzerland) adds evidence supporting external validity, though we acknowledge performance differences likely stem from variation in baseline disease activity, treatment patterns, and data completeness.

Several limitations warrant emphasis. First, reliance on registry data means that outcome capture is imperfect and potentially biased toward patients with regular follow‐up. Second, the model includes only baseline data, excluding early treatment response dynamics. Third, remission was defined as achieving a CDAI score ≤2.8 at least once between 6 and 24 months, estimating the probability of response occurrence rather than sustained disease control, and nearly half of responders subsequently had a visit with a CDAI score >2.8 before month 24. Fourth, although some patients contributed multiple treatment courses, model performance was robust in a patient‐level split sensitivity analysis, suggesting that repeated courses did not materially influence discrimination. Fifth, the external validation data set appeared to have milder disease at baseline among courses with available measurements, which may have attenuated observed model performance. Sixth, model development and interpretation were constrained to structured variables, limiting exploration of time‐series or unstructured data that may hold additional predictive value. Finally, although corrected HAQ‐DI values were used for the external validation data set (Switzerland), structural differences in the questionnaire remain. The absence of items related to specific assistive devices likely leads to underestimation in the eating, walking, and hygiene domains. This should be considered when interpreting HAQ‐DI values and external validation results.

In summary, our findings demonstrate that ML models trained on routine baseline clinical data provide limited predictive performance, with greatest utility in ruling out future RA remission, but are unlikely to deliver clinically actionable precision without richer input features or more informative longitudinal data. Future work should focus on integrating biomarkers, imaging, and patient‐reported trajectories to enable ML models to realize their full potential in RA precision medicine.

AUTHOR CONTRIBUTIONS

All authors contributed to at least one of the following manuscript preparation roles: conceptualization AND/OR methodology, software, investigation, formal analysis, data curation, visualization, and validation AND drafting or reviewing/editing the final draft. As corresponding author, Dr Salis confirms that all authors have provided the final approval of the version to be published and takes responsibility for the affirmations regarding article submission (eg, not under consideration by another journal), the integrity of the data presented, and the statements regarding compliance with institutional review board/Declaration of Helsinki requirements.

Supporting information

Disclosure form.

ART-78-2232-s001.pdf (12.1MB, pdf)

Figure S1: Top ten predictors of remission probability ranked by mean absolute SHAP (SHapley Additive exPlanations) values in the restricted cohort of patients with Clinical Disease Activity Index (CDAI) > 10 at baseline. Higher values indicate greater overall contribution of the predictor to the output of the model, irrespective of whether the effect increases or decreases remission probability. The ranking of predictors was similar to that observed in the primary analysis. Anti‐CCP: Anti‐Cyclic Citrullinated Peptide antibodies; b/ts‐DMARD: biologic/targeted synthetic Disease‐Modifying Antirheumatic Drug; HAQ‐DI: Health Assessment Questionnaire Disability Index; VAS: Visual Analogue Scale.

Table S1: Top ten predictors of remission probability ranked by mean absolute SHAP (SHapley Additive exPlanations) values, stratified by baseline biologic or targeted synthetic DMARD (b/ts‐DMARD) therapy class. Predictor rankings were derived from the training dataset using the final XGBoost model. Higher mean |SHAP| values indicate greater overall contribution to model predictions, regardless of the direction of effect. ABA: Abatacept; Anti‐CCP: Anti‐cyclic Citrullinated Peptide antibodies; CDAI: Clinical Disease Activity Index; IL‐6i: Interleukin‐6 inhibitors; JAKi: Janus kinase inhibitors; RITX: Rituximab; TNFi: Tumor Necrosis Factor inhibitors; HAQ‐DI: Health Assessment Questionnaire Disability Index; VAS: Visual Analogue Scale.

ART-78-2232-s002.docx (63.7KB, docx)

ACKNOWLEDGMENTS

We gratefully acknowledge the contribution of Romain Aymon to earlier stages of this project. Romain Aymon passed away before the final manuscript was completed. We also thank all the health professionals and patients who participated in the registries. Open access publication funding provided by COUPERIN CY26.

The JAK‐pot collaboration is supported by unconditional/unrestricted research grants from AbbVie, Alfasigma S.p.A., and Eli Lilly and was previously supported by Galapagos BV and Pfizer. The sponsors had no role in the study design or in the collection, analysis, or interpretation of the data, the writing of the manuscript, or the decision to submit the manuscript for publication. Publication of this article was not contingent upon approval of sponsors. The BIOREG registry is supported by Amgen, Astra Zeneca Astro Pharma, Biogen, Eli Lilly, Galapagos Biopharma Austria, Pfizer, Swedish Orphan Biovitrium, and UCB Pharma. Clinical work in Czech Republic was partially supported by the project from the Ministry of Health for conceptual development of research organization (MZ00023728023728, Institute of Rheumatology). BioRx.si has received funding for clinical research paid to Društvo za razvoj revmatologije from AbbVie, Roche, Medis, MSD, Biogen, Amgen, Sanofi, Celgene, and Pfizer. The GISEA registry has received financial support from AbbVie and Alfasigma. NOR‐DMARD was supported with research funding to Diakonhjemmet Hospital from AbbVie, Amgen, BMS, MSD, Novartis, Pfizer, and UCB. PT is supported by unrestricted grants from AbbVie, Biogen, Celgene, MSD, Roche, Sanofi, and Pfizer. ROB‐FIN is funded by BMS, MSD, Pfizer, Roche, and UCB. The Romanian Registry of Rheumatic Diseases uses unrestricted grants from AbbVie, Astra Zeneca, Boehringer Ingelheim, Ewopharma, Novartis, Sandoz, Sobi, Stada, and UCB. The Swiss Clinical Quality Management in Rheumatic Diseases database is sponsored by public and industrial support (https://scqm.ch/en/sponsoren/).

1University of Montpellier and Physiology and Experimental Medicine of the Heart and Muscles, Montpellier, France; 2Division of Rheumatology, Geneva University Hospital, Geneva, Switzerland; 3Geneva Center for Inflammation Research, Faculty of Medicine, University of Geneva, Geneva, Switzerland; 4Centre Hospitalier de l'Université de Montréal, Institut de Recherche en Rhumatologie, Montréal, Québec, Canada; 5Center of Rheumatic Diseases, University of Medicine and Pharmacy, Bucharest, Romania; 6Rheumatology Unit, Department of Precision and Regenerative Medicine and Ionian Area, University of Bari Aldo Moro, Bari, Italy; 7Department of Clinical Sciences and Community Health, Università degli Studi di Milano and Rheumatology Unit, ASST G. Pini‐CTO, Milan, Italy; 8Center for Treatment of Rheumatic and Musculoskeletal Diseases, Diakonhjemmet Hospital, Oslo, Norway; 9Biologica Registry BioReg, Stockerau, Austria; 10Health Care Center Mariahilf of ÖGK, Vienna, Austria; 11Rheumatology, Sigmund Freud University, Vienna, Austria; 12Departments of Medicine and Rheumatology, Helsinki University Hospital, The National Register of Antirheumatic and Biological Medicines in Finland, Helsinki, Finland; 13Institute of Rheumatology, Prague, Czech Republic; 14NOVA Medical School, Universidade Nova de Lisaboa and Reuma.pt, Sociedade Portuguesa de Reumatologia, Lisbon, Portugal; 15Department of Rheumatology, University Medical Centre Ljubljana, Ljubljana, Slovenia; 16Faculty of Medicine, University of Ljubljana, Ljubljana, Slovenia; 17Rheumatology and Clinical Immunology, Medical School, University of Crete, Rethymno, Crete, Greece; 18Institute of Molecular Biology and Biotechnology, Foundation for Research and Technology Hellas, Heraklion, Crete, Greece.

Additional supplementary information cited in this article can be found online in the Supporting Information section (https://acrjournals.onlinelibrary.wiley.com/doi/10.1002/art.70165).

Author disclosures and graphical abstract are available at https://onlinelibrary.wiley.com/doi/10.1002/art.70165.

References

  • 1. Smolen JS, Aletaha D, Barton A, et al. Rheumatoid arthritis. Nat Rev Dis Primers 2018;4(1):18001. doi: 10.1038/nrdp.2018.1 [DOI] [PubMed] [Google Scholar]
  • 2. Ostor AJ, Sawant R, Qi CZ, et al. Value of remission in patients with rheumatoid arthritis: a targeted review. Adv Ther 2022;39(1):75–93. doi: 10.1007/s12325-021-01946-w [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Pierreisnard A, Issa N, Barnetche T, et al. Meta‐analysis of clinical and radiological efficacy of biologics in rheumatoid arthritis patients naive or inadequately responsive to methotrexate. Joint Bone Spine 2013;80(4):386–392. doi: 10.1016/j.jbspin.2012.09.023 [DOI] [PubMed] [Google Scholar]
  • 4. Combe B, Rincheval N, Benessiano J, et al. Five‐year favorable outcome of patients with early rheumatoid arthritis in the 2000s: data from the ESPOIR cohort. J Rheumatol 2013;40(10):1650–1657. doi: 10.3899/jrheum.121515 [DOI] [PubMed] [Google Scholar]
  • 5. Lin CMA, Cooles FAH, Isaacs JD. Precision medicine: the precision gap in rheumatic disease. Nat Rev Rheumatol 2022;18(12):725–733. doi: 10.1038/s41584-022-00845-w [DOI] [PubMed] [Google Scholar]
  • 6. Lee S, Kang S, Eun Y, et al. Machine learning‐based prediction model for responses of bDMARDs in patients with rheumatoid arthritis and ankylosing spondylitis. Arthritis Res Ther 2021;23(1):254. doi: 10.1186/s13075-021-02635-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science 2015;349(6245):255–260. doi: 10.1126/science.aaa8415 [DOI] [PubMed] [Google Scholar]
  • 8. Guan Y, Zhang H, Quang D, et al. Machine learning to predict anti‐tumor necrosis factor drug responses of rheumatoid arthritis patients by integrating clinical and genetic markers. Arthritis Rheumatol 2019;71(12):1987–1996. doi: 10.1002/art.41056 [DOI] [PubMed] [Google Scholar]
  • 9. Tao W, Concepcion AN, Vianen M, et al. Multiomics and machine learning accurately predict clinical response to adalimumab and etanercept therapy in patients with rheumatoid arthritis. Arthritis Rheumatol 2021;73(2):212–222. doi: 10.1002/art.41516 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Rivellese F, Surace AEA, Goldmann K, et al; R4RA collaborative group. Rituximab versus tocilizumab in rheumatoid arthritis: synovial biopsy‐based biomarker analysis of the phase 4 R4RA randomized trial. Nat Med 2022;28(6):1256–1268. doi: 10.1038/s41591-022-01789-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Yoosuf N, Maciejewski M, Ziemek D, et al. Early prediction of clinical response to anti‐TNF treatment using multi‐omics and machine learning in rheumatoid arthritis. Rheumatology (Oxford) 2022;61(4):1680–1689. doi: 10.1093/rheumatology/keab521 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Koo BS, Eun S, Shin K, et al. Machine learning model for identifying important clinical features for predicting remission in patients with rheumatoid arthritis treated with biologics. Arthritis Res Ther 2021;23(1):178. doi: 10.1186/s13075-021-02567-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Alsaber AR, Al‐Herz A, Alawadhi B, 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 2024;7:1406365. doi: 10.3389/fdata.2024.1406365 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Sonomoto K, Fujino Y, Tanaka H, et al. A machine learning approach for prediction of CDAI remission with TNF inhibitors: a concept of precision medicine from the FIRST Registry. Rheumatol Ther 2024;11(3):709–736. doi: 10.1007/s40744-024-00668-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Venerito V, Angelini O, Fornaro M, et al. A machine learning approach for predicting sustained remission in rheumatoid arthritis patients on biologic agents. J Clin Rheumatol 2022;28(2):e334–e339. doi: 10.1097/RHU.0000000000001720 [DOI] [PubMed] [Google Scholar]
  • 16. Wei M, Chu CQ. Prediction of treatment response: personalized medicine in the management of rheumatoid arthritis. Best Pract Res Clin Rheumatol 2022;36(1):101741. doi: 10.1016/j.berh.2021.101741 [DOI] [PubMed] [Google Scholar]
  • 17. Plant D, Barton A. Machine learning in precision medicine: lessons to learn. Nat Rev Rheumatol 2021;17(1):5–6. doi: 10.1038/s41584-020-00538-2 [DOI] [PubMed] [Google Scholar]
  • 18. Chen T, Guestrin C. Xgboost: a scalable tree boosting system. In: KDD '16: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery; 2016:785–794. doi: 10.1145/2939672.2939785 [DOI] [Google Scholar]
  • 19. Inoue T, Ichikawa D, Ueno T, et al. XGBoost, a machine learning method, predicts neurological recovery in patients with cervical spinal cord injury. Neurotrauma Rep 2020;1(1):8–16. doi: 10.1089/neur.2020.0009 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Nguyen H, Bui XN, Bui HB, et al. Developing an XGBoost model to predict blast‐induced peak particle velocity in an open‐pit mine: a case study. Acta Geophys 2019;67(2):477–490. doi: 10.1007/s11600-019-00268-4 [DOI] [Google Scholar]
  • 21. Sheridan RP, Wang WM, Liaw A, et al. Extreme gradient boosting as a method for quantitative structure‐activity relationships. J Chem Inf Model 2016;56(12):2353–2360. doi: 10.1021/acs.jcim.6b00591 [DOI] [PubMed] [Google Scholar]
  • 22. Nam SM, Peterson TA, Seo KY, et al. Discovery of depression‐associated factors from a nationwide population‐based survey: epidemiological study using machine learning and network analysis. J Med Internet Res 2021;23(6):e27344. doi: 10.2196/27344 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Ukalovic D, Leeb BF, Rintelen B, et al. Prediction of ineffectiveness of biological drugs using machine learning and explainable AI methods: data from the Austrian Biological Registry BioReg. Arthritis Res Ther 2024;26(1):44. doi: 10.1186/s13075-024-03277-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Shwartz‐Ziv R, Armon A. Tabular data: deep learning is not all you need. Inf Fusion 2022;81:84–90. doi: 10.1016/j.inffus.2021.11.011 [DOI] [Google Scholar]
  • 25. Grinsztajn L, Oyallon E, Varoquaux G. Why do tree‐based models still outperform deep learning on typical tabular data? Paper presented at: 36th Conference on Neural Information Processing Systems (NeurIPS 2022) Track on Datasets and Benchmarks; November 28, 2022; New Orleans, LA.
  • 26. Salehi F, Salin E, Smarr B, et al. A robust machine learning approach to predicting remission and stratifying risk in rheumatoid arthritis patients treated with bDMARDs. Sci Rep 2025;15(1):23960. doi: 10.1038/s41598-025-09975-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: NIPS'17: Proceedings of the 31st International Conference on Neural Information Processing Systems. Curran Associates Inc; 2017:4768–4777. [Google Scholar]
  • 28. Lauper K, Iudici M, Mongin D, et al. Effectiveness of TNF‐inhibitors, abatacept, IL6‐inhibitors and JAK‐inhibitors in 31 846 patients with rheumatoid arthritis in 19 registers from the ‘JAK‐pot’ collaboration. Ann Rheum Dis 2022;81(10):1358–1366. doi: 10.1136/annrheumdis-2022-222586 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Pombo‐Suarez M, Sanchez‐Piedra C, Gómez‐Reino J, et al. After JAK inhibitor failure: to cycle or to switch, that is the question – data from the JAK‐pot collaboration of registries. Ann Rheum Dis 2023;82(2):175–181. doi: 10.1136/ard-2022-222835 [DOI] [PubMed] [Google Scholar]
  • 30. Aymon R, Mongin D, Bergstra SA, et al. Evaluation of discontinuation for adverse events of JAK inhibitors and bDMARDs in an international collaboration of rheumatoid arthritis registers (the ‘JAK‐pot’ study). Ann Rheum Dis 2024;83(4):421–428. doi: 10.1136/ard-2023-224670 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Aletaha D, Nell VP, Stamm T, et al. Acute phase reactants add little to composite disease activity indices for rheumatoid arthritis: validation of a clinical activity score. Arthritis Res Ther 2005;7(4):R796–R806. doi: 10.1186/ar1740 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Aletaha D, Smolen J. The Simplified Disease Activity Index (SDAI) and the Clinical Disease Activity Index (CDAI): a review of their usefulness and validity in rheumatoid arthritis. Clin Exp Rheumatol 2005;23(5)(suppl 39):S100–S108. [PubMed] [Google Scholar]
  • 33. Christodoulou E, Ma J, Collins GS, et al. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol 2019;110:12–22. doi: 10.1016/j.jclinepi.2019.02.004 [DOI] [PubMed] [Google Scholar]
  • 34. Lewis MJ, Çubuk C, Surace AEA, et al; STRAP collaborative group. Deep molecular profiling of synovial biopsies in the STRAP trial identifies signatures predictive of treatment response to biologic therapies in rheumatoid arthritis. Nat Commun 2025;16(1):5374. doi: 10.1038/s41467-025-60987-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Sauve E, Barnabe C. The usual suspects: established and emerging predictor variables for remission in rheumatoid arthritis. J Rheumatol 2026;53(2):126–133. doi: 10.3899/jrheum.2025-0476 [DOI] [PubMed] [Google Scholar]
  • 36. Norgeot B, Glicksberg BS, Trupin L, et al. Assessment of a deep learning model based on electronic health record data to forecast clinical outcomes in patients with rheumatoid arthritis. JAMA Netw Open 2019;2(3):e190606. doi: 10.1001/jamanetworkopen.2019.0606 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Aletaha D, Wang X, Zhong S, et al. Differences in disease activity measures in patients with rheumatoid arthritis who achieved DAS, SDAI, or CDAI remission but not Boolean remission. Semin Arthritis Rheum 2020;50(2):276–284. doi: 10.1016/j.semarthrit.2019.09.005 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Disclosure form.

ART-78-2232-s001.pdf (12.1MB, pdf)

Figure S1: Top ten predictors of remission probability ranked by mean absolute SHAP (SHapley Additive exPlanations) values in the restricted cohort of patients with Clinical Disease Activity Index (CDAI) > 10 at baseline. Higher values indicate greater overall contribution of the predictor to the output of the model, irrespective of whether the effect increases or decreases remission probability. The ranking of predictors was similar to that observed in the primary analysis. Anti‐CCP: Anti‐Cyclic Citrullinated Peptide antibodies; b/ts‐DMARD: biologic/targeted synthetic Disease‐Modifying Antirheumatic Drug; HAQ‐DI: Health Assessment Questionnaire Disability Index; VAS: Visual Analogue Scale.

Table S1: Top ten predictors of remission probability ranked by mean absolute SHAP (SHapley Additive exPlanations) values, stratified by baseline biologic or targeted synthetic DMARD (b/ts‐DMARD) therapy class. Predictor rankings were derived from the training dataset using the final XGBoost model. Higher mean |SHAP| values indicate greater overall contribution to model predictions, regardless of the direction of effect. ABA: Abatacept; Anti‐CCP: Anti‐cyclic Citrullinated Peptide antibodies; CDAI: Clinical Disease Activity Index; IL‐6i: Interleukin‐6 inhibitors; JAKi: Janus kinase inhibitors; RITX: Rituximab; TNFi: Tumor Necrosis Factor inhibitors; HAQ‐DI: Health Assessment Questionnaire Disability Index; VAS: Visual Analogue Scale.

ART-78-2232-s002.docx (63.7KB, docx)

Data Availability Statement

The JAK‐pot collaboration is based on data contributions from multiple rheumatology registers. Although the data are hosted and analyzed centrally in Geneva, they remain the property of the individual contributing registers. As such, data cannot be shared by the collaboration. Any external data request must be addressed directly to each contributing register and requires approval from all data owners. The R codes for the models’ creation can be found at https://github.com/Zube-Geneve/RA_Remission_JAKPOT.git.


Articles from Arthritis & Rheumatology (Hoboken, N.j.) are provided here courtesy of Wiley

RESOURCES