A robust machine learning approach to predicting remission and stratifying risk in rheumatoid arthritis patients treated with bDMARDs.

Rheumatoid arthritis (RA) is a chronic autoimmune disease affecting millions worldwide, leading to inflammation, joint damage, and reduced quality of life. Although biological disease-modifying antirheumatic drugs (bDMARDs) are effective, they are costly, and up to 40% of patients do not achieve remission within six months. Accurate prediction of treatment response is crucial for optimizing care, minimizing side effects, and enhancing cost efficiency. This study proposes a robust machine learning framework for predicting six-month remission in RA patients using baseline routine clinical data. The framework also integrates risk stratification and explainability to enhance its clinical applicability. We evaluated multiple machine learning models, AdaBoost, Random Forest, XGBoost, and Support Vector Machines, using data from Austrian RA patients. We externally validated the results on an independent dataset from the Erlangen Hospital. To improve the reliability of probability estimates for actionable risk stratification, we employed calibration techniques, including Platt scaling, Isotonic regression, Beta calibration, and Spline calibration. We generated calibration curves to assess and visualize the alignment between predicted probabilities and observed outcomes. In addition, we used SHapley Additive exPlanations (SHAP) to analyze the contributions of different patient characteristics to the prediction of RA remission. AdaBoost demonstrated stronger performance than the other models, achieving an accuracy of 85.71% and a Brier score of 0.13 with isotonic regression calibration. SHAP identified DAS28, visual analog scales (VAS), age, and swollen joint count (SJC) as important characteristics for the prediction of RA remission. We also stratified patients into low-, medium-, and high-risk categories based on model predictions to support follow-up scheduling and treatment prioritization. Our framework predicts RA remission before the initiation of bDMARD therapy. It enables personalized care, actionable risk stratification, and optimized resource allocation. Its robustness was validated on two different individual cohort datasets, which highlights its potential for integration into routine clinical workflows.