Predicting Disease Activity Score in Rheumatoid Arthritis Patients Treated with Biologic Disease-Modifying Antirheumatic Drugs Using Machine Learning Models

Fatemeh Salehi, Sudiyeh Hejri Zarifi, Sara Bayat, M. Habibpour, Amirreza Asemanrafat, Arnd Kleyer, Georg Schett, Ruth Fritsch‐Stork, Bjoern M. EskofierTechnologies 2025 · DOI

TL;DR

This paper introduces a machine learning framework to predict Rheumatoid Arthritis (RA) disease activity 12 months after starting biologic treatments. By modeling disease activity as a continuous score rather than a binary remission/non-remission outcome, their Ridge regression model offers a more flexible and clinically applicable tool for personalized treatment planning. Crucially, the model demonstrates robust generalization across different clinical settings through external validation.

What Problem Does This Paper Address?

Rheumatoid Arthritis (RA) is a chronic autoimmune disease that causes joint inflammation and progressive disability. While biologic disease-modifying antirheumatic drugs (bDMARDs) have greatly improved treatment outcomes, predicting which patients will respond best to these therapies remains a significant challenge. Current methods often require months of treatment to assess effectiveness, during which non-responders may experience continued inflammation, side effects, and financial burdens. Early prediction of treatment response could lead to more personalized and effective care, avoiding ineffective treatments and improving patient outcomes. Existing machine learning approaches often simplify the prediction task to binary classification (remission vs. non-remission), which can be less nuanced than desired, especially as remission criteria can evolve and vary. Furthermore, many models lack rigorous external validation, limiting their real-world applicability.

Methods

The researchers employed a machine learning approach using eight regression models: Ridge, Lasso, Support Vector Regression (SVR), Random Forest, Decision Tree, K-Nearest Neighbors (KNN) Regression, AdaBoost, and XGBoost. The core innovation lies in predicting the continuous Disease Activity Score for 28 joints with C-reactive protein (DAS28-CRP) at 12 months, rather than a binary remission status. This continuous prediction allows for more granular insights into disease activity.

To ensure robust model development and prevent overfitting, a nested cross-validation strategy was used. This involved an outer loop for generalization performance estimation (5-fold cross-validation) and an inner loop for hyperparameter optimization (3-fold cross-validation with randomized search). Model performance was primarily evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²). For interpretability, SHAP (SHapley Additive exPlanations) values were used to identify the most influential features.

Datasets

The study utilized two datasets:

  • Internal Dataset: Clinical data from 154 RA patients treated at University Hospital Erlangen, Germany, between 2010 and 2017. This dataset was used for training and internal validation.
  • External Dataset: Data from the Austrian BioReg registry, including RA patients treated with bDMARDs between 2010 and 2024. This dataset, representing a more diverse, real-world patient population from multiple clinical sites, was used for external validation to assess model generalizability.

Key Insights and Ideas

  • Continuous Prediction is Key: Predicting DAS28-CRP as a continuous score, rather than a binary remission outcome, offers greater flexibility. This allows clinicians to interpret disease activity more nuancedly, adapting to evolving remission criteria and individual patient needs.
  • Ridge Regression Shines: The regularized linear model, Ridge regression, consistently outperformed other models, both internally and externally. Its ability to shrink but retain all features is likely beneficial for handling collinearity in clinical data.
  • Generalizability Matters: The study strongly emphasizes the importance of external validation. The fact that Ridge regression maintained solid performance on the independent BioReg dataset highlights its potential for real-world application across different clinical settings.
  • Clinical Features are Powerful: The top predictive features identified (patient and physician VAS, ESR, TJC28, SJC28) are routinely collected in clinical practice, making the model practical for immediate implementation.

Key Results

  • Internal Performance: The Ridge regression model achieved the best internal performance with an MAE of 0.633 and an R² of 0.542.
  • External Validation: On the independent BioReg dataset, Ridge regression showed strong generalizability with an MAE of 0.678 and an R² of 0.491.
  • Classification Performance: When binarized for remission, Ridge achieved an accuracy of 0.818 and an F1 score of 0.800 internally, and 0.773 accuracy and 0.762 F1 score externally.
  • Fairness: Ridge and SVR demonstrated the most balanced performance across age groups, with minimal accuracy gaps between younger and older patients, suggesting they are not only generalizable but also fairer.
  • Feature Importance: Patient's global health assessment (VAS), ESR, tender joint count (TJC28), physician's global assessment (VAS), and swollen joint count (SJC28) were consistently the top five most influential features across both cohorts.

Limitations and Future Work

  • Sample Size: The external validation cohort had a relatively modest sample size, which could impact generalizability and model stability.
  • Baseline Data Only: The models were trained solely on baseline clinical features. Incorporating early follow-up data (e.g., 3-month DAS28-CRP) could enhance predictive accuracy and allow for more adaptive treatment planning.
  • Unmeasured Confounders: The study acknowledges that important factors like smoking status, physical activity, or socioeconomic status were not included.
  • bDMARD Class Specificity: The model did not differentiate between specific bDMARD classes (e.g., TNF inhibitors vs. JAK inhibitors), which may have varying effectiveness. Future work could explore stratified modeling.
  • Prospective Validation: While retrospective performance was strong, prospective studies are needed to validate the model's clinical utility and robustness in real-world settings.
  • Diverse Populations: Further research should examine performance across more diverse patient populations.

Summary updated

Abstract

Rheumatoid arthritis (RA) is a chronic autoimmune disease marked by joint inflammation and progressive disability. While biological disease-modifying antirheumatic drugs (bDMARDs) have significantly improved disease control, predicting individual treatment response remains clinically challenging. This study presents a machine learning approach to predict 12-month disease activity, measured by DAS28-CRP, in RA patients beginning bDMARD therapy. We trained and evaluated eight regression models, including Ridge, Lasso, Support Vector Regression, and XGBoost, using baseline clinical features from 154 RA patients treated at University Hospital Erlangen. A rigorous nested cross-validation strategy was applied for internal model selection and validation. Importantly, model generalizability was assessed using an independent external dataset from the Austrian BioReg registry, which includes a more diverse, real-world RA patient population from across multiple clinical sites. The Ridge regression model achieved the best internal performance (MAE: 0.633, R2: 0.542) and showed strong external validity when applied to unseen BioReg data (MAE: 0.678, R2: 0.491). These results indicate robust cross-cohort generalization. By predicting continuous DAS28-CRP scores instead of binary remission labels, our approach supports flexible, individualized treatment planning based on local or evolving clinical thresholds. This work demonstrates the feasibility and clinical value of externally validated, data-driven tools for precision treatment planning in RA.
BibTeX
@inproceedings{salehi2025predicting,
  author = {Fatemeh Salehi and Sudiyeh Hejri Zarifi and Sara Bayat and M. Habibpour and Amirreza Asemanrafat and Arnd Kleyer and Georg Schett and Ruth Fritsch‐Stork and Bjoern M. Eskofier},
  title = {{Predicting Disease Activity Score in Rheumatoid Arthritis Patients Treated with Biologic Disease-Modifying Antirheumatic Drugs Using Machine Learning Models}},
  booktitle = {Technologies},
  year = {2025},
  doi = {10.3390/technologies13080350},
}
Figure