Background Rheumatoid arthritis (RA) treatment guidelines recommend early initiation of disease-modifying antirheumatic drugs (DMARDs), but actual prescribing decisions are influenced by multiple clinical and contextual factors. Machine learning (ML) offers a promising tool to uncover patterns in treatment selection and support personalized decision-making. Objectives To identify the most important predictors of initial treatment in patients with newly diagnosed RA using ML algorithms applied to an international registry. Methods We conducted a secondary analysis of 16,684 patients from the METEOR registry. The primary outcome was the first treatment regimen recorded. Predictors included demographics, clinical indicators, serological markers, and country of origin. Random forest models were trained on a 70/30 split of the dataset and evaluated using accuracy, precision, recall, and generalizability metrics. Variable importance was assessed via mean decrease in Gini coefficient. Results The most common treatment regimen was methotrexate plus glucocorticoids (26.1%). Age was the most important predictor of treatment class across all models. Inflammatory burden (ESR, tender/swollen joint counts, HAQ-DI) also ranked highly, while serological markers (RF, ACPA) and imaging findings (erosions) showed limited predictive value. The best-performing model (Random Forest 2) achieved an accuracy of 0.97 and demonstrated good generalizability across countries. Conclusion In routine practice, age and clinical measures of disease activity are key determinants of initial RA treatment, often outweighing serological or imaging findings. ML models can help characterize real-world decision-making patterns and inform context-aware quality improvement and hypothesis generation; prospective validation linking predictions to outcomes is needed before clinical decision-support use. <Table Float="No" ID="Taba"> <tgroup cols="2"> <colspec align="justify" colname="c1" colnum="1" /> <colspec align="justify" colname="c2" colnum="2" /> <tbody> <row> <entry nameend="c2" namest="c1"> Key Points • Machine learning revealed age and clinical disease activity as the strongest predictors of initial RA treatment • Serological and imaging markers had limited predictive value compared to clinical measures. • Real-world prescribing patterns diverged from international treatment guidelines. • Findings support data-driven, personalized approaches in early RA care. </entry> </row> </tbody> </tgroup> </Table>

Global patterns and predictors of initial treatment in early rheumatoid arthritis: insights from a multinational machine learning study

期刊论文
发表日期：

2026年3月9日

查看全文

David Vega-Morales,
Pedro Machado,
Sytske Anne Bergstra,
Wendy Orzúa-de la Fuente,
Salvador Ruiz-Correa,
Rubén López-Revilla,
Arvind Chopra,
Ana Rodrigues,
Lai Ling Winchow

摘要

Background

Rheumatoid arthritis (RA) treatment guidelines recommend early initiation of disease-modifying antirheumatic drugs (DMARDs), but actual prescribing decisions are influenced by multiple clinical and contextual factors. Machine learning (ML) offers a promising tool to uncover patterns in treatment selection and support personalized decision-making.

Objectives

To identify the most important predictors of initial treatment in patients with newly diagnosed RA using ML algorithms applied to an international registry.

Methods

We conducted a secondary analysis of 16,684 patients from the METEOR registry. The primary outcome was the first treatment regimen recorded. Predictors included demographics, clinical indicators, serological markers, and country of origin. Random forest models were trained on a 70/30 split of the dataset and evaluated using accuracy, precision, recall, and generalizability metrics. Variable importance was assessed via mean decrease in Gini coefficient.

Results

The most common treatment regimen was methotrexate plus glucocorticoids (26.1%). Age was the most important predictor of treatment class across all models. Inflammatory burden (ESR, tender/swollen joint counts, HAQ-DI) also ranked highly, while serological markers (RF, ACPA) and imaging findings (erosions) showed limited predictive value. The best-performing model (Random Forest 2) achieved an accuracy of 0.97 and demonstrated good generalizability across countries.

Conclusion

In routine practice, age and clinical measures of disease activity are key determinants of initial RA treatment, often outweighing serological or imaging findings. ML models can help characterize real-world decision-making patterns and inform context-aware quality improvement and hypothesis generation; prospective validation linking predictions to outcomes is needed before clinical decision-support use.

Key Points

• Machine learning revealed age and clinical disease activity as the strongest predictors of initial RA treatment

• Serological and imaging markers had limited predictive value compared to clinical measures.

• Real-world prescribing patterns diverged from international treatment guidelines.

• Findings support data-driven, personalized approaches in early RA care.