Predicting the progression of difficult-to-treat rheumatoid arthritis by a machine learning scoring system, from the FIRST registry

Objectives This study aimed to develop and validate a prediction model for the future progression of difficult-to-treat rheumatoid arthritis (D2T RA) and support the precise use of biologic and targeted synthetic disease-modifying antirheumatic drugs (b/tsDMARDs). Methods Data were analysed from 1221 patients with rheumatoid arthritis in the FIRST registry, a multicentre collaborative database, who initiated their first b/tsDMARD. 25 prediction models for D2T RA were generated using 43 baseline characteristics prior to initiating the first b/tsDMARD. Model performances were tested. A scoring system based on the selected model was designed and internally validated for use in routine clinical practice. Results Among the 1221 patients, 193 (15.8%) progressed to D2T RA after a median of 54.0 months. These patients had higher tender joint count, global assessment, pain scale, Health Assessment Questionnaire score and Clinical Disease Activity Index score and more frequent coexisting lung disease. Among the machine learning models tested, Lasso logistic regression performed best (area under the curve 0.71), identifying pain scale, erythrocyte sedimentation rate and coexisting lung disease as key predictors. The scoring system based on these predictors demonstrated comparable performance. Among patients identified as high risk for D2T RA by the scoring system, interleukin-6 receptor inhibitors (IL-6Ri) significantly reduced D2T RA progression (relative risk (RR): 0.48). Conclusion This study developed and validated a scoring system to predict D2T RA progression and identify patients at high risk and suggested the advantages of IL-6Ri. This system may advance precision medicine in RA and may contribute to controlling unmet needs.