Evaluation of ultrasound-based predictive models for disease relapse in rheumatoid arthritis patients in clinical remission.

Background: Patients with rheumatoid arthritis (RA) in clinical remission face a high risk of relapse, while existing prediction models primarily rely on clinical indicators and lack imaging support. Although musculoskeletal ultrasound can effectively detect synovitis, current scanning protocols are limited by strong heterogeneity and operational complexity. This study aims to develop and validate a simplified ultrasound-clinical combined model for predicting relapse in RA patients during clinical remission. Methods: This prospective cohort study enrolled RA patients in clinical remission from a single center between August 2022 and August 2024. Participants underwent comprehensive clinical and ultrasound assessments (36 joints and 36 tendons) at baseline. Disease relapse was defined as an increase in Disease Activity Score in 28 Joints (DAS28) >0.6 from the lowest recorded value or DAS28 >2.6. Patients were randomly allocated to development (n=222) and validation (n=110) sets. We constructed three prediction models: Model I (clinical indicators only), Model II (clinical + 4 key joint ultrasound markers), and Model III (clinical + all 36 joint ultrasound markers). Model development employed the Light Gradient Boosting Machine (LightGBM) algorithm with Bayesian optimization. Results: Among the 332 analyzed patients (from 402 initially eligible), 76 (22.9%) experienced relapse during follow-up. Multivariate analysis identified eight independent predictors: disease duration, clinical remission duration, high anti-cyclic citrullinated peptide antibody (Anti-CCP) positivity, positive hand/foot X-ray findings, and ultrasound-detected synovitis in the wrist, the metacarpophalangeal joint 2 (MCP2), the knee, and the tibialis posterior (TP) tendon. Model II demonstrated excellent performance with area under the curve (AUC) =0.865 [95% confidence interval (CI): 0.816-0.914], significantly outperforming Model I (AUC =0.755, P<0.001) and approaching Model III (AUC =0.903). The simplified model showed superior clinical utility with a net reclassification index (NRI) of 0.36 (95% CI: 0.16-0.55) and an integrated discrimination improvement (IDI) =0.17 (95% CI: 0.12-0.23) compared to Model I, while reducing scanning time from 29.0 to 10.5 min (64% reduction, P<0.001). Conclusions: The simplified ultrasound-clinical model maintains high predictive accuracy while offering substantially improved clinical feasibility for identifying RA patients at high risk of relapse during clinical remission.