Machine learning-enhanced clinical and ultrasound technology for early detection of difficult-to-treat rheumatoid arthritis

Objectives : This study aimed to examine the clinical features, serological indicators, and ultrasound examination results of individuals with difficult-to-treat rheumatoid arthritis (D2T RA) and non-D2T RA, along with the development and validation of two distinct predictive models for the early detection of D2T-RA. Methods : Enrolled 85 patients with D2T-RA diagnosed as moderate or high disease activity who completed 24 months of follow-up, and randomly matched 83 non-D2T-RA patients with moderate to high disease activity. Baseline clinical information was collected, and US examinations were performed to record the single scores of gray-scale (GS) and power Doppler (PD) for 16 joints and 10 tendons, as well as the EULAR-OMERACT scores. Univariate analysis identified predictive factors, followed by machine learning to create two models: clinical/serological (Model 1) and clinical/serological/US (Model 2). We evaluated the model performance using 5-fold cross-validation, utilizing the F1 score and AUC. Results : The univariate logistic analysis showed that EULAR-OMERACT＞1 (6 variables) and clinical and serological characteristics (14 variables) were significant predictors of D2T RA. The random forest model performed best on all models, with the AUC and F1 of test set model 1 divided into 0.81 and 0.67, and the AUC and FI of model 2 with ultrasound data increased to 0.83 and 0.69, respectively. Conclusion : Multi-joint ultrasound score provides important prediction data for early identification of D2T RA, a random forest model improves prediction efficacy, and evaluating limited joints makes this method more feasible in rheumatism clinical practice.