Combined MSK US and machine learning model improves rheumatoid arthritis disease activity assessment

ObjectiveTo develop and evaluate a combined model integrating musculoskeletal ultrasound (MSK US) with a machine learning (ML) algorithm for assessing disease activity in rheumatoid arthritis (RA).MethodsA total of 203 patients with clinically confirmed RA were prospectively enrolled from December 2023 to September 2025. A cohort of 142 patients from the First Affiliated Hospital, College of Clinical Medicine of Henan University of Science and Technology served as the training cohort, while 61 patients from Affiliated Hospital of Traditional Chinese Medicine, Xinjiang Medical University (Fourth Clinical Medical College, Xinjiang Medical University) constituted the independent external test cohort. Three predictive models were developed: (1) an MSK US model incorporating two-dimensional grayscale ultrasound, power Doppler ultrasound (PDUS), and superb micro-vascular imaging (SMI); (2) a radiomics model based on two-dimensional grayscale images using the extremely randomized trees (ExtraTrees) algorithm; and (3) a combined model integrating the first two. Model performance in assessing RA disease activity was evaluated and compared using receiver operating characteristic (ROC) curve analysis. Calibration curves and decision curve analysis (DCA) were subsequently used to validate the overall performance of the optimal model.ResultsMultivariate logistic regression analysis identified erythrocyte sedimentation rate (ESR)>54 mm/h, C-reactive protein (CRP)>32.83 mg/L, and SMI synovial blood flow grade III as independent predictors of clinically active RA. The area under the ROC curve (AUC) values for the MSK US model, radiomics model, and combined model were 0.935 (95% confidence interval [CI]: 0.893-0.978), 0.976 (95% CI: 0.955-0.997), and 0.998 (95% CI: 0.998-1.000), respectively, in the training cohort; in the independent external test cohort, the AUC values for the three models were 0.904 (95% CI: 0.825-0.983), 0.823 (95%CI:0.714-0.933), and 0.929 (95%CI:0.866-0.992),respectively. The discriminative performance of the combined model was significantly superior to that of either the MSK US model or the radiomics model alone. Calibration curves demonstrated good agreement between the observed risk levels and the predicted risk probabilities. Decision curve analysis indicated that the model provided significant net benefit across threshold probability ranges of 0.02–0.80 in the training cohort and 0.12–0.78 in the test cohort.ConclusionThe combined model developed based on MSK US and radiomics demonstrated satisfactory performance for assessing disease activity in RA, enabling clinicians to dynamically monitor RA disease activity and evaluate treatment response, thereby providing a reliable imaging basis for the selection of routine medical treatment strategies.