Predicting the risk of bone erosion in rheumatoid arthritis using a SHAP-based interpretable machine learning model.
Authors
Affiliations (7)
Affiliations (7)
- Department of Ultrasound, the First Affiliated Hospital, Fujian Medical University, Fuzhou, 350005, China.
- Department of Ultrasound, National Regional Medical Center, Binhai Campus of the First Affiliated Hospital, Fujian Medical University, Fuzhou, 350212, China.
- Department of Ultrasound, Fuzhou University Affiliated Provincial Hospital, Fuzhou, China.
- Department of Ultrasound, Fuqing City Hospital, Fujian Medical University, Fuzhou, China.
- Department of Ultrasound, the First Affiliated Hospital, Fujian Medical University, Fuzhou, 350005, China. [email protected].
- Department of Ultrasound, National Regional Medical Center, Binhai Campus of the First Affiliated Hospital, Fujian Medical University, Fuzhou, 350212, China. [email protected].
- Department of Ultrasound, Fuzhou University Affiliated Provincial Hospital, Fuzhou, China. [email protected].
Abstract
Bone erosion (BE) is a critical prognostic indicator in rheumatoid arthritis (RA). This study aimed to develop and validate an interpretable machine learning (ML) model for predicting BE risk in RA patients using the Shapley Additive exPlanations (SHAP) framework. This multi-center retrospective study enrolled 412 RA patients without baseline BE. Patients were stratified into BE and non-bone erosion (NBE) groups confirmed by MRI or musculoskeletal ultrasound after a 2-year follow-up. After data preprocessing, including K-nearest neighbors imputation and synthetic minority over-sampling technique-based resampling to address class imbalance, least absolute shrinkage and selection operator regression was applied for feature selection. Ten ML algorithms were trained using fivefold cross-validation with GridSearchCV hyperparameter optimization. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration (Brier score), and decision curve analysis (DCA). For the optimal model, the SHAP framework was utilized to interpret both global and individual feature contributions. A total of 318 patients from Center 1 were randomly divided into a training set (n = 222) and an internal-test set (n = 96) at a 7:3 ratio. The 94 patients from Center 2 constituted an external-test set. The eXtreme Gradient Boosting (XGBoost) model demonstrated superior performance, achieving an AUC of 0.896 on the internal-test set and 0.893 on the external-test set. SHAP global analysis revealed that synovial Power Doppler Imaging (PDI) (mean SHAP value = 2.761), synovial hyperplasia (2.168), disease duration (1.359), and anti-cyclic citrullinated peptide (1.260) were the four most important predictors. DCA showed the XGBoost model provided net benefit across 10%-80% threshold probabilities. The SHAP-interpretable XGBoost model demonstrates strong discriminative performance in predicting BE risk in RA, highlighting synovial PDI and synovial hyperplasia as pivotal factors.