An interpretable machine learning model integrating ultrasound and clinical variables for predicting osteoporosis in patients with rheumatoid arthritis.
Authors
Affiliations (9)
Affiliations (9)
- Department of Ultrasound, The First Affiliated Hospital of Fujian Medical University, No. 20 Chazhong Road, Taijiang District, 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, No. 134 East Street, Gulou District, Fuzhou, 350001, China.
- Department of Neurosurgery, The First Affiliated Hospital of Fujian Medical University, No. 20 Chazhong Road, Taijiang District, Fuzhou, 350005, China.
- Department of Ultrasound, The First Affiliated Hospital of Fujian Medical University, No. 20 Chazhong Road, Taijiang District, 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, The First Affiliated Hospital of Fujian Medical University, No. 20 Chazhong Road, Taijiang District, 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, No. 134 East Street, Gulou District, Fuzhou, 350001, China. [email protected].
Abstract
Rheumatoid arthritis (RA) significantly increases the risk of osteoporosis (OP) and fractures, yet dual-energy X-ray absorptiometry (DXA) is underused in routine care. This study aims to develop and explain a machine learning model to identify individuals with OP risk among RA patients. In this single-center retrospective study, 299 RA patients were enrolled and split 7:3 into training and test sets. Candidate features included clinical variables, laboratory indices, and musculoskeletal ultrasound semi-quantitative scores. After performing correlation analysis and variance inflation factor (VIF) screening, recursive feature elimination combined with random forest (RF) and five-fold cross-validation was used for feature selection. Nine machine learning models were constructed and compared in terms of performance. The optimal model was further validated through calibration curves, precision-recall (P-R) curves, and SHapley Additive exPlanations (SHAP) analysis. The extra trees (ET) model demonstrated the most stable performance, with an area under the curve (AUC) of 0.914 (95% CI 0.854-0.961), specificity of 0.951, sensitivity of 0.586, accuracy of 0.833, and a Brier score of 0.120 in the test set. Calibration and P-R curve analyses indicated good model performance. SHAP analysis revealed bone erosion (BE), age, and disease duration as key driving factors for RA-OP. An interpretable machine learning model integrating clinical variables, laboratory indices, and semi-quantitative ultrasound scores may support risk stratification for DXA-defined osteoporosis in patients with RA. Given its high specificity but moderate sensitivity, the model should be regarded as an adjunctive tool for prioritizing DXA assessment and bone-health management rather than as a stand-alone screening method. Key Points • An explainable machine learning model integrating clinical variables and semi-quantitative ultrasound scores was developed for RA-specific OP risk stratification. • The final extra trees model demonstrated good discrimination and high specificity, but moderate sensitivity, supporting its use as an adjunctive rather than stand-alone screening tool. • Semi-quantitative ultrasound scores may provide complementary information beyond routine clinical indicators for prioritizing DXA assessment.