Abstract
Osteoporotic fractures are a major complication of osteoporosis and pose a substantial global health burden, particularly in postmenopausal women. Although bone mineral density (BMD) is widely used for fracture risk assessment, its predictive accuracy is limited, and integrating multidimensional clinical indicators may improve risk prediction. This retrospective study included 1,717 postmenopausal women from two tertiary hospitals in Shaanxi Province, China, who were classified into fracture (n=797) and non-fracture (n=920) groups based on a history of low-energy fractures. Thirty-two clinical variables, including BMD, bone turnover markers (BTMs), serum electrolytes, age, and body mass index, were analyzed. Recursive feature elimination was applied, and ten machine learning models were developed using a training dataset (70%) and evaluated on a testing dataset (30%). Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Model interpretability was explored using SHapley Additive exPlanations (SHAP). Among all models, the Random Forest model demonstrated the best performance (AUC = 0.872), outperforming the Extra Trees (AUC = 0.841) and XGBoost (AUC = 0.836) models. SHAP analysis identified BMD, serum chloride (Cl(-)), age, albumin-to-globulin ratio, and neutrophil percentage as the most influential predictors, with osteocalcin N-mid fragment contributing more prominently than other BTMs. In conclusion, this machine learning-based model effectively identified key risk factors for osteoporotic fractures in postmenopausal women, and integrating BMD with biochemical and clinical indicators may improve fracture risk prediction and support clinical screening and risk stratification.