An explainable machine learning model for predicting osteoporotic fragility fractures: a retrospective study in South China

The study examined the clinical characteristics of hospitalized patients with osteoporotic fractures in South China. The research included 1125 hospitalized patients with osteoporotic fractures who underwent DXA examination between January 2019 and December 2025. Seventy-three features were collected from patients, including clinical, laboratory, and radiological data. Researchers tested six machine learning algorithms to create the best predictive model. The LightGBM model achieved the best results with ROC-AUC of 0.903 on the training set and 0.840 on the test set. SHAP analysis identified age, hepatitis C virus antibodies, and serum sodium levels as the main factors increasing fracture risk. The model demonstrates good ability to predict osteoporotic fragility fractures and opens the way for earlier intervention in patients susceptible to these fractures.