Machine learning incorporating trabecular bone score improves lumbar fracture risk stratification in type 2 diabetes

Conventional fracture risk assessment tools frequently underestimate lumbar fragility fracture risk in type 2 diabetes mellitus (T2DM). This study developed and validated a machine learning (ML) model integrating trabecular bone score (TBS) to enhance discrimination accuracy for lumbar fragility fractures in T2DM patients. We retrospectively analyzed 261 T2DM patients (≥ 50 years, 77.4% female) who underwent dual-energy X-ray absorptiometry (DXA) at Jinjiang Hospital (August 2023–May 2024). Six features (lumbar spine bone mineral density (LS-BMD), TBS, age, body mass index (BMI), gender, and fasting blood glucose) were selected via Least Absolute Shrinkage and Selection Operator (LASSO) regression to train ten machine learning algorithms. Model performance was evaluated using 5-fold cross-validation and an independent test set (20% holdout). Model interpretability was analyzed using the SHapley Additive exPlanations (SHAP) framework. Among 261 patients, 64 (24.5%) had lumbar fragility fractures. The multivariate integrated model (AUC = 0.797) significantly outperformed LS-BMD alone (AUC = 0.739, P < 0.001) and TBS alone (AUC = 0.724, P < 0.001). Among the ten algorithms, the logistic regression (LR) model demonstrated optimal performance with a cross-validation AUC of 0.889 and an independent test set AUC of 0.818. SHAP analysis identified TBS as the most important discriminative feature (mean |SHAP| value = 0.1245). TBS serves as a critical microstructural predictor, addressing the limitations of conventional BMD in assessing lumbar fragility fracture risk in T2DM patients. Machine learning models incorporating TBS, particularly the logistic regression model, significantly enhance discrimination accuracy for lumbar fragility fractures in T2DM patients.