Development and validation of a machine learning model for predicting refracture risk in older adults with osteoporotic vertebral compression fractures

Osteoporotic vertebral compression fractures (OVCFs) are common among older adults and are associated with a high risk of subsequent vertebral refracture, which may result in functional decline and reduced quality of life. However, effective tools for identifying Older Adults at high risk of refracture remain limited. This study aimed to develop and validate a machine learning-based prediction model for estimating one-year vertebral refracture risk among Older Adults with OVCFs in China. Older Adults with OVCFs admitted to three hospitals affiliated with Zunyi Medical University were included in this study. Patients admitted between July 2021 and July 2023 formed the retrospective development cohort ( n = 582), which was randomly divided into a training cohort ( n = 406) and an internal validation cohort ( n = 176). Patients admitted between August 2023 and August 2024 formed the prospective external validation cohort ( n = 402). After excluding 13 patients lost to follow-up, 389 patients constituted the external validation cohort. Seven machine learning models were developed using hyperparameter optimization and five-fold cross-validation. Model performance was assessed in the internal validation cohort using AUC, sensitivity, specificity, and accuracy. The machine-learning model with the most consistent validation performance was used to select predictors, which were then entered into a multivariable logistic regression model for nomogram development. The nomogram was subsequently externally validated for discrimination, calibration, and clinical utility using AUC, calibration metrics, Brier score, and decision curve analysis. Among the 971 included patients, 185 (19.05%) developed vertebral refracture within one year. Among the seven machine-learning models evaluated, ENET achieved the highest AUC in the internal validation cohort (0.709, 95% CI: 0.621–0.789) and the smallest difference in AUC between the training and internal validation cohorts. ENET was used for variable selection, retaining three predictors with non-zero standardized coefficients: lumbar bone mineral density (BMD) (− 0.157), percutaneous vertebroplasty (PVP) (0.111) and previous fracture history (0.103). These predictors were then entered into a multivariable logistic regression model, which constituted the final prediction model. Higher lumbar BMD was associated with a lower risk of refracture (OR = 0.553, 95% CI: 0.435–0.704, P < 0.001), whereas previous fracture history (OR = 2.276, 95% CI: 1.458–3.554, P < 0.001) and PVP (OR = 2.602, 95% CI: 1.672–4.050, P < 0.001) were associated with an increased risk. The final model was presented as a nomogram and externally validated, with an AUC of 0.698 (95% CI: 0.641–0.755), a Brier score of 0.1404, a calibration slope of 0.9864, and a calibration intercept of 0.0098. The ENET-assisted nomogram incorporating lumbar BMD, previous fracture history, and PVP provides an individualized approach for estimating one-year vertebral refracture risk in Older Adults with OVCFs. Although the model showed moderate discrimination, it may serve as a practical tool for early risk stratification when combined with comprehensive clinical assessment.