External validation of predictive models for new vertebral fractures following percutaneous vertebroplasty

Abstract

Abstract Objective To investigate the external validation and scalability of four predictive models regarding new vertebral fractures following percutaneous vertebroplasty. Methods Utilizing retrospective data acquired from two centers, compute the area under the curve (AUC), calibration curve, and Kaplan–Meier plot to assess the model’s discrimination and calibration. Results In the external validation of Zhong et al.’s 2015 predictive model for the probability of new fractures post-vertebroplasty, the AUC for re-fracture at 1, 2, and 3 years postoperatively was 0.570, 0.617, and 0.664, respectively. The AUC for Zhong et al.’s 2016 predictive model for the probability of new fractures in neighboring vertebrae was 0.738. Kaplan–Meier plot results for both models indicated a significantly lower incidence of re-fracture in low-risk patients compared to high-risk patients. Li et al.’s 2021 model had an AUC of 0.518, and its calibration curve suggested an overestimation of the probability of new fractures. Li et al.’s 2022 model had an AUC of 0.556, and its calibration curve suggested an underestimation of the probability of new fractures. Conclusion The external validation of four models demonstrated that the predictive model proposed by Zhong et al. in 2016 exhibited superior external generalization capabilities.