Development and Validation of a Multivariable Prediction Model for Recurrent Osteoporotic Fractures in Elderly
Wenwen Shi1,2, Ratana Sapbamrer2, Dali Liang1
1Department of Nursing, Youjiang Medical University for Nationalities, 98 Chengxiang Road, Baise, Guangxi, 533000, China, ymcn.gx.cn.
Background/Objectives:
The objective of this study is to investigate the determinants of subsequent osteoporotic fractures (OPF) in elderly patients with Type 2 diabetes mellitus (T2DM) who have sustained an initial fracture and to construct and validate a multivariable predictive model for estimating individual recurrence risk.
Methods:
A total of 373 patients with OPF attributed to T2DM who received treatment at four public medical institutions in southwest China between June 2019 and July 2022 were included as the derivation cohort for model development. Demographic, lifestyle, and disease-related data were collected, and recurrent fracture outcomes were monitored. Model performance was evaluated using the optimism-corrected consistency index (C-index, 200 bootstrap resamples), time-dependent receiver operating characteristic (ROC) curves with area under the curve (AUC) (95% confidence interval [95% CI]; Heagerty method), calibration plots (deciles of predicted risk at 12, 24, and 36 months; slope and intercept), decision curve analysis (DCA), and a nomogram. Sensitivity analysis adjusting for hospital heterogeneity was performed. For external validation, an independent cohort of 140 patients from a different hospital (2023-2025) was used to assess model generalizability, based on the same key variables and follow-up outcomes.
Results:
The final multivariable model identified several independent predictors of recurrent fracture in patients with OPF and T2DM, including gender, treatment, alcohol consumption, postdischarge medication duration, fall risk, calcitonin, alkaline phosphatase, and homocysteine. Internal validation yielded an optimism-corrected C-index of 0.792, with time-dependent AUC at 12, 24, and 36 months of 0.815 (95% CI 0.770-0.859), 0.869 (95% CI 0.829-0.910), and 0.941 (95% CI 0.891-0.990), respectively. Calibration slopes (intercepts) at the same time points were 0.9 (0.03), 1.0 (0.01), and 1.19 (-0.12), and DCA showed net benefit superior to the treat-all and treat-none strategies across clinically relevant thresholds. In external validation, the model achieved a C-index of 0.692 and time-dependent AUC of 0.701 (95% CI 0.603-0.799), 0.654 (95% CI 0.536-0.772), and 0.803 (95% CI 0.467-1.000) at 12, 24, and 36 months, respectively; calibration slopes (intercepts) were 0.5 (0.06), 0.75 (0.07), and 0.44 (0.56). DCA remained favorable across clinically relevant threshold probabilities. Sensitivity analysis adjusting for hospital heterogeneity confirmed model robustness.
Conclusions:
The final model incorporated key modifiable predictors, showing excellent internal performance and clinical utility. Although external validation revealed moderate discrimination with miscalibration, the model remained robust after adjusting for hospital heterogeneity. Recalibration is recommended for absolute risk prediction in different populations; nonetheless, it offers a clinically applicable and generalizable solution for risk stratification in elderly patients with T2DM and OPF.
