Prediction of Vertebral Recollapse After Percutaneous Pedicle Screw Fixation for Single- Level Traumatic Thoracolumbar Fractures: Model Development and Internal Validation

Abstract Background Progressive loss of restored vertebral height can occur after percutaneous pedicle screw fixation (PPSF) for traumatic thoracolumbar fractures, yet individualized prediction of vertebral recollapse remains limited. We aimed to develop and internally validate an early postoperative prediction model for vertebral recollapse after short-segment PPSF. Methods We retrospectively included 420 neurologically intact patients with single-level T11–L2 traumatic fractures treated with a uniform four-screw, two-rod PPSF construct. Recollapse was defined as a > 20-percentage-point increase in normalized anterior vertebral height loss (ΔHL) from the early postoperative assessment to follow-up before planned implant removal. Twelve candidate predictors were evaluated using least absolute shrinkage and selection operator (LASSO) regression followed by multivariable logistic regression. Internal validation used 1,000 bootstrap resamples. Post hoc comparisons of four nested models used a common stratified 10-fold cross-validation partition. Sensitivity analyses examined alternative recollapse thresholds and continuous ΔHL using multivariable linear regression with HC3 robust standard errors. Results Recollapse occurred in 98 patients (23.3%). Six predictors were retained: age, vertebral compression degree, injured-vertebral Hounsfield unit (HU) value, screw-to-endplate distance, facet joint violation (FJV) grade, and lower endplate destruction. The apparent and optimism-corrected areas under the receiver operating characteristic curve (AUCs) were 0.851 and 0.829, respectively; the optimism-corrected Brier score was 0.135 and the calibration slope was 0.883. In nested comparisons, cross-validated AUC increased from 0.734 for age and compression to 0.834 for the full model, while Brier score decreased from 0.158 to 0.132. The largest incremental gain occurred after addition of the postoperative imaging block, whereas HU added little discrimination. Continuous-outcome analysis showed concordant association directions, with residual uncertainty for HU. Conclusions The six-predictor model showed promising internally validated performance for early postoperative prediction of vertebral recollapse. Post hoc comparisons suggested additional predictive information from endplate integrity and the combined postoperative imaging variables, whereas the observed incremental contribution of HU was small. External validation and prospective clinical-impact assessment are required before clinical implementation.