Abstract / Summary
Abstract Background Adjacent vertebral fracture (AVF) is an important complication after percutaneous vertebroplasty (PVP) for osteoporotic vertebral compression fracture (OVCF). Existing prediction models frequently combine variables obtained at different clinical time points, which may limit their applicability for preoperative risk assessment. This study aimed to develop and internally validate a strictly preoperative prediction model and to determine whether perioperative cement-distribution information provides additional prognostic value after PVP. Methods We retrospectively included 314 consecutive patients with OVCF who underwent PVP between December 2022 and December 2024 and completed at least 18 months of follow-up. The primary outcome was AVF within 18 months after PVP. Model 1 incorporated six preoperative predictors: age, target vertebral CT Hounsfield units (CT-HU), pre-existing vertebral compression fracture (VCF) count, paraspinal fat infiltration rate (FIR), preoperative sagittal imbalance, and preoperative anterior vertebral compression rate. Model 2 retained these predictors and added cement distribution pattern. Internal validation included 1,000 bootstrap resamples, 100 repetitions of stratified five-fold cross-validation, and fixed stratified 10-fold out-of-fold (OOF) prediction. Discrimination, calibration, Brier score, decision curve analysis, and incremental model performance were evaluated. LASSO and DEXA T-score analyses were performed as sensitivity analyses. Results Among 314 patients, 80 (25.5%) developed AVF within 18 months. In Model 1, older age, lower CT-HU, greater pre-existing VCF burden, higher paraspinal FIR, sagittal imbalance, and greater anterior vertebral compression were independently associated with AVF. The apparent AUC was 0.872, the bootstrap-corrected AUC was 0.857, and the 10-fold OOF AUC was 0.854. After addition of cement distribution, contiguous O-shaped and fragmented patterns were associated with higher AVF odds than bilateral H-shaped distribution. Model 2 improved overall model fit (likelihood-ratio χ²=11.99, P = 0.0025), but the increase in OOF discrimination was small (AUC 0.859; ΔAUC = 0.005, 95% CI − 0.014 to 0.025; P = 0.622). Model 1 OOF risk strata showed observed AVF rates of 6.0%, 25.8%, and 63.2% in the low-, intermediate-, and high-risk groups, respectively. Adding DEXA T-score did not materially improve Model 1. Conclusions A six-predictor preoperative model integrating bone quality, previous fracture burden, paraspinal muscle degeneration, sagittal alignment, and vertebral morphology showed good internal performance for predicting 18-month AVF after PVP. Cement distribution provided additional perioperative prognostic information and improved overall model fit, but only minimally increased discrimination. These findings support a two-stage framework in which preoperative risk is estimated first and subsequently updated using procedural information. External validation is required before clinical implementation.