Structural Injury Severity Predicts Poor Prognosis After Posterior Pedicle Screw Fixation for Thoracolumbar Burst Fractures: A Retrospective Modeling Study

Objective: To identify dominant predictors and compare the performance of multiple machine learning models in predicting delayed union and poor functional outcomes following posterior pedicle screw fixation for thoracolumbar burst fractures. Methods: This retrospective single-center study initially screened 552 patients, with 379 eligible patients included in the final analysis. Seventeen clinical and radiographic variables were analyzed. The primary endpoint (poor outcome) was defined as: (1) Oswestry Disability Index improvement rate <50% with CT-confirmed nonunion at 12 months, or (2) Vertebral height loss ≥20% after implant removal. After identifying key predictors using multivariable logistic regression, we performed a head-to-head comparison of four machine learning models: Random Forest, XGBoost, LightGBM, and Support Vector Machine (SVM). Results: Poor outcomes occurred in 51 patients (13.5%). Multivariable logistic regression identified four dominant predictors: endplate injury grade (OR = 6.61, p = 0.0002), disc injury grade (OR = 6.37, p = 0.0021), vertebral compression ratio (OR = 0.86, p = 0.0001), and vertebral HU (OR = 0.94, p = 0.0025). The logistic regression model exhibited exceptional discrimination (AUC = 0.961, 95% CI: 0.944–0.967). A nomogram integrating these predictors was developed to personalize risk assessment. In the machine learning comparison, XGBoost achieved the most balanced and superior overall performance (AUC = 0.934, Accuracy = 0.956), while all models confirmed the paramount importance of structural injury severity in prognostic prediction. Conclusion: Structural injury severity and bone quality are the strongest predictors of poor prognosis. Among the five algorithms compared, the XGBoost model provided the optimal balance between sensitivity, specificity, and calibration. The findings validate the robustness of using structural metrics for risk stratification and provide a reliable, multi-model benchmark for developing clinical decision-support tools.