Explainable AI assisted vertebral refracture diagnosis after percutaneous vertebroplasty through effective feature engineering and stacked ensemble learning

BACKGROUND AND OBJECTIVE: To develop and validate a machine learning model based on stacking ensemble learning and feature selection strategies to predict vertebral refracture risk after percutaneous vertebroplasty. METHODS: Clinical data from osteoporotic vertebral compression fractures (OVCF) patients treated with percutaneous vertebroplasty were retrospectively collected. Preprocessing steps included outlier removal, feature encoding, normalization, and Synthetic Minority Oversampling Technique (SMOTE) for dataset balancing. Univariate and multivariate analyses identified 13 key predictive features. Five machine learning algorithms-Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and XGBoost-were employed within a two-layer stacking ensemble framework for dynamic weighted fusion. Model performance was evaluated using 10-fold cross-validation and an independent validation set, with explainability enhanced through SHapley Additive exPlanations (SHAP) for global and individual-level explanations. RESULTS: The Stacking Ensemble Model demonstrated exceptional performance in internal cross-validation, achieving an area under the curve (AUC) of 0.984 and Average Precision (AP) of 0.987, with superior recall and Brier scores. In the independent validation cohort, it maintained robust performance (AUC = 0.959, AP = 0.949). SHAP analysis identified brace usage, age, anti-osteoporosis drug, daily sunlight exposure, and 25-OH-VD levels as primary risk determinants. CONCLUSION: The Stacking Ensemble Model demonstrates high predictive accuracy and robust generalization performance in assessing vertebral refracture risk following percutaneous vertebroplasty, providing clinicians with effective decision support for risk assessment and personalized rehabilitation planning.