Comprehensive Summary
In this study, researchers aimed to develop machine learning models to accurately predict the risk of refracture after percutaneous kyphoplasty (PKP) in patients with osteoporotic vertebral compression fractures (OVCF). 3,942 OCVF patients who underwent PKP were identified and classified into non-refracture (3,518) and refracture (424) groups. The primary diagnostics for refracture were new onset back pain, limited mobility, or findings on postoperative spinal MRIs. Machine learning models were trained using the two groups, and class weights were applied to address imbalance. 9 different models (Logistic Regression, Decision Tree, Random Forest, SVM, XGBoost, CatBoost, Gradient Boosting, Balanced Bagging, and MLPClassifier) were tested, and performance was evaluated using an independent validation set. The core metrics tested were Area Under the Curve (AUC), specificity, and sensitivity. Identified risk factors for refracture included sex, age, BMI, bone mineral density (BMD), fracture level, cement distribution and leakage, sarcopenia ,anterior vertebral height restoration rate (AVHRR), and osteoporosis treatment. Overall, the Random Forest model had a strong performance (AUC = 0.993, specificity = 0.994) and identified the key risk factors, which were AVHRR, BMD, and osteoporosis treatment. Collectively, the machine learning models accurately predicted refracture risk and identified key clinical contributors.
Outcomes and Implications
The use of machine learning models in this study provides an efficient way to predict refracture risk following PKP. Earlier detection of patients at risk for refracture allows clinicians to intervene sooner, potentially reducing complications and preventing further spinal damage. By identifying key risk factors, these models enable more specific preoperative planning, surgical technique adjustments, and postoperative care. A more personalized approach can improve patient outcomes and increase efficiency in clinical settings. Further research should validate these models in wider clinical settings and larger patient populations to ensure generalizability.