Machine learning prediction of early reoperation following lower extremity tumor resection and endoprosthetic reconstruction: A PARITY trial secondary analysis

Oncologic resection and endoprosthetic reconstruction of malignant bone tumors carries a high risk of complication and secondary surgery. Given the significant morbidity associated with reoperation in systemically compromised patients, accurate risk stratification is critical to patient counseling and shared decision-making. The purpose of this study was to develop a machine learning (ML) model for prediction of reoperation within one year of lower extremity tumor resection and endoprosthetic reconstruction. Using data from the PARITY trial, 54 features across 604 lower extremity endoprosthetic reconstructions were evaluated as predictors of all-cause reoperation within one year. Logistic regression (LR), Random Forest, gradient boosting, AdaBoost, and XGBoost were used for model building. Standard metrics of area under receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and Brier scores were used to evaluate model performance. Important features for the top-performing model were determined. Of 604 lower extremity endoprosthetic reconstructions performed in the study period, 155 patients (25.7%) underwent at least one reoperation. The Gradient Boosting model had the highest discrimination (AUROC = 0.817, AUPRC = 0.690) of the tested models and was well-calibrated. Surgical site infection (SSI), operative time, white race, negative pressure wound therapy (NPWT) use, and female sex were the five most important features for model performance. We report a well-calibrated ML-driven algorithm with high discriminatory power for the prediction of all-cause early reoperation following lower extremity tumor resection and endoprosthetic reconstruction. Preventing SSI remains paramount to avoiding the morbidity of reoperation after complex oncologic limb salvage surgeries.