An explainable prognostic model after vascularized bone grafting for hip preservation based on CT radiomics combined with SHAP.

The purpose of this study is to develop a CT radiomics-based interpretable prognostic diagnostic model for vascularized bone graft hip preservation, with the objective of predicting postoperative hip preservation outcomes. The study recruited 107 patients, collecting preoperative CT scans and preoperative blood biochemistry data. Among these patients, 27 had a good prognosis, while 80 had a poor prognosis. Five machine learning algorithms were employed to develop predictive models evaluating the effectiveness of modified vascularized bone implants in hip preservation. The interpretability of the top-performing models was assessed using SHapley Additive exPlanations (SHAP). Nine radiomic features were extracted from preoperative CT scans to develop a radiomic score. Through univariate and multivariate logistic regression analyses, clinical indicators, including patient age and preoperative platelet-to-lymphocyte ratio (PLR), were retained. Fifteen models were constructed, incorporating clinical, radiomic, and combined approaches across various algorithms. The combined model utilizing the XGBoost algorithm demonstrated superior performance, achieving an AUC of 0.90 (95% CI 0.81-0.98) on the training set and 0.87 (95% CI 0.75-1.00) on the test set. These results showed improvements of around 31% and 28%, respectively, compared to the top performing clinical and radiomic models (p < 0.05). High radiomics scores, a high PLR, and older age were identified as significant predictors of poor prognosis. A robust joint clinical and radiomics model was developed using the XGBoost algorithm for predicting the prognosis of hip-preserving surgery. The predictions of this model were interpreted using SHAP to enhance clinical applications.