A novel machine learning-based model for precise prediction of postoperative recurrence in gallbladder cancer: development and multicenter validation across four centers

Gallbladder cancer (GBC) is a highly aggressive malignancy with an extremely poor prognosis, and its high postoperative recurrence rate is a major challenge affecting long-term patient survival. Accurately predicting individualized recurrence risk is crucial for formulating adjuvant treatment and follow-up strategies. This study aimed to develop and validate a high-accuracy predictive model by integrating multi-center clinicopathological data and systematically comparing multiple machine learning algorithms. The SHAP (SHapley Additive exPlanations) framework was employed to enhance model interpretability, quantifying the contribution of key predictive factors, thereby providing an accurate and transparent tool for individualized postoperative recurrence risk assessment in GBC. This study retrospectively included GBC patients who underwent radical resection from four centers. Clinicopathological characteristics (e.g., TNM stage, tumor differentiation, vascular invasion, perineural invasion), serum tumor markers, and preoperative imaging data were collected. Data from one entire center served as the training set, while data from the remaining three centers were combined as the validation set, achieving a case number ratio of approximately 7:3. Four machine learning algorithms—Extreme Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Machine (SVM), and Cox Proportional Hazards Regression model—were employed to construct prediction models. Model performance was evaluated by the Area Under the Receiver Operating Characteristic Curve (AUC) for discrimination, calibration curves for accuracy, and Decision Curve Analysis (DCA) for clinical utility. SHAP analysis was applied to the best-performing model for interpretability, quantifying the contribution of each feature to the predictions. On the training set, the XGBoost model demonstrated the best predictive performance, with an AUC of 0.969, significantly superior to RF (AUC = 0.941), SVM (AUC = 0.711), and the Cox model (AUC = 0.676). The calibration curve indicated high consistency between predicted probabilities and actual recurrence probabilities for the XGBoost model. DCA further confirmed its highest clinical net benefit across a wide range of threshold probabilities. SHAP analysis revealed that the T-stage was the most important risk factor affecting model predictions, followed by tumor differentiation grade, N-stage, CEA, and indirect bilirubin levels. The analysis also provided individualized risk explanations, demonstrating complex interactions among features. This study successfully constructed a postoperative recurrence prediction model for GBC based on the XGBoost algorithm, which demonstrated excellent performance in discrimination, calibration, and clinical utility. The model’s interpretability analysis not only validated known clinical risk factors but also provided a powerful tool for individualized risk assessment. This model holds promise for assisting clinicians in early postoperative identification of high-risk patients, thereby offering decision support for formulating individualized adjuvant treatment strategies (e.g., more active recommendation of adjuvant chemotherapy for high-risk patients) and differentiated follow-up plans.