Multicenter machine learning model to predict very early recurrence after curative-intent resection of hilar
Background:
Surgical resection remains the main curative option for hilar cholangiocarcinoma (HCCA), but long-term survival is often limited by very early recurrence (VER, ≤6 months postoperatively). We aimed to develop a model to predict VER after curative-intent resection.
Methods:
In this retrospective multicenter cohort study, we included patients who underwent curative-intent resection for HCCA at three Chinese centers between April 1, 2010, and March 1, 2023. Patients were randomly assigned to training and validation cohorts. Logistic regression, random forest, support vector machine, and extreme gradient boosting (XGBoost) models were constructed to predict VER. Performance was evaluated using receiver operating characteristic, Brier score, calibration plots, and decision curve analyses. SHapley Additive exPlanations were used to assess feature importance.
Results:
Among 474 patients, 106 (22.4%) developed VER. The training and validation cohorts comprised 331 and 143 patients, respectively. VER was associated with hepatic artery invasion, portal vein invasion, positive margins, lymph node metastasis, elevated CA19-9, and poor differentiation. XGBoost achieved the highest AUC (0.85, 95% CI 0.78-0.92) and was selected for online deployment.
Conclusion:
The XGBoost model predicts VER after HCCA resection and may facilitate postoperative risk stratification and management.
