Background & AimsEarly recurrence is the leading cause of death for patients with perihilar cholangiocarcinoma (pCCA) after surgery. Identifying high-risk patients preoperatively is important. This study aimed to construct a preoperative prediction model for the early recurrence of pCCA patients planned treatment with curative resection. MethodsThis study ultimately enrolled 400 pCCA patients after curative resection in five hospitals between 2013 and 2019. They were randomly divided into training (n=300) and testing groups (n=100) at a ratio of 3:1. Associated variables were identified via LASSO regression. Four machine learning models were constructed: support vector machine (SVM), random forest (RF), logistic regression, and K-nearest neighbors (KNN). The predictive ability of the models was evaluated via receiver operating characteristic (ROC) curves, precision–recall curve (PRC) curves, and decision curve analysis (DCA). KaplanMeier survival curves were drawn for the high/low-risk population. ResultsFive factors, CA19-9, tumor size, total bilirubin, hepatic artery invasion, and portal vein invasion, were selected by LASSO regression. In both the training and testing groups, the ROC curve (AUC: 0.983 vs 0.952) and the PRC (0.981 vs 0.939) showed that RF was the best. The cutoff value for distinguishing high- and low-risk patients was 0.51. KM survival curves revealed that in both groups, there was a significant difference in RFS between high- and low-risk patients (P<0.001). ConclusionThis study used preoperative variables from a large, multicenter database to construct a machine learning model that could effectively predict the early recurrence of pCCA patients planned treatment with curative resection and help clinicians make better treatment decisions.