Preoperative prediction of intrahepatic cholangiocarcinoma lymph node metastasis by means of machine learning: a multicenter study in China

Background: Hepatectomy is currently the most effective modality for the treatment of intrahepatic cholangiocarcinoma (ICC). The status of lymph node affects the choice of surgical method and the formulation of postoperative treatment plan. Therefore, preoperative judgment of lymph node status is of great significance. Previous prediction models mostly adopted logistic regression modeling, and few studies applied random forests in the prediction of ICC lymph node metastasis(LNM). Methods: : A total of 149 ICC patients who met clinical conditions were enrolled in the training group. Taking into account preoperative clinical data and imaging features, 21 indicators were included for analysis and modeling. Logistic regression was used to filter variables through multivariate analysis, and random forest was used to rank the importance of variables through algorithm. The prediction accuracy of the model was assessed by the C-index and calibration curve and validated with external data. Result: Multivariate analysis shows that CEA, CA199, and lymphadenopathy on imaging are independent risk factors for lymph node metastasis. The random forest algorithm identifies the top four risk factors as CEA, CA199, lymphadenopathy on imaging and AST. The predictive power of random forest is significantly better than the nomogram established by logistic regression in both the validation group and the training group (AUC reached 0.758 in the validation group) Conclusions: : We construct a random forest model for predicting lymph node metastasis. Compared with the traditional nomogram, it has higher prediction accuracy, and at the same time, it has an auxiliary role in imaging examinations.