Machine learning prediction of perineural invasion in intrahepatic cholangiocarcinoma

Perineural invasion (PNI) significantly influences postoperative recurrence and survival in intrahepatic cholangiocarcinoma (ICC) patients. This study aims to develop and validate an interpretable model that can be used to predict PNI in ICC cases before surgery.

Retrospective clinical information was gathered from ICC patients (n = 250) at our hospital, covering the period from January 2012 to January 2022. The patients were randomly assigned to the training group (n = 176, 70.4 %) and validation group (n = 74, 29.6 %). We employed four machine learning algorithms to establish prediction models, each model's performance was assessed via a receiver operating characteristic (ROC) curve. Decision Curve Analysis (DCA) was performed to evaluate the models' risks and benefits. SHapley Additive exPlanations (SHAP) were used to quantify the contributions of model features, providing both global and local interpretations.

Significant differences in tumor size, tumor number, lymph node metastasis, CA199, distant metastasis ratio, HBsAg, PLR, and NLR were observed between the PNI[-] (n = 172, 68.8 %) and PNI[+](n = 78, 31.2 %) groups. The PFS and OS rates in the PNI[-] group were better than those in the PNI[+] group. Based on the evaluation of the validation group, the XGBoost model demonstrated the best predictive performance. SHAP analysis identified tumor number, tumor size, and lymph node metastasis as the top three factors predicting PNI in ICC patients.

We developed a reliable predictive model that effectively predicts PNI status in patients with ICC and facilitates personalized clinical decision-making.