Abstract: Objective To develop a discrimination model for differentiating between pancreatic cancer and benign pancreatic lesions using machine learning methods. Methods The study population consisted of 251 patients diagnosed with pancreatic diseases and treated at the Second Affiliated Hospital of Nanchang University between January 2018 and December 2023. Six machine learning models were developed, including logistic regression, random forest, eXtreme gradient boosting (XGBoost), support vector machine, multilayer perceptron, and Gaussian Naive Bayes, to distinguish pancreatic cancer from benign pancreatic lesions. The models' discriminatory capabilities were assessed using the receiver operating characteristic (ROC) curve. Model consistency was evaluated using a calibration curve, clinical applicability was assessed through a decision curve, and model interpretation was facilitated by the  SHapley Additive exPlanations (SHAP) method. Results Out of the  251 patients with pancreatic diseases, 100 were diagnosed with pancreatic cancer, while 151 were diagnosed with benign pancreatic lesions. Six machine learning models were successfully developed, with the area under the ROC curve (AUC) for the random forest, XGBoost, support vector machine, and multilayer perceptron models demonstrated superior performance compared to the carbohydrate antigen 19⁃9 (CA19⁃9) marker (all P<0.05). Notably, the XGBoost model exhibited the highest AUC （AUC=0.886）, and analyses using decision and calibration curves further confirmed its substantial clinical net benefit and consistency. SHAP analysis identified CA19⁃9 as the most significant contributor to the XGBoost model. Conclusions　XGBoost model developed using tumor markers and clinical test data, significantly enhances the ability to discriminate  between pancreatic cancer from benign pancreatic lesions, indicating potential for clinical application.

Key words: Pancreatic cancer, Benign pancreatic lesions, Machine learning, Discrimination model