A machine learning-based interpretable model for predicting pancreatic cancer in chronic pancreatitis patients with focal pancreatic lesions

BACKGROUND: Pancreatic cancer (PC) is deadly and distinguishing it from inflammatory conditions in chronic pancreatitis (CP) patients is challenging. We aimed to develop machine learning models to predict PC in CP patients with focal pancreatic lesions. METHODS: For this bicentric retrospective study, CP patients with indeterminate focal pancreatic lesions discovered through contrast-enhanced computed tomography scans were enrolled. Final diagnosis of focal pancreatic lesions was established by surgical pathology or follow-up outcomes. We used Boruta algorithm for feature screening, and conducted six machine learning models (Logistic Regression, Random Forest, eXtreme Gradient Boosting, Light Gradient Boosting Machine, K-Nearest Neighbors and Naive Bayes). The input data this study used were clinical information and laboratory data. Finally, SHAP was employed for interpretation. Receiver operating characteristic curve, area under curve (AUC), accuracy, sensitivity and specificity were used to evaluate model performance. RESULTS: A total of 187 participants were enrolled, and 44 patients (23.5%) were diagnosed as PC. Six important features were identified by the Boruta algorithm. Among the six machine learning models based on these features, Logistic Regression had the best diagnostic performance (AUC: 0.875 (95% CI: 0.801-0.949) in the training set and 0.908 (95% CI: 0.818-0.997) in the testing set). The ranking of SHAP variables importance from highest to lowest were carbohydrate antigen 19 - 9, hemoglobin, alanine transaminase, carcinoembryonic antigen, aspartate transaminase and the maximum diameter of focal pancreatic lesions. CONCLUSION: Six features-based machine learning models, especially the Logistic Regression, had satisfactory performance in predicting PC in CP patients with focal pancreatic lesions. This approach was crucial for enhancing early detection rate and reducing mortality associated with PC.