Application of machine learning for prognostic modeling in unresectable pancreatic cancer treated with chemoradiotherapy.

Background: Patients with unresectable pancreatic cancer have poor outcomes despite chemoradiotherapy (CRT). Traditional prognostic tools lack accuracy in predicting survival. This study aimed to develop an artificial intelligence (AI)-based model to improve survival prediction. Methods: We retrospectively included 214 patients treated with CRT between 2018 and 2024. Five models-Cox, LASSO, RSF, SVM, and XGBoost-were trained to predict overall survival. Model performance was evaluated using the C-index, time-dependent ROC, calibration, and decision curve analysis. SHAP was used to interpret feature importance. Results: The median overall survival (mOS) for the entire cohort was 18.4 months (95% CI, 16.3-28.1). XGBoost showed the best performance (C-index = 0.949). It also achieved higher area under the receiver operating characteristic curves at 6 and 12 months (0.751 and 0.732) compared to other models. Calibration and clinical benefit were superior. SHAP analysis identified CA199, tumor size, platelet count, and age as the most important predictors. The model stratified patients into risk groups with significant survival differences (p < 0.001). Conclusion: The XGBoost-based model accurately predicted survival in unresectable pancreatic cancer patients receiving CRT. It may serve as a useful tool for personalized risk assessment and treatment planning.