Interpretable machine learning prediction of 1-year overall survival in pancreatic cancer patients aged 65 years and older.

Background: Pancreatic cancer (PC) has an extremely poor prognosis, and elderly patients aged ≥65 years account for the majority of PC cases with more complex clinical characteristics. Accurate prediction of 1-year survival rate is crucial for individualized treatment decision-making in this subgroup. This study aimed to construct and validate machine learning (ML) models for predicting 1-year survival in PC patients aged ≥65 years, and to identify key prognostic factors using least absolute shrinkage and selection operator (LASSO) regression and SHapley Additive exPlanations (SHAP) analysis. Methods: Clinical data of PC patients aged ≥65 years diagnosed between 2011 and 2015 were extracted from the Surveillance, Epidemiology, and End Results (SEER) database. Demographic, tumor-related, and treatment-related indicators were included as predictive variables. Six ML algorithms [decision tree (DT), random forest (RF), support vector machine (SVM), light gradient boosting machine (LightGBM), logistic regression (LR), and K-nearest neighbor (KNN)] were used to construct prediction models. LASSO regression was employed for feature selection and optimal model screening. Model performance was evaluated using area under the curve (AUC), accuracy, sensitivity, and F1-score. SHAP analysis was performed to interpret the optimal model and clarify the importance of key features. Results: A total of 5,601 eligible patients were included in this study. LASSO regression identified 6 key prognostic features: surgery, grade, T stage, chemotherapy, radiation, and N stage. The RF model showed the best predictive performance, with an AUC of 0.875 in the training set and 0.799 in the validation set. The confusion matrix of the RF model on the validation set demonstrated good discrimination between survived and non-survived patients. SHAP analysis revealed that surgery was the most influential factor promoting 1-year survival, while higher grade was the dominant factor reducing survival probability. T stage and chemotherapy also had significant impacts on the predictive outcome. Conclusions: The RF model constructed based on SEER database data can accurately predict the 1-year survival rate of PC patients aged ≥65 years. Surgery, grade, T stage, and chemotherapy are the key prognostic factors. This interpretable ML model provides a reliable tool for clinicians to assess short-term prognosis and formulate individualized treatment strategies for elderly PC patients.