Development and validation of a machine learning model for predicting liver metastasis in pancreatic cancer: a retrospective cohort study.

Background: Liver metastasis (LM) is the leading cause of treatment failure and poor survival rates in pancreatic cancer (PC). Early identification of patients at high risk for LM is crucial for optimizing clinical management. This study aims to develop and validate a machine learning (ML)-based predictive model to accurately identify the risk of LM in PC patients and to determine the core risk factors contributing to LM. Methods: We collected data from eligible PC patients from the Surveillance, Epidemiology, and End Results (SEER) database (2010-2015), which was divided into a training set (70%) and a validation set (30%). Feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO), recursive feature elimination (RFE), and Boruta-Shap algorithms, with model parameters optimized through cross-validation. Six machine learning (ML) algorithms were developed to predict LM risk in PC patients. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, including the area under the curve (AUC), sensitivity, specificity, and negative predictive value (NPV). The best-performing model was further interpreted using SHapley Additive exPlanations (SHAP). Additionally, based on the optimal ML model, an online calculator was developed to provide personalized LM risk assessments for PC patients. Results: A total of 12,098 eligible PC patients were registered in the SEER database, with 8,469 assigned to the training set and 3,629 to the validation set. Using a combination of LASSO, RFE, and Boruta-Shap methods, key predictors of LM in PC were identified, including surgery, radiotherapy, tumor size, age, histological grade, primary site, pathological type, T stage, lung metastasis, and bone metastasis. All six ML models performed well in the training set, with the final model selection based on the validation set results. In the validation set, the Gradient Boosting Machine (GBM) model achieved a competitive AUC value of 0.875 [95% confidence interval (CI): 0.862-0.889], with balanced sensitivity (0.851), specificity (0.776), and NPV (0.959). Furthermore, precision-recall (PR) curves, calibration curves, and decision curve analysis (DCA) consistently confirmed the superiority of the GBM model. Additionally, the SHAP framework indicated that surgery, radiotherapy, and tumor size were the primary factors influencing the predictions made by the ML model. Conclusions: The GBM model was proven to be the most effective in predicting LM in PC patients. Its robust predictive performance and interpretability provide a reliable tool for supporting personalized clinical decision-making.