Development and internal validation of an explainable random forest model for predicting 1-Year mortality after pancreatic resection

Objective Pancreatic resection remains one of the most technically demanding abdominal surgery procedures and is associated with substantial perioperative morbidity and long-term mortality. Although minimally invasive pancreatic surgery has been increasingly adopted, its association with longer-term outcomes remains uncertain. This study aimed to develop and internally validate a machine learning model for predicting 1-year mortality after pancreatic resection and to use SHapley Additive exPlanations (SHAP) to explore the contribution of key clinical predictors. Methods This retrospective cohort study was based on the publicly available INSPIRE database. Adult patients who underwent pancreatic resection were included, and the primary outcome was all-cause mortality within 1 year after surgery. After preprocessing, least absolute shrinkage and selection operator (LASSO) regression with class balancing was used to select the most informative predictors. Nine machine learning algorithms were evaluated within a 5-fold cross-validation framework with a synthetic minority over-sampling technique (SMOTE) applied only to the training folds. Model discrimination was primarily assessed using the area under the receiver operating characteristic curve (AUC), while calibration was evaluated by the Brier score and calibration curve. Decision curve analysis (DCA) was used to assess potential clinical utility. The best-performing model was further interpreted using SHAP. Results A total of 414 patients were included, of whom 46 (11.1%) died within 1 year after surgery. LASSO identified 10 core predictors. Among the nine evaluated algorithms, the random forest model achieved the best overall discrimination, with an AUC of 0.722 (95% CI: 0.662–0.777). The model showed acceptable calibration with a Brier score of 0.120 and demonstrated favorable net benefit across a broad range of threshold probabilities in the DCA. SHAP analysis revealed surgical approach, lymphocyte count, and age as the most influential predictors. In the interpretation of the SHAP-based model, the laparoscopic approach appeared to be associated with lower predicted 1-year mortality, and this pattern seemed more pronounced among elderly patients (aged > 65 years). Conclusions This model showed acceptable performance for predicting 1-year mortality after pancreatic resection in the internal validation cohort, and an association between the laparoscopic approach and lower predicted mortality was detected, particularly among elderly patients. These findings may support individualized postoperative risk stratification and generate hypotheses for future prospective studies.