Internal Evaluation of a Machine Learning-Based Random Forest Model for Clinically Relevant Postoperative Pancreatic Fistula After Pancreaticoduodenectomy: Limited Incremental Value of Inflammatory–Nutritional Markers

Background and Objectives: This study evaluated whether preoperative neutrophil-to-lymphocyte ratio (NLR) improves prediction of clinically relevant postoperative pancreatic fistula (CR-POPF) after pancreaticoduodenectomy. Materials and Methods: This single-center retrospective cohort included 120 of 158 patients operated on between January 2019 and March 2026; 38 lacked documented pancreatic duct diameter. CR-POPF followed the 2016 International Study Group on Pancreatic Surgery definition. A random forest incorporating duct diameter, estimated blood loss, and NLR was compared with a model omitting NLR and penalized logistic regression using 10-times-repeated stratified five-fold cross-validation. Post hoc clinicopathological sensitivity analyses added pancreatic texture and tumor origin, then tested inflammatory–nutritional markers individually. Discrimination, average precision, Brier score, calibration, and exploratory decision-curve analysis were assessed. Results: CR-POPF occurred in 16 patients (13.3%). Biochemical leak occurred in 28 (23.3%). The NLR-integrated model achieved an area under the receiver operating characteristic curve (AUC) of 0.721 (95% confidence interval [CI], 0.576–0.850), versus 0.760 for the comparator and 0.736 for penalized logistic regression. Expanded baseline AUC was 0.743; adding NLR yielded 0.715. Monocyte-to-lymphocyte ratio increased AUC numerically to 0.779, but the paired difference remained uncertain (ΔAUC, 0.037; 95% CI, −0.017 to 0.096). Other evaluated markers showed no clear incremental benefit. Postpancreatectomy acute pancreatitis occurred in seven (5.8%). Conclusions: NLR did not demonstrate incremental predictive value. The monocyte-to-lymphocyte ratio finding was hypothesis-generating. Small event numbers and predictor selection outside cross-validation limit interpretation; performance may remain optimistic. These findings do not support clinical implementation without independent external validation.