Machine learning models and nomogram based on clinical, laboratory profiles and skeletal muscle index to predict pancreatic fistula after pancreatoduodenectomy.

Background: Postoperative pancreatic fistula (POPF) is a perilous complication that may arise subsequent to pancreaticoduodenectomy (PD). In recent times, there has been an escalating interest in employing machine learning (ML) techniques to aid in treatment decision-making. The purpose of this research is to assess the effectiveness of ML in comparison to conventional models, while also conducting an initial evaluation of the predictive capability of skeletal muscle index (SMI) concerning POPF. Methods: This retrospective observational study was carried out at The First Affiliated Hospital of Wenzhou Medical University from January 2012 to January 2021, encompassing data from 269 patients who underwent PD. After identifying independent factors associated with the condition, a logistic regression model was employed to construct a nomogram, alongside the establishment of five ML models. To assess their effectiveness, the best-performing ML model and nomogram were evaluated on a separate test group comprising 77 additional patients. The evaluation involved comparing the area under the curve (AUC) and Brier score. Results: Among the 269 patients studied, the incidence of POPF was found to be 56.9%, with 106 patients (69.3%) experiencing clinically-relevant POPF. We identified six independent factors associated with POPF, including body mass index (BMI), SMI, pancreatic duct dilatation, tumor size, triglyceride levels, and the ratio of aspartate aminotransferase to alanine aminotransferase (AST/ALT) on the first postoperative day. When evaluated on the test set, the Gaussian Naive Bayes (GNB) model, which was the best-performing ML model, achieved an AUC of 0.824 and a Brier score of 0.175. The corresponding performance indicators for the nomogram were 0.844 for AUC and 0.165 for the Brier score. Conclusions: This study found that there is minimal difference between ML and the nomogram based on logistic regression in predicting POPF. Additionally, SMI shows promise as a potential and practical tool for assessing the risk of POPF.