Using an artificial neural network to predict biliary fistula after pancreaticoduodenal resection

Aim . To determine the risk factors of biliary fistula after pancreaticoduodenal resection. Materials and methods. 128 pancreaticoduodenal resections were performed in the period of 2018–2023. Biliary fistula was predicted using a neural network and logistic regression. Prediction accuracy was evaluated by ROC analysis (Receiver Operator Characteristics). The DeLong test was used to compare ROC curves. Results . Biliary fistula developed in 16 patients (12.5%). Univariate analysis showed that risk factors of biliary fistula included the patient's age >70 years, Charlson comorbidity index >7 points, diabetes mellitus, postsurgical anemia, common bile duct diameter <5 mm, and pancreatic fistula. In multivariate analysis, diabetes mellitus, common bile duct diameter <5 mm, and anemia after pancreaticoduodenal resection increased the risk of biliary fistula. A prognostic multivariate model of biliary fistula development, constructed using an artificial neural network demonstrated higher sensitivity (87.5%) and specificity (95.5%) compared to the logistic regression model (68.8% and 90.2%; p = 0.03). Conclusion . The use of neural networks in predictive analysis of pancreaticoduodenal resection results can increase the efficiency of biliary fistula prediction.