Prediction of pancreatic fistula after pancreatoduodenectomy using machine learning

Abstract

Objective: to analyze the results of pancreatoduodenectomy (PD) and identify predictive risk factors for postoperative pancreatic fistula (PF) using machine learning (ML) technology.Material and Methods. A nonrandomized study of treatment outcomes in 128 patients, who underwent PD for periampullary carcinoma between 2018 and 2023, was conducted. To predict PF, the ML models based on the multilayer perceptron and binary logistic regression (BLR) in SPSS Statistics v.26, were used. The Receiver Operator Characteristics (ROC) analysis was used to assess the accuracy of the models. To compare ROC curves, the DeLong test was used.Results. Clinically significant PF occurred in 19 (14.8 %) patients (grade B according to ISGPS 2016 – in 16 (12.5 %), grade C – in 3 (2.3 %)). The data of 90 (70.3 %) patients were used to train the neural network, and 38 (29.7 %) were used to test the predictive model. In multivariate analysis, the predictors of PF were a comorbidity level above 7 points on the age-adjusted Charlson scale, a diameter of the main pancreatic duct less than 3 mm, and a soft pancreatic consistency. The diagnostic accuracy of the ML model estimated using the area under the ROC curve was 0.939 ± 0.027 (95 % CI: 0.859–0.998, sensitivity: 84.2 %, specificity; 96.3 %). The predictive model, which was developed using BLR, demonstrated lower accuracy: 0.918±0.039 (95 % CI: 0.842–0.994, sensitivity: 78.9 %, specificity: 94.5 %) (p=0.02).Conclusion. The use of machine learning technologies makes it possible to increase the probability of a correct prediction of the occurrence of pancreatic fistula after pancreatoduodenectomy.

Keywords

• pancreatoduodenectomy
• pancreatic fistula
• machine learning