Machine learning approach in diagnosis and risk factors detection of pancreatic fistula

Abstract

Abstract Purpose. The aim of the study is to develop a predictive ML model of development of postoperative pancreatic and to detect the main risk factors of the complications. Methods. We performed a single-centre retrospective clinical study. 150 patients, who underwent pancreatoduodenal resection in FSBI NMRRC, were included in the study. We developed ML models, basing on the pre- and intraoperative data and the 3-5 postoperative days data. Binary model classes were no fistula and biochemical leak or fistula B/C. 3-dimentional model distinguished no fistula, biochemical leak and fistula B/C. Logistic regression, Random forest and CatBoost algorithms were employed. The risk factor were evaluated basing on the most accurate model, roc auc, and Kendall correlation, p<0.05. Results. We detected significant positive correlation of blood and drain amylase level increase in association with biochemical leak and pancreatic fistula B/C. Catboost algorithm was detected as the most accurate, roc auc 74%-86%. Risk factors were evaluated with model parameter “importance”. Binary model, roc auc 71%, detected the main risk factors of all the fistulas on the first postoperative day: tumor vascular invasion, age and BMI. Risk factors of fistula B/C were BMI, age, tumor size and vascular invasion, the 3-dimensional model roc auc 70%. Basing on the 3-5 days data, binary model risk factors were blood and drain amylase levels, blood leukocytes, roc auc 86%. Fistula B/C risk factors were the same, the 3-dimensional model roc auc 75%. BMI and age were also important. Conclusion. We developed sufficient quality ML models of postoperative pancreatic fistulas. Blood and drain amylase level increase were the major risk factors of further fistula B/C development. Young age and high tumor size were common factors of fistulas development.