Study On Risk Factors And Clinical Prediction Model Of Pancreatic Fistula After Pancreaticoduodenectomy Based On Machine Learnin

Backgrounds and Objectives:Postoperative pancreatic fistula(POPF)is a frequent complication after pancreaticoduodenectomy.Machine-learning technique was applied to identify and validate a minimum number of predicting factors in pancreaticoduodenectomy patients,aiming to capture most information of all factors modeled when predicting POPF risk.Methods:A total of 378 patients who underwent pancreaticoduodenectomy between August 2012 and July 2022 in our department were consecutively enrolled.POPF is defined according to the International Study Group for Pancreatic Fistula(ISGPF2016).The risk factors of POPF were analyzed by univariate and multivariate logistic regression analysis.Eleven machine-learning algorithms were trained.Statistical analyses were done by community PyCharm(Edition 2018.1 x64)and SPSS 27.Results:In the end,we identified a total of 47 postoperative pancreatic fistula(POPF)cases,accounting for approximately 12.4%of all patients.Univariate and multivariate analyses showed that diagnostic classification and pancreatic texture were independent risk factors for POPF.We evaluated the predictive ability of 11 machine learning methods from five perspectives.The support vector machine(SVM)showed superior performance compared to other machine learning methods,with an accuracy of 0.875,precision of 0.273,recall of 0.214,F1 score of 0.240,and AUROC of 0.673.After evaluating the importance of various factors,we ultimately selected seven factors:pancreatic texture,pathological diagnosis,intraoperative blood loss,decompression of the residual end of the stump,history of alcohol consumption,vascular resection,and preservation of the pylorus.These factors were validated for their predictive performance using a deep learning sequence model.Conclusions:Our study results indicate that diagnostic classification and pancreatic texture are independent risk factors that contribute to the development of postoperative pancreatic fistula(POPF).The SVM algorithm outperforms other commonly used machine algorithms and can effectively predict significant risks of POPF with a minimum of seven factors.Its predictive performance is comparable to that of predicting all relevant factors,providing strong support for risk management of POPF.