The value of pancreatic imaging in predicting diabetes in patients with Hepatitis B cirrhosis

WANG Siqi; GUO Dongqiang; WU Zhifeng

Abstract：Objective:This study aims to analyze the imaging characteristics of pancreatic CT plain scans and dual-energy CT(DECT)enhanced scans in patients with hepatitis B cirrhosis.Additionally,it seeks to evaluate the efficacy of machine learning approaches in predicting the comorbidity of hepatogenic diabetes mellitus(HD)in hepatitis B cir-rhosis patients.Methods:Clinical data,including Child-Pugh grading,and images from pancreatic CT plain scans and DECT enhanced scans were collected from patients with hepatitis B cirrhosis.The imaging features were extracted using the scientific research platform and the software of"Huiyi Huiying".Subsequently,the optimal imaging features were processed by the machine learning approach,Gaussian naive Bayesian(Gaussian NB),support vector machine(SVM)and passive aggressive algorithm(PA),respectively,to predict the presence of HD in patients with hepatitis B cirrho-sis.Lastly,we compared the predictive diagnostic efficacy under three conditions:using only CT imaging data;combi-ning CT imaging and clinical features;applying three machine learning approaches.Results:In 141 patients with hepati-tis and cirrhosis,134 patients were included in this study after screening for inclusion and exclusion criteria.They were divided into HD Group(n=49)and nHD group(n=85).There was significant difference in Child-Pugh grading be-tween the two groups(P＜0.05).When combining CT imaging and clinical features,the predictive diagnostic efficacy,as indicated by the AUC values,for Gaussian NB,PA and SVM in the training group were 0.919,0.821 and 0.845,re-spectively.The data of 134 patients were randomly divided into the training group(n=94)and the test group(n=40)in a 7∶3 ratio,based on imaging features,histological features and clinical variables,respectively(Combined Model).The predictive diagnostic AUC of Gaussian NB,PA and SVM in test group was 0.579,0.614 and 0.670 respectively.Delong test showed that there was no significant difference in the efficiency of the joint model constructed by the three machine learning classification algorithms(P＞0.05).Compared with the combined model,the radiomics model had a lower predictive power.Conclusion:Machine learning approaches,utilizing the imaging characteristics of CT plain scans and DECT enhanced scans in conjunction with clinical data,demonstrate the potential to predict whether patients with hepatitis B cirrhosis have the comorbidity with HD.