Application And Value Of Artificial Neural Network Model In Noninvasive Diagnosis Of Hepatitis B Related Cirrhosis

Objective:To explore the application and value of Artificial neural network(ANN)model in noninvasive diagnosis of hepatitis B related cirrhosis.Methods:A retrospective study was conducted on the data of 1152 hepatocellular carcinoma(HCC)patients who underwent hepatectomy in the Guangxi Medical University Cancer Hospital from September 18,2013 to December 21,2017.These patients were divided into a modeling group(864cases)and validation group(288 cases)using the 3:1 random matching principle.Logistic regression analysis was used to identify independent risk factors related to occurrence of liver cirrhosis in the modeling group,and then Artificial Neural Network(ANN)model of hepatitis B associated cirrhosis noninvasive diagnosis model was established.The accuracy of the model in predicting cirrhosis in two groups was evaluated respectively by the area under the receiver operating characteristic curve(AUC),and compared with the following commonly used predictive systems: Child-pugh score,the model for end-stage liver disease(MELD)score,albumin-bilirubin(ALBI)score,aspartate aminotransferase to platelet ratio index(APRI)score,brosis index based on 4factors(FIB-4)score,aspartate aminotransferase to alanine aminotransferase ratio(AAR)score,gamma-glutamyl transpeptidase to platelet ratio(GPR)score,and then the relationship between ANN model and cirrhosis was analyzed.Results:Univariate and multivariate Logistic regression analysis in the modeling group indicated that total bilirubin(T-Bil),platelet count(PLT),prothrombin time(PT)and age were closely related to cirrhosis.Combined with the above four indicators,an ANN model for the diagnosis of cirrhosis was established.The receiver operating characteristic(ROC)curve analysis results suggest that the accuracy of the ANN model(AUC=0.757)in diagnosing cirrhosis in the modeling group was better than Child-Pugh score(AUC=0.532),MELD score(AUC=0.594),ALBI score(AUC=0.575),APRI score(AUC=0.621),FIB-4 score(AUC=0.644),AAR score(AUC=0.491),and GPR score(AUC=0.604).Similar results were obtained in the validation group(AUC: 0.767(ANN)vs 0.538(Child-Pugh score)vs0.544(MELD score)vs 0.512(ALBI score)vs 0.603(APRI score,)vs0.582(FIB-4 score)vs 0.470(AAR score)vs 0.623(GPR score)).With an optional cut-off value of 0.45,the sensitivity and specificity of the ANN model for predicting cirrhosis were 75.0% and 60.7%,respectively.The risk of cirrhosis in the high-risk group(> 0.45)was higher than that in the low-risk group(≤0.45).Conclusion:The accuracy of artificial neural network(ANN)model in diagnosing cirrhosis is better than Child-Pugh score,MELD score,ALBI score,APRI score,FIB-4 score,AAR score and GPR score,which has certain guiding significance for clinical diagnosis of cirrhosis patients and treatment decision.