A Study Of Short Term Influencing Factors For Death And Prognosis Evaluation Models For HBV-related Acute On Chronic Liver Failure

Objective: To analyse risk factors and establish the short-term prognosis model for patients with HBV-related acute-on-chronic liver failure(HBV-ACLF)using Machine Learning classification models.Compare the predictive values of Machine Learning classification models and MELD score.Methods: A total of 131 consecutive HBV-ACLF patients that followed up for 3 month from January 2010 to April 2017 in the First Affiliated Hospital of Xinjiang Medical University were enrolled.Patients were divided into survival group and death group according to their short-term survival within 3-month.whose clinical data,biochemistry characteristics were analyzed.To analyse risk factors and establish the classification model of the patient’s survival situation useing Random learning forest,support vector machine,Adaboost algorithm,Bagging algorithm,decision tree and logistics regression.Receiver operating characteristic(ROC)curve used to assess the value of the above model and MELD score in predicting the 3-month survival.Results:Among 131 patients 61 died within three months after admission,The mortality rate is46.6%.The average accuracy of Bagging algorithm,Adaboost algorithm,decision tree,support vector machine,random forest and Logistic regression is 75.43%,65.41%,68.03%,71.46%,69.84%,71.70%,respectively,and the average accuracy of Bagging algorithm is the highest in several classification models.Under the ROC curve,the area under the ROC curve of Bagging algorithm,Adaboost algorithm,decision tree,support vector machine,random forest,Logistic regression and improved MELD score is 0.8928,1,0.8248,0.8736,1,0.85,0.6985,respectively.Under the ROC curve,the traditional MELD score is the worst,AUC=0.6985.The Bagging algorithm of the HBV-ACLF short-term influencing factors of death in order of age,PTA,PT,albumin,BUN and these variables have a high importance on short-term prognosis of HBV-ACLF,and the MELDindex score(bilirubin,creatinine,international normalized ratio,there are differences in etiology).Conclusion: the machine learning classification models have better evaluation effect on the short-term prognosis of HBV-ACLF.age,PTA,PT,albumin and blood urea were the significant risk factors of 3-month death in patients with HBV-ACLF.