Background and ObjectivesChronic liver diseases,which induced by kinds of causes such as viral hepatitis,alcoholism,non-alcoholic fatty liver disease(NAFLD),autoimmune liver diseases,drug-induced liver injury,inherited metabolic diseases,are resulted in liver injury and hepatocyte inflammation,will progress to liver fibrosis,and liver cirrhosis.Liver cirrhosis has become the 11 th most common cause of mortality and consists of compensated and decompensated stages.Decompensated cirrhosis is characterized by liver dysfunction and portal hypertension.Complications of liver cirrhosis including hepatic encephalopathy(HE),upper gastrointestinal bleeding(UGIB),spontaneous bacterial peritonitis,acute-on-chronic liver failure(ACLF),cirrhotic cardiomyopathy(CCM)are associated with enhanced risk of mortality.HE is one of the most severe complications in liver cirrhosis,which will increase the economic burden and reduce the quality of life due to frequent readmission.The mortality of patients with severe HE achieves 50% in the first year.UGIB is a life-threatening emergency in liver cirrhotic patients and variceal bleeding is the most common cause with a mortality rate of 20% in 6 weeks.UGIB will increase the load of heart in cirrhotic patients,and CCM is often ignored in cirrhotic patients.QTc interval prolongation is one of the characteristic electrocardiogram findings in patients with CCM,which is also used as an indicator to evaluate the prognosis of patients.The accuracies of the prognostic models such as Child-Pugh,MELD,MELD-Na,and ALBI in different liver cirrhotic patients are still controversial.Therefore,it is particularly important to clarify the risk factors and the diagnostic value of prognostic models in patients with different liver cirrhosis complications.As the development of anti-viral treatment,the prevalence of viral-related liver diseases is decreased and the prevalence of non-viral-related liver diseases such as autoimmune hepatitis(AIH),NAFLD and alcohol-related liver disease(ALD)are gradually increased.It is known that NAFLD is associated with metabolic syndrome and ALD is associated with alcohol consumption.However,the aetiology of AIH remains unclear.Diagnosis of AIH mainly depends on clinical symptoms,laboratory tests,and specific pathological features,and excluding the other liver diseases.Patients at advanced fibrosis stage are associated with increased risk for disease progression,severe liver dysfunction,and poor prognosis.Early identification and treatment may improve patients’ prognosis.Liver biopsy remains the golden criterion for diagnosis and staging fibrosis in AIH,whereas the invasiveness,sample error,differences among pathologists’ evaluation,and the low acceptance of patients limit its clinical application.Moreover,it is not a suitable tool for monitoring and follow-up.Non-invasive fibrosis tests including fibrosis index-4(FIB-4),aspartate aminotransferase to platelet ratio index(APRI),and aspartate aminotransferase-to-alanine aminotransferase ratio(AAR)based on serum indicators,which can be calculated through serum biomarkers and age and so on,but carry unsatisfactory diagnostic accuracies.Vibration controlled transient elastography,2Dshear wave elastography,and magnetic resonance elastography based on imaging technology showed better diagnostic capabilities but could not be widely used because the results may be influenced by liver inflammation,obesity,and ascites,as well as unavailable facility or skilled operator or high costs.Therefore,current non-invasive tests remain an unmet need for clinical practice.There is an urgent need to establish newly non-invasive diagnostic models to accurately predict liver fibrosis.Emerging analytic procedures such as machine learning have been proved better than traditional ways for the diagnosis,prevention,and therapeutic response of various diseases.In this study,we aimed to probe the risk factors and evaluate the accuracies of predictive models in chronic liver diseases.The study mainly divided into three parts:(1)To explore the risk factors and the diagnostic abilities of prognostic models in cirrhotic patients with HE;(2)To explore the diagnostic values of prognostic models in cirrhotic patients with UGIB;(3)To predict advanced liver fibrosis in patients with AIH using machine learning methods.Methods1.Patients admitted to our hospital between January 2016 and August 2020 were retrospectively enrolled.Patients diagnosed with liver cirrhosis and HE were extracted through electronic database.We gathered medical history,laboratory tests and HE grade.Child-Pugh,Model for End-stage Liver disease(MELD),Model for End-stage Liver disease-Sodium(MELD-Na),neutrophil to lymphocyte ratio(NLR),and albumin-bilirubin(ALBI)scores were calculated for evaluation of liver function.We applied multivariate regression analysis to explore the independent risk factors of HE severity,ACLF occurrence and in-hospital mortality.The predictive capabilities of models were calculated using the receiver operating characteristic(ROC)curve analyses and compared by the De Long tests.2.We retrospectively collected patients diagnosed with liver cirrhosis consecutively admitted to our hospital from May 2017 to May 2018.The baseline characteristics consisted of medical history,laboratory tests and QTc interval.We calculated Child–Pugh class/score,MELD,and ALBI score/class.Outcomes were defined as in-hospital mortality.We also analyzed patients with hepatitis B virus(HBV)subgroup according to aetiology.We also compared baseline characteristics of these groups.ROC curve analyses were executed for assessing the accuracies of Child-Pugh,MELD,ALBI,and QTc interval in predicting inhospital mortality of all UGIB patients and HBV groups.3.Patients with AIH who underwent liver biopsy were retrospectively analyzed.Least absolute shrinkage and selection operator(LASSO)was applied for selection of variables.The optimal hyperparameters was ascertained by 10-fold cross-validation.Six machine learning methods included decision tree(DT),logistic regression(LR),multilayer perceptron(MLP),random forest(RF),support vector machine(SVM),and extreme gradient boosting(XGB),were used to establish diagnostic models for predicting advanced fibrosis.We chose the highest areas under the curve(AUC)in the training set as the optimal model for predicting advanced fibrosis,and further compared with the traditional non-invasive models.Ultimately,we transferred the RF model into online calculator,clinicians could obtain the predictive probability of advanced fibrosis by inserting the indicators.Results1.A total of 304 patients were eligible for this study after exclusion.In multivariate regression analyses,neutrophil and total bilirubin(TBIL)were independently correlated with in-hospital death.Alanine aminotransferase(ALT)and blood urea nitrogen(BUN)were independent serum biomarkers associated with HE severity.Hemoglobin,TBIL,BUN,and international normalized ratio(INR)were significant indicators associated with ACLF incidence.For prediction of in-hospital mortality,Child-Pugh was superior to the others in the whole patients,while NLR showed the best capability in the ACLF group.2.A total of 167 cirrhotic patients with UGIB were analyzed.QTc interval prolongation were presented in 111 patients(66.5%).AUCs of Child-Pugh,MELD,ALBI,and QTc in predicting in-hospital death in whole population was 0.886,0.858,0.795,and 0.699.In the HBV subgroup,AUCs of QTc was 0.865,0.824,0.812,and 0.722.No significant differences were observed among the four prognostic models in two groups.3.A total of 233 AIH patients were enrolled in the study.Among the six methods of machine learning,RF model,which consisted of albumin(ALB),PT,and ultrasonic spleen thickness(UTST),showed the highest diagnostic performance(AUC = 0.951)in the training set.In the testing set,RF model also performed better than the other models(AUC = 0.869).In addition,machine learning presented significantly greater prediction than traditional noninvasive models,including FIB4,AAR and APRI,the AUCs of which were 0.775,0.765,and0.669,respectively in the testing set.Finally,we transferred the RF model into a web calculator,thus clinicians could obtain the prediction rate by input the indicators.Conclusion1.In cirrhotic patients present with HE,neutrophil,BUN,and liver function indicators were associated with prognosis.We also found NLR might be an effective score in the assessment of short-term prognosis in HE patients who complicated with ACLF.Child-Pugh and NLR scores may be effective prognosticators in patients with HE.2.QTc interval prolongation was prevalent in cirrhotic patients with UGIB and correlated with liver dysfunction.QTc may not be a substitute marker for prognostic models such as ChildPugh,MELD,and ALBI to predict the short-term prognosis of cirrhotic patients with UGIB.Hepatologists and cardiologists should work together to launch large sample,multicenter,and well-designed prospective studies to explore the diagnostic values of liver function models and QTc prolongation in chronic liver diseases.3.Among the six machine learning models,RF model showed the best performance in the discrimination of advanced liver fibrosis in patients with AIH,which consisted of ALB,PT,and UTST.Web calculator was helpful for improving the diagnostic efficiency in clinical practice.However,RF model should be validated in external cohort to improve its generalization performance. |