Analysis And Comparison Of Noninvasive Diagnosis Of Liver Fibrosis In Patients Using Machine Algorithms

ObjectiveThe purpose of this study is that summarizes the existing researched results, collect the liver biopsied patients in the Second Xiangya Hospital of Central South University, and takes the advantage of machine algorithms which developed in recent years for noninvasive indicators model of chronic hepatitis B. The established models would address a liver biopsy sampling error, and diagnose most of the CHB patients as absence or presence of cirrhosis and then be free from liver biopsy.MethodThe patients were selected with CHB and underwent liver biopsy in the Second Xiangya Hospital of Central South University from January2009to May2011. Inclusion Criteria:patient received a liver biopsy, and according to Chinese medical Association the latest "Proclaim Prevention and Cure Guide For Chronic Hepatitis B" in2010, the patient has a history of hepatitis B or HBsAg positive more than6months, and the hepatitis B virus surface antigen (HBsAg) and (or) HBV DNA is still positive. Exclusion criteria:the patient who merger of hepatitis C virus, human immunodeficiency virus (HIV) infection, and alcoholic liver disease, autoimmune liver disease, cholestasis liver disease and other chronic liver disease and liver tumors; Patients have receive antiviral treatment, and the liver biopsy samples do not conform to the standard. In order to ensure the time validity of the model, the inspected objects were taken morning fasting venous blood within1week before the liver biopsy. Serum indicators include:1) Blood routine:White Blood Cells(WBC), Platelets(PLT), Red Blood Cells(RBC), Hemoglobin(Hb), Neutrophils(Neu), Liver function:Serum Alanine Aminotransferase (ALT), Aspartate Aminotransferase (AST), AST/ALT ratio, Albumin, Globulin, Total Bilirubin (TBIL), Direct Bilirubin(DBIL), Total Bile Acid(TBA), Alkaline Phosphatase(APL), gamma Glutamyl Transpeptidase (γ-GGT), Renal function:Urea Nitrogen(UN), Creatinine(Cre), Coagulation function:Prothrombin Time (PT), prothrombin time activity percentage(PTA) and international normalized ratio(INR).2) Serum liver fibrosis indexes (ria method) hyaluronic acid(HA), Laminin (LN), type IV of Collagen (Ⅳ-C), PrecollagenⅢPeptide(PⅢP);3) All cases underwent liver biopsy after the admission were guided by the ultrasound. The length of liver tissue specimen is not less than1 cm, and containing at least6portal areas, tissue by formaldehyde fixation, paraffin embedding, sectioning, conventional HE staining and mesh dyeing. The fibrosis score of portal tract was graded by Knodell scoring system (Inflammatory necrosis grade G1～4, degree of fibrosis stages S1～4). A pathology expert identifies all the pathological images independently.According to the American Liver Disease Practice Guidelines (2004), the degree of liver fibrosis as no significant liver fibrosis was determined with S1to S2, and significant liver fibrosis was classified with S3to S4. The inflammation necrosis of G1and G2were classified as less severity, and G3～G4were classified as severity. The correlation of multiple diagnostic indexes and the degree of fibrosis (G), and the correlation of multiple diagnostic indexes and the degree of inflammatory activity (S) were analyzed using ROC curve. And the related indexes were sorted, and the high correlation indexes were selected as discriminant indicators. Based on selected indicators, advanced machine algorithm (i.e. Support Vector Machines and Random Forests) were applied to establish the noninvasive diagnosis model of liver fibrosis. The noninvasive diagnosis model of the inflammatory activity and fibrosis degree were established, and the diagnostic accuracy of various methods were compared and analyzed.ResultResults of ROC characteristic curve analysis show that the areas under ROC curves of22indicators including age, ALT, AST, ALB, GLO, TBil, DBil, ALP, GGT, BUN, Cre, WBC, Neu, RBC, Hb, PLT, PT, INR, HA, Ⅳ-C, LN, and PⅢP for the identification of significant liver fibrosis are:0.5457,0.5517,0.6562,0.3501,0.7695,0.6126,0.6107,0.7041,0.7232,0.5163,0.4051,0.3199,0.3934,0.3510,0.3761,0.3019,0.6616,0.5834,0.7525,0.6968,0.6616,0.8213; the areas under ROC curves for the degree of inflammatory activity recognition are:0.5236,0.5942,0.7038,0.3504,0.7360,0.6263,0.6492,0.7411,0.7538,0.5107,0.42020.3253,0.3770,0.3578,0.3856,0.3106,0.6776,0.6073,0.7289,0.7376,0.7046,0.8310. The indexes, of which the areas for inflammation degree and for the significant liver fibrosis are greater than0.5, are16indicators including age, ALT, AST, ALB, GLO, TBil, DBil, ALP, GGT, BUN, PT, INR, HA, Ⅳ-COL, LN, and PⅢP.The highest sensitivity index is PⅢP. The optimized indicators were selected as discriminant indexes, the models of Fisher discriminant analysis, Random Forests, Support Vector Machines were established. The accuracy of Fisher discriminant analysis, Random Forests, and Support Vector Machines for inflammation degree are0.75,0.79and0.75, respectively; The accuracy of Fisher discriminant analysis, Random Forests, and Support Vector Machines for significant fibrosis are0.86,0.93and0.89, respectively.ConclusionResults of clinical data applications show that Fisher discriminant analysis, Random Forests and Support Vector Machine can identify degree of inflammation and significant hepatic fibrosis effectively, and the accuracy of Random Forest method is the highest one, which is reach great than90%for the recognition of significant liver fibrosis, and it can be used in clinical diagnosis after the further validation.