LIVER
Οriginal Paper
A Practical Scoring System for Predicting Cirrhosis in Patients with Chronic Viral Hepatitis Jae Youn Cheong1, Soon Ho Um2, Yeon Seok Seo2, Seung Soo Shin3, Rae Woong Park4, Dong Joon Kim5, Seong Gyu Hwang6, Youn Jae Lee7, Mong Cho8, Jin Mo Yang9, Young Bae Kim10, Young Nyun Park11 and Sung Won Cho1 Department of Gastroenterology, 3Pulmonary and Critical Care Medicine, 4Department of Medical Informatics and 11Department of Pathology, Ajou University School of Medicine, Suwon, South Korea 2 Department of Internal Medicine, Korea University College of Medicine, Seoul, South Korea 5 Department of Internal Medicine, Hallym University College of Medicine, Chuncheon, South Korea 6 Department of Internal Medicine, Pochon CHA University College of Medicine, Sungnam, South Korea 7 Department of Internal Medicine, Inje University College of Medicine, Pusan, South Korea 8 Department of Internal Medicine, Pusan National University College of Medicine, Pusan, South Korea 9 Department of Internal Medicine, Catholic University College of Medicine, Suwon, South Korea 10 Department of Pathology, Yonsei University College of Medicine, Seoul, South Korea 1
Corresponding author: Sung Won Cho, M.D., Department of Gastroenterology, Ajou University School of Medicine, San-5 Wonchon-Dong, Youngtong-ku, 442-721 Suwon, South Korea; Tel.: +82-31-219-6939, Fax : +82-31-219-5999; E-mail:
[email protected]
ABSTRACT Background/Aims: The purpose of the current study was to develop a simple model for predicting cirrhosis in chronic viral hepatitis patients and to evaluate the usefulness of decision tree algorithms. Methodology: Serum markers of fibrosis were compared with the stage of fibrosis in liver biopsy specimens prospectively obtained from 526 subjects with chronic HBV and HCV infections (estimation set, 367; validation set, 159). Results: Univariate analysis revealed that age, bilirubin, platelet count, APRI, ALP, hyaluronic acid (HA), α2-macroglobulin, MMP-2, TIMP-1, and procollagen III N-terminal peptide (PIIINP) were significantly different between patients with (F4) and without cirrhosis (F0123). Multivariate logistic regression analysis identified platelet count, HA and PIIINP as independent predictors of cirrhosis. We categorized the individual variable into the most appropriate cut-off INTRODUCTION Liver fibrosis is the excessive deposition of fibrillar extracellular matrix components in the tissue as a consequence of chronic liver damage. Histological examination of a liver biopsy specimen is regarded as the reference standard for detecting liver fibrosis, but the majority of patients are reluctant to undergo liver biopsy. Thus, cirrhosis is usually diagnosed by serum biochemical data and imaging findings in daily clinical practice. There is a strong demand for reliable, organ-specific, non-invasive biomarkers of liver fibrosis to replace the invasive method of liver biopsy, which is accompanied by a high degree of sampling error (1). The sampling error is even more pronounced when the size of the biopsy is small; a clear relationship has been established between the size of the liver tissue and the accuracy of the staging of fibrosis (2,3). As serum markers offer the opportunity of sampling from the entire organ, they offer a Hepato-Gastroenterology 2012; 59:00-00 doi 10.5754/hge10157 © H.G.E. Update Medical Publishing S.A., Athens
Key Words:
Chronic hepatitis; Liver fibrosis; Hyaluronic acid; Procollagen III N-terminal peptide; Decision tree algorithm.
value by calculating the likelihood ratio for predicting cirrhosis and constructed a score system expressed by the following simple formula: PHP index = platelet score + HA score + PIIINP score. For predicting cirrhosis, the area under the receiver operating characteristic curve (AUROC) was 0.824 and 0.759 in the es- Abbreviations: timation and validation set, respectively. Using a cut- γ-Glutamyl Transoff score of 4, the presence of cirrhosis was predicted peptidase (GGT); Area Under The with high accuracy. The diagnostic performance of the Receiver OperatPHP index was similar to decision tree algorithms (AU- ing Characteristic ROC=0.819) for predicting liver cirrhosis, but more (AUROC); Alanine useful in clinical situations. Conclusions: Compared Aminotransferase (ALT); Aspartate to a decision tree model, a simple score system using Aminotransa categorized value based on a combination of platelet ferase (AST); count, HA and PIIINP identified patients with liver cir- Alkaline Phosphatase (ALP); rhosis with a higher clinical usability.
α-Fetoprotein (AFP); Hyaluronic Acid (HA); Procollagen III NTerminal Peptide clear benefit over a biopsy that is associated with sam- (PIIINP); Positive pling error. Furthermore, liver biopsy is an impractical Predictive Value Negative way of monitoring changes in liver fibrosis in patients (PPV); Predictive Value on antiviral or anti-fibrotic treatment because of poten- (NPV); Chronic tial adverse events. Hepatitis B (CHB); Serum markers of liver fibrosis have been catego- Chronic Hepatitis C (CHC).
rized as either direct or indirect (4). Direct serum markers are constituents of hepatic matrix or enzymes involved in matrix turnover. In contrast, indirect biomarkers simply reflect abnormalities of hepatic structure or liver function and do not reflect extracellular matrix turnover. Some serum markers may be influenced by inflammation, co-morbid conditions, or medications. Therefore, combinations of serum markers are more accurate than single markers in the assessment of liver fibrosis. A growing number of indices of fibrosis incorporating combinations of direct and/or indirect markers have been developed (5-14). The most widely published
Hepato-Gastroenterology 59 (2012)
JY Cheong, SH Um, YS Seo, et al.
serum marker panel is the Fibrotest, which includes the indirect markers, α2-macroglobulin, apolipoprotein A1, haptoglobin, γ-glutamyl transpeptidase (GGT) and bilirubin (12,15-17). No serum marker panel is clearly superior to the others and most indices have an area under the receiver operating characteristic (AUROC) for significant fibrosis of 0.80-0.85. There is no currently available biomarker of fibrosis that can completely replace the liver biopsy. A decision tree is a flowchart for modeling a decision analysis. A decision tree is a reliable and effective decision-making technique that provides high diagnostic accuracy (18). No study has previously addressed the performance of decision tree algorithms for predicting liver cirrhosis. The purpose of the current study was to develop a practical and clinically useful scoring system and to compare diagnostic performance of the prediction model with decision tree algorithms for predicting liver cirrhosis in patients with chronic viral hepatitis.
Serological markers of liver fibrosis for analysis were selected to include surrogate markers of matrix synthesis or degradation, based on knowledge of the basic mechanisms involved in liver fibrosis. The serum α2macroglobulin level was measured with an automatic nephelometer (Dadebehring, Marburg, Germany). The serum apolipoprotein A1 level (Roche, Basel, Switzerland) was determined with an ELISA kit. The serum hyaluronic acid (HA) concentrations were measured with a Corgenix HA quantitative test kit (Corgenix, Inc., Westminster, CO, USA). HA tests results are expressed in ng/mL. The reference values were 1.75-4.2g/L for α2macroglobulin, 1.08-1.76g/L for apolipoprotein A1 and 10-800ng/mL for HA (detection limit, 10ng/mL). MMP-2 and TIMP-1 levels were determined with Quantikine® Immunassay kits (R&D Systems, Inc., Minneapolis, MN, USA). The procollagen III N-terminal peptide (PIIINP) level was assessed by UniQ PIIINP radioimmunoassay (Orion Diagnostica, Inc., Espoo, Finland); the reference range was 2.3-6.4μg/L.
Laboratory tests Blood analysis was performed using standard methodologies. Serum biochemical parameters included total bilirubin, alanine aminotransferase (ALT), aspartate aminotransferase (AST), GGT, alkaline phosphatase (ALP), albumin, blood urea nitrogen, creatinine, α-fetoprotein (AFP), prothrombin time, blood glucose, triglycerides and total cholesterol. The markers of hepatitis virus included HBsAg, HBeAg, anti-HBc, anti-HBs, anti-HBe, and anti-HCV antibodies, HBV DNA and HCV RNA. Serum HBV DNA concentrations ≥2х103 copies/mL were referred to as HBV DNA-positive. HCV infection was defined by a positive third-generation anti-HCV test and detection of HCV RNA on a qualitative assay.
Statistical analysis Within the estimation group the variables which differed between the patients with (F4) and without cirrhosis (F0123) were identified by univariate χ2 and ttest analyses. All variables that were significant in the univariate analysis were entered in a logistic regression model with stepwise forward selection to develop a model for predicting cirrhosis. To derive the binary or categorical predictors required for a clinically useful scoring system, we categorized continuous variables. The analysis of multi-level likelihood ratios was introduced to estimate the likelihood of having cirrhosis according to each laboratory measurement. This procedure empirically determined the cut-off point for each of the measures that best distinguished patients with cirrhosis from those patients without cirrhosis. Finally, superior cut-off points were determined by using a likelihood ratio for predicting cirrhosis. Two decision tree models were constructed using a C5.0 algorithm (Clementine 10.1; SPSS, Inc., Chicago, IL, USA) with the estimation data set. C5.0 builds decision trees using the concept of information entropy and examines the normalized information gain (difference in entropy) that results from choosing an attribute for splitting the data. The attribute with the highest normalized information gain is the one used to make the deci-
METHODOLOGY Patient selection This project was organized and performed by the clinical research center for liver cirrhosis in Korea. The cohort was composed of the estimation and validation sets for derivation of the prediction model for liver cirrhosis and subsequent validation. Between October 2005 and July 2007, 367 consecutive subjects (estimation set) with chronic hepatitis B virus (HBV) or hepatitis C virus (HCV) infections were prospectively recruited from 6 centers (Ajou University Hospital, Hallym University Chuncheon Sacred Heart Hospital, Pochon CHA University Hospital, Pusan Inje University Hospital, Pusan National University Hospital and Catholic University St. Vincent’s Hospital) in South Korea. All comparable 159 patients from the Korea University Hospital were enrolled during the same period of time and represented the external validation cohort. A total of 526 subjects were included in this study. Clinical data were collected and serum samples were obtained at the time of liver biopsy and processed immediately. Patients were excluded if they had other causes of liver disease or decompensated cirrhosis or had received antiviral treatment within the previous 6 months. The study protocols were approved by the Institutional Review Board of Human Research of Ajou University Hospital and all participating hospitals. Informed consent to participate in the study was obtained from all patients.
Histological examination The biopsy specimens were fixed with 10% formalin, routinely embedded in paraffin and the tissue sections were processed with hematoxylin and eosin, Masson’s trichrome and reticular fiber staining. A minimum length of at least 1.0cm from the liver biopsy and at least 6 complete portal tracts were required for diagnosis (19). The stage of fibrosis was determined using the METAVIR staging system, as follows: F0, no fibrosis; F1, enlarged fibrotic portal tracts; F2, enlargement of portal tracts with rare periportal or portal-portal septa; F3, numerous septa without cirrhosis; and F4, cirrhosis (20). All biopsy specimens were analyzed by two hepatopathologists (Kim YB and Park YN) who were blinded for clinical data. At the end of the study, discrepancies between the two pathologists were resolved by joint discussion and consensus at the microscope.
A Practical Scoring System for Predicting Cirrhosis sion. The algorithm then recurs on the smaller sub-lists (21). For comparison, another simple decision tree model using the same variables selected from the above logistic regression was also constructed. The diagnostic value of prediction model for predicting cirrhosis was assessed with the AUROC analysis. The sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) for various values in the final model were calculated to determine the optimal cut-off values that would predict or exclude cirrhosis. All statistical tests were two-sided and performed with SPSS software, version 11.5 (SPSS, Inc.).
RESULTS Patient characteristics A total of 526 patients were enrolled in this study (367 estimation set and 159 validation set). Patients with cirrhosis were more common in the validation set than in the estimation set (27.7% vs. 12.8%, respectively, p=0.001). In the estimation set, chronic liver disease was secondary to HBV in 288 patients and to HCV in 79 patients. The distributions of the METAVIR fibrosis stages were as follows: F0=11 (3.0%), F1=67 (18.3%), F2=126 (34.3%), F3=116 (31.6%), and F4=47 (12.8%) in the estimation set and F0=7 (4.4%), F1=29 (18.2%), F2=40 (25.2%), F3=39 (24.5%), and F4=44 (27.7%) in the validation set. None of these patients had any clinical or sonographic features of decompensated cirrhosis. Comparisons between the estimation and validation sets showed differences in platelet count, bilirubin, ALP and glucose (Table 1). There were no significant differences among the subjects between the two groups in terms of age, gender, WBC count, AST, ALT, GGT, cholesterol and prothrombin time.
Predictive index of liver cirrhosis from the estimation set The predictive index for predicting liver cirrhosis was constructed with data from the estimation set. To derive the binary or categorical predictors required for a clinically useful scoring system, we categorized continuous variables. The analysis of multi-level likelihood ratios was introduced to estimate the likelihood of having cirrhosis according to each laboratory measurement. This procedure empirically determined the cut-off point for each of the measures that best distinguished patients with cirrhosis from those patients without cirrhosis. Finally, superior cut-off points were determined by using a likelihood ratio for predicting cirrhosis. Univariate analysis revealed age (
