无明显肝纤维化的HBV病人的简单非损伤性预见模型（2005.3.19）

zasxz

Identification of Chronic Hepatitis B Patients Without Significant Liver
Fibrosis by a Simple Noninvasive Predictive Model

Alex Yui Hui, M.D.; Henry Lik-Yuen Chan, M.D.; Vincent Wai-Sun Wong, M.D.;
Choong-Tsek Liew, M.D.; Angel Mei-Ling Chim, B.Sc.; Francis Ka-Leung Chan,
M.D.; Joseph Jao-Yiu Sung, M.D., Ph.D.

Abstract and Introduction
Abstract
Patients and Methods: This was a retrospective study on 235 treatment-na飗e
viremic CHB patients. Univariate analysis of data from the training cohort
(n = 150) followed by multivariate logistic regression were performed to
identify independent predictors of significant fibrosis and generate
predictive models. The models were validated with the remaining patients or
validation cohort (n = 85) and by receiver operating characteristics (ROC)
analysis.
Results: Body mass index (BMI), platelet count, serum albumin, and total
bilirubin levels were identified as independent predictors of bridging
fibrosis or cirrhosis (Ishak stage 3-6). ROC analysis was performed using
the predictive probabilities derived from the regression models. The area
under the ROC curve of the best model was 0.803 (95% CI: 0.729-0.878) for
the training cohort, 0.765 (95% CI: 0.644-0.885) for the validation cohort,
and 0.791 (95% CI: 0.728-0.854) for the entire cohort. Using the low cut-off
probability of 0.15, significant fibrosis could be excluded in 83 patients
of the total patient population (negative predictive value 0.92).
Conclusions: Our noninvasive model comprising BMI and three routine
laboratory tests was accurate in predicting absence of significant fibrosis.
Application of this model could provide useful additional information on the
stage of disease, guide future management decisions, and potentially
decrease the need for liver biopsy in some CHB patients.

Introduction
Liver biopsy has long been the gold standard in assessing histological
disease of chronic hepatitis B (CHB). It is an invasive procedure with
inherent risk.[1] The accuracy of the histological assessment of
necroinflammation and fibrosis depends on size of the specimens.[2]
Furthermore, interpretation of the biopsies carries intraobserver and
interobserver variation of 10-20% even among experienced pathologists.[3]

Noninvasive tests of liver fibrosis in chronic hepatitis C (CHC), especially
serum-based measurements, have attracted much attention in recent years.[4]
Forns et al. developed an index based on age, platelet count, cholesterol,
and ?-glutamyltransferase level.[5] The test was independently validated by
at least two other groups of investigators[6, 7]. A simple index using only
aspartate transaminase (AST) and platelet count was proposed by Wai et al.
while the algorithm used by Poynard et al. has been commercialized into
Fibrotest[8-10].

Significant differences exist between CHB and CHC not only in the etiology
but also in natural history, laboratory parameters, liver histology, and
associated medical conditions. For example, elevated alanine transaminase
(ALT) reflects accurately the necroinflammation of CHB and is used as one of
the criteria for therapy while the same could not be applied in CHC.[11]
Steatosis is an important feature of CHC but its role and importance in CHB
are less well defined.[12] The strong association of diabetes mellitus with
specific genotypes in CHC has not been found in CHB.[13]

In the current study, we aimed at identifying independent clinical and
laboratory parameter to predict liver fibrosis. A regression model was
derived based on these predictors. Using this model, a significant
proportion of patients without significant fibrosis could be detected.

Patients and Methods
Patients
The current study included patients with CHB who were recruited to
therapeutic drug trials between 1998 and 2003 in the Prince of Wales
Hospital, Hong Kong. All patients were treatment-na飗e and had HBV-DNA level
>105 copies/ml prior to entry into the trials. The ALT levels of these
patients were between 1.5? and 10? of the upper limit of normal (ULN) for
noncirrhotic patients. Level of ALT was not part of the inclusion criteria
for cirrhotic patients. Clinical, biochemical, and hematological data were
recorded from each patient within 6 wk prior to liver biopsy. Only
pretreatment biopsies were used for the current study. A total of 256
patients were recruited into the trials. Liver biopsies of 21 patients were
excluded from the current study due to either inadequate specimen for
histological diagnosis or failure to retrieve the slides. Biopsies of the
remaining 235 patients were used. All patients gave written consent on entry
to the trials for use of these data for research purposes, and the trials
were approved by the Clinical Research Ethics Committee, the Chinese
University of Hong Kong.

Histological Staging
Liver biopsy was performed using 16G biopsy needles. The specimens were
fixed, paraffin-embedded, and stained with hematoxylin and eosin (H&E). A
minimum of 1.5 cm of liver tissue with at least five portal tracts were
required for diagnosis. Histological grading of necroinflammation and
staging of liver fibrosis were performed using Knodell inflammatory score
and Ishak fibrosis score, respectively, by a single pathologist (CTL)
blinded for the clinical data. Significant fibrosis was defined as Ishak
score of 3 or more, that is, presence of bridging fibrosis or cirrhosis.[9]

Statistical Analysis
Data from a randomly generated split sample of 150 patients (65%), the
training cohort, were used to develop the model. The remaining 85 patients
constituted the validation cohort. Data were expressed as mean ?SE unless
otherwise stated. Continuous variables were compared using the Student's
t -test whereas the categorical variables were compared by ?2 test. To
formulate the predictive models for detecting significant fibrosis,
univariate analysis was performed on variables between patients with and
without significant fibrosis from the training set. The significant
variables ( p < 0.05) were then subjected to multivariate stepwise logistic
regression, using both forward and backward approaches.

Modeling
The diagnostic value of each regression model was assessed with the area
under the curve (AUC) of receiver operating characteristic (ROC) analysis
using the predictive probabilities (PP). The predictive probabilities (1 < p
< 0) are calculated according to the following formula:

PP = exp(b0 + b1x1 + b2x2 +......bnxn) / (1 + exp(b0 + b1xx + b2x2
+......bnxn))

where x is the significant predictor and b is the corresponding regression
coefficient.

Formulas that could predict significant fibrosis were constructed by
entering different sets of independent variables into the regression model.
The predictive probabilities were then used to generate ROC curves for each
individual formula. The diagnostic value was assessed by the area under ROC
curve (AUROC). The models that have AUC >0.8 were then subjected to further
validation. Data from the validation set were entered into the best models
to test the accuracy of the model. All statistical tests were two-sided and
performed with SPSS software version 11.5 (SPSS Inc., Chicago, IL).

zasxz

Identification of Chronic Hepatitis B Patients Without Significant Liver
Fibrosis by a Simple Noninvasive Predictive Model (2 of 2)

Results
Patients' Characteristics
The demographic data of the entire cohort of 235 patients are shown in Table
1 . The mean age of the cohort was 38.6 ?0.6 yr and 180 (77%) were male.
Twenty-six percent of the patients had significant fibrosis. There was no
significant difference between the training group and the validation group
in demographic, laboratory, and histological parameters.

Predictive Models of Significant Fibrosis from Training Set
Univariate analyses identified ALT and AST as predictors of Knodell score >7
( p < 0.001 for ALT and p = 0.006 for AST), that is, significant
necroinflammation. Since only aminotransferase levels were found to be
predictive of necroinflammation, which was consistent with previous
studies,[15] we did not attempt to establish a predictive model for
histological inflammatory activity. Twelve variables were associated with
significant fibrosis ( Table 2 ). These included age, body mass index (BMI),
serum albumin, total bilirubin, ALP, AST, ALT/AST ratio, alpha fetoprotein
(AFP), platelet count, INR, HBeAg positivity, and HBV-DNA. Multivariate
logistic regression was performed with different sets of the variables to
identify the independent predictors of fibrosis.

The following two models were identified as most sensitive in predicting
significant fibrosis.

(1) Model 1: PP = exp(3.148 + 0.167 ? BMI + 0.088 ? bilirubin[礛]-f"/> 0.151
? albumin[g/l]-f"/> 0.019 ? platelet[109/l])/(1 + exp(3.148 + 0.167 ? BMI +
0.088 ? bilirubin[礛]- 0.151 ? albumin[g/l]- 0.019 ? platelet[109/l]))


(2) Model 2: PP = exp(1.23 + 0.167 ? BMI + 1.191 ? ALP[/ULN]+ 0.081 ?
bilirubin[礛]- 0.139 ? albumin[g/l]- 0.017 ? platelet[109/l])/(1 + exp(1.23
+ 0.167 ? BMI + 1.191 ?ALP[/ULN]+ 0.081 ? bilirubin[礛]- 0.139 ?
albumin[g/l]- 0.017 ? platelet[109/l]))

The ROC curves of the two models for the training set, validation set, and
the entire cohort are represented in Figures 1-3. Since the two models had
comparable AUROC, the first one with only four variables was preferred.
Cut-off predictive probability was chosen based on the ROC analysis of the
training set to obtain sensitivity of at least 90% in predicting significant
fibrosis. A high cut-off point was also chosen to provide a specificity of
at least 85%.


Figure 1. (click image to zoom) ROC curves of ( A ) model 1 and ( B ) model
2 in the prediction of significant fibrosis in training set.



Figure 2. (click image to zoom) ROC curves of ( A ) model 1 and ( B ) model
2 in the prediction of significant fibrosis in validation set.



Figure 3. (click image to zoom) ROC curves of ( A ) model 1 and ( B ) model
2 in the prediction of significant fibrosis in the entire cohort.


Applying the lower cut-off PP value of 0.15 to the training set, 38 (93%) of
41 patients with significant fibrosis were correctly identified ( Table 3 ).
Using the higher cut-off value of 0.5, 91% (96/106) of patients without
significant fibrosis were correctly identified. Nevertheless, with either
cut-off points, the positive predictive values (PPV) were relatively low at
41% and 63%, respectively. When the cut-off PP values were applied to the
validation set and the entire cohort, similar results of low PPV but high
negative predictive values (NPV; 81-92%) were obtained ( Table 4 and Table
5 ). A patient with PP less than 0.15 is unlikely to have significant
fibrosis (Ishak score >3) with the NPV of 92%. Patients with PP falling
between 0.5 and 0.15 (n = 92 or 41% of all patients with available PP, Table
5 ) are also unlikely to have significant liver fibrosis with NPV of 81%.

The results strongly suggested that the utility of this model was to
identify patients without significant fibrosis. It was thus important to
review individual cases that were discordant in the prediction of absence of
significant fibrosis. Seven patients with significant fibrosis had PP <
0.15, that is, false negative results. Of these patients, 2 were cirrhotic
on liver biopsy. Both patients had sonographic features of cirrhosis and 1
had esophageal varices on endoscopy. For the remaining 5 patients, 2 had
Ishak scores of 3, and 3 had Ishak scores of 4. None of these patients had
any clinical or sonographic feature of severe fibrosis or cirrhosis.
Therefore, for all the discordant cases, at least 2 were definitely due to
misclassification by the model. For the remaining 5, it remains uncertain
whether there was genuine misclassification by the model or
misinterpretation of the biopsy.

Lastly, we compared the performance of this model directly with that of the
predictive model developed by Wai et al. for CHC.[9] In the latter, the AST
to platelet ratio index (APRI) was used with two cut-off points, 0.5 and
1.5. The AUROC of the APRI in predicting significant fibrosis for our entire
cohort was 0.673 (95% CI: 0.581-0.764). The positive predictive value and
negative predictive value using the cut-off of 0.5 were 30% and 87%,
respectively. In other words, the main utility of APRI also lay in the
exclusion significant fibrosis. We did not compare our index with Forns'
model since cholesterol and ?-glutamyltransferase levels were not routinely
checked in our patients.[5] The same applied to ?2 macroglobulin and
apolipoprotein A1, two of the components of the Fibrotest, which is not
available in our locality.[8]
Discussion
The latest AASLD practice guidelines on CHB in 2004 recommend that patients
with HBV-DNA >105 copies/ml and persistent or intermittent elevation in
aminotransferase levels should be evaluated further with liver biopsy.[11]
Histological assessment provides valuable information on grade of
necroinflammatory activity and extent of fibrosis. While decision on
treatment may be based on HBV-DNA level, biochemical and serological data
combined, in the presence of moderate or significant necroinflammation on
histology, treatment is still warranted regardless of the transaminase and
HBV-DNA levels. Liver biopsy, however, is associated with a finite, albeit
small, risk of complication of ~0.5%, patient discomfort, and expense. It is
therefore not suitable for regular monitoring of disease progression. With
advances in treatment of CHB and the now well-accepted fact that liver
fibrosis and cirrhosis are reversible, such monitoring is highly desirable.

The present study aimed at establishing a simple model, based on routine
laboratory tests, which assess the degree of fibrosis in patients with CHB.
We found that total bilirubin level, serum albumin level, platelet count,
and BMI were independent predictors of significant fibrosis. The model
generated a predictive probability valued between 0 and 1 for each patient.
We chose two cut-off values to differentiate patients with significant
fibrosis from those without. The lower cut-off value of 0.15 identified over
one-third of patients in our entire cohort as having only mild fibrosis
(Ishak stage 0-2) with a NPV of over 90%. The accuracy of the model is 60%.
This degree of accuracy compares well with similar models for CHC.[5,6] For
example, the index of Forns et al. was able to identify CHC patients with
METAVIR stage 0-1 at high certainty with NPV of 96% and an overall accuracy
of 58%.[5] More importantly, when Wai et al. 's APRI model was applied to
our cohort, it was able to identify patients without significant fibrosis
with only a NPV of 87% and diagnostic accuracy of 57%.

Thrombocytopenia as a predictive factor of severe fibrosis has been
repeatedly demonstrated by studies on CHC. In CHB, it is also a poor
prognostic indicator in patients with acute icteric reactivation of the
disease.[16] The underlying mechanism could be related to the decreased
production of thrombopoietin in fibrotic liver, and sequestration and
destruction of platelets in the enlarging spleen.[17-19] Total bilirubin and
serum albumin are components of a number of prognostic models for liver
fibrosis, cirrhosis, and hepatocellular carcinoma including the Child-Pugh
scoring system.[8,20,21] In Fibrotest developed by MULTIVIRC group, total
bilirubin is one of the six markers that make up the algorithm. In CHB
patients with decompensation, albumin is an important prognostic indicator
of survival.[22] Biological function of hepatocytes is inevitably affected
by changes in the quantity and composition of the extracellular matrix
during fibrogenesis, which could explain the predictive values of these two
parameters.[23]

An interesting finding of our study is the association between high BMI and
significant fibrosis. In CHC, high BMI is associated with less favorable
response to antiviral therapy and increased risk of significant fibrosis
after treatment.[24,25] Among CHC patients with diabetes mellitus, BMI is an
independent predictor of fibrosis.[26] The relationship between obesity and
nonalcoholic steatosis and steatohepatitis is well established.[27] In CHC,
high BMI has been shown to correlate with degree of steatosis, which in turn
could be an important cofactor in accelerating fibrosis.[12] Interestingly,
BMI has not been included in the recent studies on noninvasive predictive
models of fibrosis in CHC.[5,8,9]

Though steatosis is not regarded as a common or important histological
feature of CHB, it was present in about 40% of liver biopsies from CHB
patients in one study.[28] There was no previous investigation on possible
correlation between BMI or body weight and histological disease of CHB. Our
finding that BMI is an independent predictor of significant fibrosis
suggests that further work is warranted to look into the correlation among
BMI, steatosis, and disease activity and stage in CHB.

Aspartate transaminase is useful in the model predicting liver fibrosis in
chronic hepatitis C as proposed by Wai et al. .[9] In our cohort of CHB
patients, higher AST levels were associated with significant fibrosis on
univariate analysis but not on multivariate analysis. This may be related to
the intermittent necroinflammation pattern in CHB particularly among
HBeAg-negative patients. The severity of liver fibrosis is also affected by
the duration of necroinflammation, which in turn is dependent on the
duration of immune clearance. It is therefore logical that a snapshot
assessment of an inflammatory marker cannot offer accurate prediction of the
liver fibrosis in CHB.

Myers et al. evaluated the utility of Fibrotest in CHB patients.[15] Though
it was originally developed for CHC, the Fibrotest achieved a similar degree
of diagnostic accuracy compared to our model (AUROC = 0.78 in predicting
METAVIR F2-F4 fibrosis). Nevertheless, Fibrotest comprises biomarkers that
are not commonly measured, such as ?2 macroglobulin and apolipoprotein A1.
The same study also demonstrated that the aminotransferase alone had similar
predictive value for necroinflammation, when compared with Fibrotest plus
ALT (Actitest). Our finding of AST and ALT being the only predictors of
necroinflammation was consistent with their results.

There are limitations to our study. Our analysis included only patients who
were recruited into drug trials. Therefore, our model is not applicable to
patients with inactive virological disease (HBV-DNA <105 copies/ml). Our
cohort however contained treatment-na飗e patients with normal or elevated
ALT and all different stages of fibrosis were represented. Such patient
characteristics are similar to those of CHB patients whom physicians see and
consider treating in daily practice. They are also the ones who may require
frequent monitoring of disease progression. In this study, HBV genotypes of
patients have not been assessed. Genotype C HBV has been consistently shown
to be associated with more active liver disease than genotype B HBV in Asian
patients.[29,30] However, the determination of HBV genotype requires a
molecular diagnostic tool, which may not be convenient in routine clinical
practice. Though our model comprises complicated formulas, in this era of
easy access to sophisticated and portable computing power, it should not be
a deterrent to routine use.

As with Forns' predictive model for CHC, our index also lacks good positive
predictive value for significant fibrosis, despite the use of different
cut-off points. Even among those with PP value larger than 0.5, the accuracy
of the model in predicting significant fibrosis is unsatisfactory. Similar
pitfall could be found in CHC models in which a large percentage of patients
still fall into an indeterminate group. It thus seems that a common
shortcoming of simple predictive models based on routine tests is the
inability to classify a significant proportion of patients. Preliminary
studies on use of serum proteomics and protein glycomics in diagnosing liver
fibrosis and cirrhosis, respectively, have yielded promising results and may
help overcome this problem eventually.[31,32]

Accurate staging of the disease is just as important as the assessment of
necroinflammation because it reflects the scar formation as a result of cell
necrosis and tissue damage. The major application of our model is to predict
the absence of significant fibrosis. As such it provides valuable
information to guide management decisions. It is possible that our
noninvasive fibrosis index combined with the biochemical, serological, and
virological data will provide sufficient information that liver biopsy could
be avoided or postponed in some patients. More importantly, the noninvasive
nature of this model provides additional data for regular monitoring of
disease progression in patients who may or may not have a baseline liver
biopsy and thus facilitates revision of the management plan, which would not
be feasible with invasive and costly biopsies.

In summary, we showed that CHB patients without significant fibrosis can be
identified with high accuracy using a simple model that is composed of one
clinical and three routine laboratory variables. Our findings also suggest
significant correlation between BMI and degree of fibrosis in CHB. External
validation of our model is needed in other institutes, in patients with
lower viral load, and in patients of other ethnic groups. Prospective
studies will also be needed to evaluate its application in longitudinal
monitoring of patients undergoing therapy.

CLICK HERE for subscription information about this journal.

Acknowledgements

We thank our biostatistician, Mr. Albert Cheung, for his advice.

Funding Information

This study was supported by the Cheng Suen Man Shook Foundation for
Hepatitis Studies, Hong Kong, and Clinical Research Fellowship Scheme
(A.Y.H.) jointly sponsored by Research Grants Council and The Chinese
University of Hong Kong, Hong Kong.

Reprint Address

Henry LY Chan, M.D., Department of Medicine & Therapeutics, Prince of Wales
Hospital, 30-32 Ngan Shing Street, Shatin, Hong Kong.

Am J Gastroenterol.  2005; 100 (3): 616-623.  ?005 Blackwell Publishing