Cost-effectiveness of non-invasive methods for assessment and monitoring of liver fibrosis and cirrhosis in patients with chronic liver disease: systematic review and economic evaluation

Liver biopsy

Currently, histological examination of a tiny piece of liver tissue (liver biopsy) is considered the reference standard for the diagnosis and monitoring of liver fibrosis and cirrhosis. This is usually performed through the skin under the guidance of ultrasound6–8 and involves taking a small section of the lesion using a sharp hollow needle. This can usually be performed under local anaesthesia. 6–8 The main risks of percutaneous biopsy are clinically significant bleeding (1.1–1.6%),6,7 which can be fatal. 7

Histological examination provides a spectrum of information, including liver architecture, presence and extent of steatosis, presence and grade of necroinflammation and presence and extent of liver fibrosis. It can also provide a diagnosis in cases of unexplained liver function test abnormalities. This amount of information is not provided by any non-invasive test, as they are mainly confined in the assessment of liver fibrosis. Therefore, liver biopsy will remain essential in many cases, whereas non-invasive liver function tests will be used in cases where the aetiology of liver disease is known and the clinical question is the extent of fibrosis.

Liver fibrosis is assessed in liver histological scoring systems using various staging systems that assess liver architecture and fibrosis. Such systems include Ishak, Knodell, Sheuer and METAVIR. 9,10 The METAVIR scoring system stages fibrosis in five categories, from 0 to 4, while the Ishak system stages fibrosis in seven categories, 0 to 6. Cirrhosis always represents the end stage of the spectrum and is characterised by bridging fibrosis and regenerative nodules.

It should be stressed that histological stages are descriptive semiquantitative categories that assess both liver architecture and liver fibrosis and do not provide a quantitative assessment of liver fibrosis. 10,11 The numbers that have been assigned to histological stages have no quantitative relationship between them, i.e. METAVIR fibrosis stage 2 does not mean twice the amount of fibrosis of stage 1. 12 Therefore, non-invasive fibrosis markers, which assess fibrosis quantitatively, should be ideally developed and validated with reference to a histological quantitative assessment of liver fibrosis. 13,14 Such histological methods have indeed been developed and quantify fibrosis by measuring liver collagen using digital image analysis. 15–17

As liver biopsy assesses only a tiny amount of the liver, sample variability could potentially misclassify the extent of fibrosis. In addition, histological staging is also prone to intra- and interobserver variability, even when senior liver histopathologists are involved. A French study found that, in patients with chronic hepatitis C (HCV), 35% of biopsies 15 mm in length were not categorised correctly. 18 The study suggested that a sample at least 25 mm in length is necessary to evaluate fibrosis accurately with a semiquantitative score, with the possible exception of cirrhosis. Biopsies of such length are not always feasible with one needle pass in a percutaneous biopsy and, therefore, the patient’s discomfort and also the complication rate might increase. The misclassification rate (percentage of incorrect staging of fibrosis) of liver biopsy is the source of the myth that non-invasive fibrosis tests cannot achieve a high concordance with histological stages. This is only true for non-invasive tests for which their development was independent from liver histology, such as transient elastography (TE), although the diagnostic test accuracies of such tests are also evaluated using histology; it could be argued that in certain cases the false positive or false negative of such a test compared with the result of a liver biopsy is a fault of the biopsy rather than the test itself, i.e. the test diagnosed correctly what was missed by the biopsy. However, serum non-invasive fibrosis markers have been developed and calibrated with direct reference to a set of liver biopsies. Therefore, the perfect serum marker in this case would replicate the ‘golden’ histological standard and could theoretically reach an area under the receiver operator curve (AUROC) of 1, replicating even the misclassifications of a liver biopsy. 19

Non-invasive fibrosis tests

During the last few years, there has been an explosive development and use of non-invasive fibrosis tests. 20,21 These tests in many cases have replaced liver biopsy in clinical practice in the staging of fibrosis and follow-up of patients with established chronic liver disease, especially in patients with chronic HCV. The non-invasive liver tests (NILTs) can be broadly divided into three categories: simple or indirect serum markers, direct serum markers and imaging modalities.

Indirect serum markers or class II biomarkers consist of the combination of routine biochemical or haematological tests, such as transaminases, platelet count and albumin, and patient demographics that are associated with fibrosis, such as age or the presence of diabetes. 20 These tests usually have dual cut-offs: a high cut-off with high specificity and a low cut-off with high sensitivity. Depending on the clinical scenario and the disease prevalence, the low or high cut-off is used at the expense of increased false positives and false negatives, respectively. If these cut-offs are combined, then the number of false positive and false negatives are minimised; however, a number of patients will fall in the indeterminate range of fibrosis (i.e. their score will be between the low and the high cut-off) and will need either further non-invasive testing or a liver biopsy. Commonly used indirect serum markers are FIB-4, aspartate aminotransferase (AST)-to-alanine aminotransferase (ALT) ratio and APRI (AST to platelet ratio index).

Direct serum non-invasive tests or class I biomarkers are intended to detect extracellular matrix turnover and/or fibrogenic cell changes. The most common markers used in current assays involve measuring products of extracellular matrix synthesis or degradation, and the enzymes that regulate their production or modification, such as hyaluronic acid, serum collagenases and their inhibitors and profibrogenic cytokines. It should be noted that these markers are not exclusively found in liver tissue; therefore, they reflect fibrogenic processes in various other organs. For instance, the enhanced liver fibrosis (ELF) biomarker is influenced by age and sex. 22 Moreover, their sensitivity is low in the initial stages of fibrosis.

Various direct and indirect tests have been combined in patented commercial algorithms that improve the diagnostic accuracy of tests when used singly. These are ELF, Fibrotest, Fibrospect, Fibroindex and Fibrometer. Of the tests, Fibrotest (Fibrosure in the USA) is the most widely validated panel; it consists of five parameters, namely total bilirubin, haptoglobin, gamma-glutamyl-transpeptidase, α2-macroglobulin and apolipoprotein A1, and has been studied in viral hepatitis, non-alcoholic fatty liver disease (NAFLD) and alcoholic liver disease (ALD). 23 The Fibroindex was developed for patients with chronic HCV and uses platelet count, AST and g-globulin levels. 24 Fibrospect includes hyaluronate, tissue metalloproteinase (TIMP)-1 and α2-macroglobulin, and is validated in chronic HCV. 25 Fibrometers are a family of six blood tests: one for staging and one for quantifying liver fibrosis in each of the three main causes of liver disease (chronic viral hepatitis, ALD and NAFLD). 26 The ELF biomarker is a panel of direct noninvasive markers that includes hyaluronic acid, type III collagen and TIMP-1. 27 It has been used in patients with chronic HCV and NAFLD.

New imaging modalities offer better sensitivity and specificity than conventional techniques, such as ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI). The last of these can only identify cirrhosis, based on imaging findings of coarse echo-texture, collaterals suggestive of portal hypertension and nodularity. These new modalities measure liver elasticity or liver stiffness based on ultrasound or magnetic resonance (MR) techniques. The most widely used imaging modality is TE or Fibroscan (Echosens, Paris). 28 Briefly, vibrations of mild amplitude and low frequency are transmitted by an ultrasound transducer, inducing an elastic shear wave that propagates within the liver. Pulse-echo ultrasonic acquisitions are performed to follow the shear wave and measure its speed, which is directly related to the tissue stiffness (the harder the tissue, the faster the shear propagates). Results are expressed in kilopascals (kPa) and correspond to the median value of 10 validated measurements ranging from 2.5 to 75 kPa, with 5.5 kPa reported to define normality. The volume of liver tissue evaluated by TE approximates a cylinder 4 × 1 cm which is at least 100 times bigger than a liver biopsy. Moreover, TE is painless and rapid (< 5 minutes) and thus highly acceptable for patients.

Other modalities include acoustic radiation force impulse (ARFI)29 and MR elastography. 30 ARFI allows the evaluation of liver stiffness in a region of interest (ROI) involving mechanical excitation of tissue by the use of short-duration (≈262 μs) acoustic pulses while performing a real-time B-mode conventional hepatic ultrasound. Results are expressed in m/s. Although the volume of liver explored is smaller than that for TE (10 mm long × 6 mm wide), a critical advantage is the possibility to choose the representative area of interest, thereby avoiding large vessels and ribs. An advantage over TE is that it can be easily incorporated into a modified ultrasound machine. MR elastography uses a modified phase-contrast method to evaluate the propagation of the shear waves within the liver. It is a very promising technique but is not yet widely available and cost might be an important limiting factor.

Finally, algorithms of sequential or contemporary use of NILTs have been used mainly in chronic HCV, to improve the diagnostic accuracy of single tests. 31 These are typically based on an agree–disagree scenario or the sequential use of a second test if the result of the first test falls in the grey zone of an indeterminate result.

A major limitation of all the above NILTs is the absence of uniformly established and validated cut-offs for specific aetiologies of liver disease and fibrosis stages and the poor methodological quality of many of the published studies. In a recent meta-analysis on TE, only 6 of 41 included studies had both histological evaluation and Fibroscan measurements optimally performed, while all studies had a high risk of bias based on quality assessment by the Quality Assessment of Diagnostic Accuracy Studies (QUADAS) tool. 32

Criteria for considering studies for review

The aim of the systematic review was to identify papers comparing the diagnostic accuracy of different non-invasive tests in the diagnosis and monitoring of liver fibrosis and cirrhosis with liver biopsy, and to synthesise the outcomes where possible. We included studies providing cross-sectional information of the index test(s) and reference test. In other words, we included all studies that reported staging of fibrosis by index test(s) and reference standard so that it is possible to know how many patients had a certain stage of fibrosis by index test and reference test (true positive), how many had that stage by index test but not on the reference test (false positive), how many did not have that particular stage by index test but were found to have that stage by reference test (false negative), and how many patients did not have a certain stage of fibrosis by index test or reference test (true negative) in the appropriate patient population, irrespective of language or publication status, or whether data were collected prospectively or retrospectively.

We also included comparative studies in which the different index tests were performed in the same study population, or studies in which different individuals in the study population received different index tests, and the choice of tests that the different individuals received were determined in a random manner or if all the patients underwent both the index tests that were assessed. We excluded diagnostic case–control studies from the analysis if there were at least four cross-sectional or comparative studies for that test. We also excluded studies where the maximum interval between the reference standard (liver biopsy) and the non-invasive fibrosis test (index test) was > 6 months.

Participants

Adult patients with chronic liver disease (irrespective of the aetiology and clinical presentation). Studies reporting on paediatric patients were excluded.

Index tests

Ultrasound, CT scan, MRI, elastography (TE by ultrasound or MR elastography), and direct and indirect serum markers (such as AST–ALT ratio, APRI, ELF test, Fibrotest, etc.).

Target condition

Liver fibrosis and cirrhosis.

Reference standards

Histopathological examination of liver tissue (percutaneous or transjugular or laparoscopic biopsy). The staging and grading of liver biopsy can be performed by various histological scoring systems such as Ishak, METAVIR, Knodell and others. 57 We included studies irrespective of the histological scoring system used. For data synthesis and analysis we transformed the histological scores used in individual studies to METAVIR for HBV, HCV and alcohol and to Kleiner for NAFLD/NASH as these are the most commonly used histological scores. Conversion of various histological stages to METAVIR is shown in Table 1.

Electronic searches

The following databases were searched from 1988 until April 2012: MEDLINE (PubMed), EMBASE, Science Citation Index Expanded, Bioscience Information Service (BIOSIS), Cochrane Central Register of Controlled Trials (CENTRAL), Latin American and Caribbean Health Sciences Literature (LILACS) and Cumulative Index to Nursing and Allied Health Literature (CINAHL). 58,59

The search strategies for the different databases are provided in Appendix 1.

Initially, we did not use any filter; however, this yielded 200,000 references and a compromise had to be arranged, as it would not be possible to complete the analysis within the time scale allowed for this study. Therefore, a methodological filter is included but does not act as a filter for all search results (see Appendix 1). This represents a potential limitation in our search strategy.

Searching other sources

Reference lists of identified studies and reviews, and conference proceedings from the recent hepatobiliary conferences (last 2 years), were hand-searched to identify further studies.

Selection of studies

The references were searched by two researchers independently for identification of relevant studies. No restrictions were placed on the language or the publication status (full text vs. abstract from conference proceedings). However, studies which reported on a total of fewer than 10 patients with fibrosis or cirrhosis were excluded. Full texts were obtained for the references that at least one of the reviewers considered relevant. Full-text articles were then used to include or exclude studies for the review.

Data extraction and management

Data were extracted by two reviewers independently. Any differences in the data extraction were resolved by the lead applicant, Professor Burroughs, and Dr Gurusamy. Data necessary to calculate the true positive, false positive, true negative and false negative diagnostic test results were extracted using the reference standard of liver biopsy. If the information on true positive, false positive, false negative and true negative diagnostic test results were not available directly, these were calculated from information available in the study. Data were entered into a Microsoft Excel (Microsoft Corporation, Redmond, WA, USA) file created for the purpose.

The following data were extracted:

1. year of publication

2. country/ethnicity of included patients

3. inclusion criteria

4. exclusion criteria

5. total number of patients

6. patients included in the analysis

7. mean age

8. mean body mass index (BMI)

9. sex

10. mean ALT

11. aetiology of liver disease

12. technical failure in undertaking liver biopsy or non-invasive tests

13. non-invasive test used

14. fibrosis histological scoring system used

15. non-invasive test cut-off for diagnosing specific fibrosis stages

16. distribution of patients across histological stages

17. sensitivity, specificity, true positive, false positive, false negative, true negative of non-invasive test for diagnosing different histological stages

18. number of patients with uninterpretable liver biopsies or index tests

19. number of patients with indeterminate non-invasive test for a specific fibrosis stage

20. methodological quality using the QUADAS-2 assessment tool.

Assessment of methodological quality

The quality of the studies was assessed independently by two reviewers using the QUADAS-2 assessment tool. 60–62 This tool comprises four domains: patient selection, index test, reference standard, and flow and timing. Each domain is assessed in terms of risk of bias, and the first three domains are also assessed in terms of concerns regarding applicability. Signalling questions are included to help judge the risk of bias. The quality criteria that were derived from the QUADAS-2 tool and were assessed are shown in Table 2.

Statistical analysis and data synthesis

The data obtained from the various studies are presented in the form of summary sensitivity and specificity with their corresponding 95% confidence intervals (CIs). The data were combined using the bivariate random-effects model with correlation between sensitivity and specificity63 using the METADAS macro developed by the Systematic Review Diagnostic Test Accuracy Group64 in the SAS 9.2 statistical software (SAS Institute Inc., Cary, NC, USA). We calculated the summary sensitivity and specificity at specific thresholds for tests with explicit thresholds such as serum markers and calculated the overall summary sensitivity and specificity for tests that do not have an explicit threshold (such as ultrasound).

The bivariate model allows for meta-analysis of diagnostic test accuracy studies to be conducted in which both the index test under study and the reference test (gold standard) are dichotomous. Bivariate analysis involves statistical distributions at two levels. At the lower level, it models the cell counts in the 2 × 2 tables extracted from each study using binomial distributions and logistic (log-odds) transformations of proportions. At the higher level, random study effects are assumed to account for heterogeneity in diagnostic test accuracy between studies beyond that accounted for by sampling variability at the lower level. 65

If the results did not converge using the above random-effects model with correlation between sensitivity and specificity, we performed the meta-analysis with variations of bivariate analysis. The variations included different assumptions such as no correlation between the sensitivity and specificity in the studies; random-effects model for sensitivity but fixed-effect model for specificity; fixed-effect model for sensitivity but random-effects model for specificity; and fixed-effect models for both sensitivity and specificity (Takwoingi, University of Birmingham, March 2013, personal communication).

It must, however, be pointed out that the assumptions used to perform the above analysis (e.g. if one assumes that there is no correlation between the sensitivity and specificity, one has to ensure this from a scatterplot and correlation coefficient, and when one assumes a fixed-effect model, the values should be relatively close to each other) were not always met and the summary values of a model that converged was used. This could have resulted in a biased effect estimate. The alternative was not to conduct a meta-analysis for those tests which would have meant that the information could not be used in the cost-effectiveness analysis.

We also calculated the median, the lowest and the highest prevalence for the specific stages of fibrosis in the studies included.

Investigations of heterogeneity

The following sources of heterogeneity were explored.

1. Studies of high methodological quality versus low methodological quality.

2. Different stages of fibrosis (different scoring systems were converted to comparable stages in METAVIR in viral diseases and alcohol, and to Kleiner scoring system in NAFLD).

3. Different reference histological scoring systems (e.g. Ishak scoring, METAVIR, Knodell score, etc.). 57

4. Different aetiological diagnosis (e.g. ALD, HCV infection, etc.).

5. Different threshold levels for classification of positive and negative results. We performed a meta-analysis for every possible cut-off in each fibrosis stage of the reference standard.

6. Studies not published in full text were compared with studies published in full text.

7. Different ranges of transaminases (normal, between normal and up to three times the normal level, and more than three times the normal level).

Comparators

Where a large number of applicable NILTs were located by the systematic literature review (HBV and HCV), a two-stage approach to the analysis was conducted. The first stage compared each NILT identified from the systematic review (per aetiology) with each other and with liver biopsy. Where analyses involved treatment (HBV, HCV and cirrhosis), two additional testing approaches were included: a ‘treat all’ approach, where everyone is treated, and a ‘no treatment’ approach, where no diagnostic tests or treatments are administered.

The second stage of the analysis evaluated comparisons of sequential testing strategies, again compared with each other, biopsy and the treat-all and treat-no-one strategy. For this, combinations of the two most cost-effective tests within each category were chosen based on an incremental analysis using a cost-effectiveness threshold value of £20,000. 66 We assumed a decision rule whereby the two most cost-effective tests from each category were combined with tests from the other categories (reflecting combinations which would happen in actual current practice or potential future practice). Some of the NILTs evaluated have defined ‘low’ or ‘high’ cut-off thresholds and were analysed as separate test options. Combinations of tests considered to be clinically implausible were excluded; for example, a NILT with a low cut-off diagnostic threshold would not be followed by a second NILT with a low cut-off diagnostic threshold in practice but could be followed by a test with a high cut-off threshold. The following assumptions were made when combining the tests:

• If the first NILT used was an indirect serum marker, a patented or direct serum marker or an imaging modality or liver biopsy could be administered as a second test.

• If the first NILT used was a direct or patented serum marker, an imaging modality or liver biopsy could be administered as a second test.

• If the first test used was an imaging modality, a liver biopsy could be administered as a second test.

The analysis also assumed that the sensitivity and specificity of each test were independent of each other, i.e. there was no correlation of sensitivities and specificities of the tests used in the first stage and the second stage. The combinations were assumed to take four possible sequential testing strategies (Table 3).

The probabilities of having each of the four possible diagnoses (true negative, true positive, false negative, false positive) for the four sequential testing strategies were determined by multiplying the probabilities (i.e. using decision tree calculation methodology: multiplying probabilities along pathways from left to right to estimate the probability of each pathway).

Each of the sequential tests were compared with each other; liver biopsy alone; ‘treat all’ and ‘no treatment’ approaches; each cost-effective test singly; reported tests which used a combined cut-off; and any reported tests whose efficacy was estimated using a published algorithm derived from two or more tests used sequentially.

Synthesis of economic evidence

A decision tree model was constructed to estimate the cost-effectiveness of all comparators. Sensitivity and specificity data included in the decision tree were extracted from the meta-analysis (see Chapter 4).

Long-term costs and QALYs were taken from the literature if estimates specifically matching the decision tree pathways were available. Where this was not possible, long-term costs and outcomes were estimated using a series of Markov models. Figure 1 depicts the flow of data between the different modelling elements for the models estimating incremental cost per QALY.

FIGURE 1.

Illustration of data flow (input into decision tree from other modelling elements). FN, false negative; FP, false positive; TN, true negative; TP, true positive.

The watchful waiting strategy incorporated a retest every 2 years. We assumed that the retest would have perfect sensitivity and specificity in the base case for modelling practicality due to the large number of applicable NILTs identified.

Literature review

Literature searches were undertaken to identify incremental-cost-per-QALY analyses of the non-invasive tests for each aetiology. Titles and abstracts were reviewed and full papers were retrieved if deemed relevant. If existing systematic reviews were available, these were reviewed and the searches updated and/or amended as required.

Studies were excluded if not in the English language due to resource limitations. We gave preference to UK-based studies for cost data as there may be transferability issues using data from other populations due to underlying differences between the populations.

Literature searching was undertaken to populate input parameters for the models (for natural history, costs and QALY inputs). Titles and abstracts were reviewed and full papers were retrieved if deemed relevant. We started by identifying existing recent reviews. The papers identified in these were reviewed. The searches were updated, amended if needed, and rerun.

For data on natural history, inclusion criteria related to the population of interest. Judgements about the relevance of studies also took into account the country of origin (preference for UK data), high-quality and recent studies. For cost studies, those reporting data from a NHS perspective were preferred. For studies reporting health-related utility inclusion criteria requiring data from the population of interest (depending on aetiology), information on health had to be collected directly from patients and the method of preference elicitation had to be a choice-based method (e.g. time trade-off) in a UK population. As per standard NICE methods guidance,66 data obtained through the EQ-5D measure were preferred.

Further details of the search results are described in the cost-effectiveness chapters (see Chapters 5–9) and the individual search strategies are listed in Appendix 2.

Costs

All unit costs reported in the analysis for health states and liver biopsy are priced for the year 2012. Where required, costs were inflated to 2012 prices using NHS inflation indices. 67 Test costs for the NILTs are costed for the year 2012–13 as costs for some of the components for the NILTs were sourced in early January 2013.

Incremental-cost-per-quality-adjusted-life-year analyses

All analyses were fully incremental. In the incremental analyses, test strategies were ordered according to the least effective and test strategies which were found to be more costly and less effective (‘dominated’) than another strategy were ruled out of the analysis. Incremental cost-effectiveness ratios (ICERs) were calculated for the tests that were not dominated and test strategies with an ICER greater than that of a more effective intervention (‘extendedly dominated’) were also ruled out; the ICER was calculated using the formula

where C₁ equals the cost of strategy 1, C₀ equals the cost of (the next best) strategy 0, E₁ equals QALYs from strategy 1 and E₀ equals QALYs from (the next best) strategy.

The cost-effectiveness results for the remaining strategies which were not ruled out (not ‘dominated’ or ‘extendedly dominated’) were presented as ICERs. 68

Probabilistic sensitivity analysis (PSA) was conducted. With a PSA, rather than using the average values for each parameter input, the value is instead drawn from a distribution. The probability distribution for each input variable (natural history data, mortality rates, costs, QALYs, treatment effectiveness and test effectiveness) was constructed using estimates of the mean value and standard error (if required for probability distribution) and Monte Carlo simulation was used to randomly sample from each input distribution simultaneously for 1000 runs of the models. For each of the decision tree model outputs (1000 simulation runs), an average total lifetime cost and QALY was calculated for each testing strategy.

To summarise the uncertainty around the cost-effectiveness result, we constructed cost-effectiveness acceptability curves (CEACs) which are derived from a joint distribution of the costs and effects (QALYs) to represent the probability that a testing option is cost-effective (had the highest net monetary benefit) at different levels of a cost-effectiveness threshold (varied from £0 to £60,000 in analysis). Net benefit is calculated using the formula

where E is equivalent to the health outcome for a testing strategy, CR equals the ceiling ratio which is the cost-effectiveness threshold (range between £0 to £60,000) and C equals the cost of the testing strategy. 69

The CEAC represents the probability that a testing option has the highest probability of being cost-effective over a range of threshold values. However, as Fenwick et al. 69 have shown, the testing option with the highest probability of being cost-effective may not necessarily have the highest expected net benefit. In this case, the CEAC should not be used to identify the optimal option; instead, the cost-effectiveness acceptability frontier (CEAF) which plots the uncertainty associated with the optimal testing option (option with highest expected net benefit) for different cost-effectiveness threshold values may be more applicable.

We also present the CEAFs to illustrate the probability of any testing strategy being optimal (has the highest expected net benefit) compared with each other over a range of different cost-effectiveness thresholds (threshold value range varied from £0 to £60,000). 69,70

Cost per correct diagnosis (alcoholic liver diseases and non-alcoholic fatty liver disease)

The cost-per-correct-diagnoses analyses are presented incrementally. We carried out a probabilistic analysis where we estimated the number of correct true responses for each tests (positive and negative responses). We then compared the results of each test incrementally using the cost for each test to rule out tests which were more costly and provided less correct results. Liver biopsy was included as a comparator in the cost-per-correct-diagnosis analyses.

Description of studies

The search strategy initially retrieved 114,071 studies, or after duplicate exclusion, 91,097 studies. The flow chart is shown in Figure 2. Finally, data from 302 studies were analysed (HCV n = 162, HBV n = 52, NAFLD n = 48, radiology n = 60, ALD n = 12). 23–31,71–363 All but five of the included studies were captured by the search strategy79,256,276,332,334 These five studies were retrieved by manually searching the reference lists of included studies and published meta-analyses.

FIGURE 2.

Flow chart of literature review search.

Results: hepatitis C virus

Data on patients with HCV were extracted from 162 studies. 23–29,31,71–224 Meta-analysis was performed separately for each non-invasive test which had been assessed at each METAVIR stage (F1–F4). Summary sensitivity and specificity for F2 and F4 are shown in Tables 5 and 6, while the sensitivity and specificity estimates for F1 and F3 are reported in Appendix 3. Individual study characteristics are shown in Appendix 4. The median prevalence (minimum–maximum) of fibrosis stages F1–F4 in included studies was for F1 0.875 (0.157–0.968), F2 0.522 (0.063–0.893), F3 0.291 (0.051–0.778) and F4 0.17 (0.026–0.681). Forest plots and summary receiver operating characteristic (SROC) plots of different NILTs across fibrosis stages are presented in Appendices 5 and 6, respectively.

For the diagnosis of fibrosis stage ≥ F2, which was the one mainly used in economic modelling, 19 non-invasive tests were evaluated in single studies. Of 47 different evaluated tests, only 18 converged with the bivariate random-effects model [APRI low and high cut-offs, AST–ALT ratio, FIB-4 low and high cut offs, Forns index low and high cut-off, Göteborg University Cirrhosis Index (GUCI), Lok’s index, platelet count, hyaluronic acid, Hepascore, Fibrometer, Fibrotest standard, low and high cut-offs, platelet-to-spleen-diameter ratio and Fibroscan]. The most commonly evaluated non-invasive tests were APRI (low cut-off), which was evaluated in 47 studies,24,31,73,74,79,81,84,85,89,90,91,94,97,98,103,107,109,121,126,127,130,131,134,137,140,143,144,146,150,152–154,156–158,163,164,168,182,185,187,189,194,195,210,218,220,223 followed by TE in 37 studies28,29,75,76,86–88,91,95,98–100,102,105,106,110,116,119,130,137,141,147,153,155,159,161,164,170,172,173,194,199–201,211,223,224 and APRI (high cut-off) in 37 studies. 24,31,72–74,79,81,89,90,91,94,97,100,103,121,123,126,131,134,140,143,146,150,152–154,156–158,182,187,195,209,210,218,220,223

For the diagnosis of cirrhosis, there were 37 different evaluated tests; however, only nine converged with the bivariate random-effects model (APRI low and high cut-offs, AST–ALT ratio, platelet count, hyaluronic acid, Hepascore, Fibrotest and Fibroscan).

For the diagnosis of fibrosis stage ≥ F1, there were only five tests that reported diagnostic accuracy; however, none converged with the bivariate random-effects model.

For the diagnosis of fibrosis stage ≥ F3, there were 37 different evaluated tests, of which six converged with the bivariate random-effects model (APRI high cut-off, FIB-4 low and high cut-offs, Hepascore, Fibrotest and Fibroscan).

Uninterpretable NILT results were very rare in serum markers (< 1%) and were more frequently encountered in patients who were undergoing Fibroscan examination. The rate of uninterpretable results with Fibroscan (due to < 10 valid measurements, success rate < 60% and interquartile range > 30%) was 8.5%; however, this could be underestimated due to under-reporting.

Cut-offs of non-invasive tests for specific disease stages varied among studies and were predetermined in only 51 studies (31.4%). 25,31,72–74,83,85–87,90–92,94,95,97,99,100,106,109,117,119,127,132–134,143,144,146,150,154,156,157,167,169,171,172,175,186,187,192–194,203,213–216,220,222,223,366 We did not include data on APRI for cirrhosis from some studies in the meta-analysis because some cut-offs differed significantly from what is used in the literature. 82,137,168,194 Liver biopsy was of acceptable quality (≥ 15 cm in length with ≥ 6 portal tracks) in only 20 studies (12.3%),75,86,110,112,115–117,119,124,127,129,137,141,143,153,186,188,194,222,223 while minimum sample requirements were not reported in 84 (51.8%) studies. 26,27,72–74,77–80,84,85,88,91–93,96,100,101,103,104,106,107,111,114,118,121–123,125,126,130,131,133,135,138,140,147,148,151,152,159–163,165,167–169,172–175,177,178,180,181,183–185,190–192,196–204,208–215,218,221 Overall, only three studies86,143,222 had a low risk of bias in all of the domains of the QUADAS-2 tool; therefore, all our estimates may be biased. Quality assessment of included studies based on QUADAS-262 is shown in Table 7. Studies that were judged as low risk of bias or unknown in the three most important QUADAS domains, namely patient selection, index test and reference standard, were still a modest fraction of the total number of studies (29 out of 152; 19%). 72,73,79,80,86,91,92,106,110,133,135,143,148,165,167,173,197–199,201,203,206–208,211,214,215,221,222

We explored potential sources of heterogeneity as outlined in the methods section. As all but five studies were of low methodological quality,86,127,143,186,222 this potential source of heterogeneity could not be assessed. Significant heterogeneity was found mainly in relation to the transaminases level, without, however, a particular pattern in relation to the transaminase elevation. More specifically and according to source of heterogeneity, details of significant associations are:

• Full text versus abstract publication: Fibrotest for ≥ F2 (significantly higher specificity if published in full text) and borderline for ≥ F3 (likelihood ratio test, p = 0.059) and platelet count for F4 (significantly lower specificity if published in full text).

• Transaminases levels (comparator is normal transaminases): for ≥ F2 Fibrotest, hyaluronic acid, Hepascore, Lok’s index, Fibroscan; for ≥ F3 FIB-4 low and high cut-off, TE; and for F4 AST–ALT ratio. There was no specific pattern of influence; therefore, the test could have improved or worse diagnostic accuracy if the transaminases levels were high.

• Histological score used (comparator is the METAVIR system): for Hepascore in ≥ F3, use of Ludwig scoring system in one study76 resulted in significantly lower sensitivity. This might reflect the particular study rather than the histological score used; for AST–ALT ratio in F4, use of Ludwig scoring system resulted in significantly higher sensitivity; and for platelet count in F4, use of Scheuer or Knodell resulted in significantly higher sensitivity and specificity, respectively.

Results: hepatitis B virus

Data on patients with HBV were extracted from 52 studies. 116,119,120,200,225–272 Meta-analyses were performed separately for each non-invasive test assessed at each METAVIR stage (F1–F4). Summary sensitivity and specificity for fibrosis stages F2 and F4 are shown in Tables 8 and 9, whereas summary sensitivity and specificity for fibrosis stages F1 and F3 are reported in Appendix 3. The median prevalence (minimum–maximum) of fibrosis stages F1–F4 in included studies was for F1 0.617 (0.416–0.884), F2 0.528 (0.269–0.915), F3 0.370 (0.171–0.780) and F4 0.209 (0–0.604). Individual study characteristics, and forest plots and SROC plots of the different NILTs across fibrosis stages, are presented in Appendices 4, 5 and 6, respectively.

Overall, there were 18 different non-invasive tests reported for the diagnosis of ≥ F2, of which 13 (72%) were reported in single studies. Of 18 different evaluated tests, only five converged with the bivariate random-effects model (APRI low and high cut-off, FIB-4 low cut-off, Fibrotest and Fibroscan). The most commonly evaluated non-invasive tests were Fibroscan (13 studies),116,200,225,230,241,242,246,247,250,251,263,264,272 APRI (low cut-off, eight studies),225,233,256–259,269,272 APRI (high cut-off, six studies)225,256–259,270 and Fibrotest (six studies). 225,247,249,255–257

For the diagnosis of cirrhosis, there were 14 different evaluated tests; however, only Fibroscan converged with the bivariate random-effects model.

For the diagnosis of fibrosis stage ≥ F1, there were eight tests that reported on diagnostic accuracy; however, none converged with the bivariate random-effects model.

For the diagnosis of fibrosis stage ≥ F3, there were 10 different evaluated tests, of which only Fibroscan converged with the bivariate random-effects model.

We explored potential sources of heterogeneity as outlined in the methods section. As all but one study200 were of low methodological quality, this potential source of heterogeneity could not be assessed. Significant heterogeneity was only found in relation to the transaminases level (comparator is normal transaminases): lower specificity for FIB-4 low cut-off in F2, and borderline lower specificity for Fibrotest in F2 (p = 0.055).

Cut-offs for specific histological stages were predetermined in 11 studies (21%). 119,200,225,230,232,252,253,255,258,264,271 Liver biopsy was of acceptable quality in 12 studies (23%). 116,119,200,225,227,230,234,251,263,266,271,272 We did not include data from some studies on APRI for F2 and F3,242 on Forns index for F2269 and on AST–ALT ratio for cirrhosis254 in the meta-analysis because of cut-offs that differed significantly from what is used in the literature. Only one study257 had low risk of bias in all of the domains of the QUADAS-2 tool; therefore, all our estimates may be biased and should be assessed with caution. Studies that were judged as low risk of bias or unknown in the three most important QUADAS domains, namely patient selection, index test and reference standard, were still a fraction of the total number of studies (11 out of 52; 23%). 225,235,236,241,242,244,246,257,263,265,271 Quality assessment of included studies based on QUADAS-262 is shown in Table 10.

Results: non-alcoholic fatty liver disease

Data on patients with NAFLD were extracted from 48 studies. 117,119,165,229,284–327 Meta-analysis was performed separately for each non-invasive test assessed at each Kleiner stage (F1–F4). Summary sensitivity and specificity for F3 and F4 are shown in Tables 11 and 12, while summary sensitivity and specificity for F1 and F2 are reported in Appendix 3. The median prevalence (minimum–maximum) of fibrosis stages F1–F4 in included studies was for F1 0.588 (0.367–0.814), F2 0.319 (0.119–0.526), F3 0.186 (0.050–0.440) and F4 0.128 (0.039–0.907). Individual study characteristics are presented in Appendix 4. The prevalence of F1–F4 in NAFLD is lower than the prevalence of such stages in the other evaluated aetiologies of liver disease; this is probably due to the relatively low prevalence of the progressive steatohepatitis among patients with NAFLD. 47 Forest plots and SROC plots of different NILTs across fibrosis stages are presented in Appendices 5 and 6, respectively.

Overall, there were 24 different non-invasive tests reported for the diagnosis of ≥ F3, of which 11 (46%) were reported in single studies. 117,229,284,286,287,289,290,299,304,306,327 Of 24 different evaluated tests, 10 converged with the bivariate random-effects model [BARD (BMI, AST-ALT ratio, diabetes), AST–ALT ratio low cut-off, NAFLD fibrosis score low and high cut-offs, FIB-4 low and high cut-off, hyaluronic acid, type IV collagen, Fibrotest and Fibroscan]. The most commonly evaluated non-invasive tests were NAFLD fibrosis score (low cut-off, 10 studies), 290,300,301,309,311,315,320,322 NAFLD fibrosis score (high cut-off, nine studies),285,290,300,301,309,311,315,320,322 Fibroscan (eight studies)119,236,288,296,298,308,323,324 and BARD (seven studies). 284,291,300,301,312,315,319

For the diagnosis of cirrhosis, there were 14 different evaluated tests; however, only platelet count and Fibroscan converged with the bivariate random-effects model.

For the diagnosis of fibrosis stage ≥ F1, there were 12 tests that reported on diagnostic accuracy, and three of them converged with the bivariate random-effects model (NAFLD fibrosis score low and high cut-offs, TE).

For the diagnosis of fibrosis stage ≥ F2, there were 20 different evaluated tests, of which three converged with the bivariate random-effects model (NAFLD fibrosis score low and high cut-offs, Fibrotest).

Uninterpretable NILT results were very rare in serum markers (< 1%) and were more frequently encountered in patients who were undergoing Fibroscan examination. The rate of uninterpretable results with Fibroscan (due to < 10 valid measurements, success rate < 60% and interquartile range> 30%) was 9.6%; however, this could be underestimated due to under-reporting.

Cut-offs for specific histological stages were predetermined in 10 studies (21%). 117,119,284,291,309,311,312,315,321,322 Liver biopsy was of acceptable quality in 10 studies (21%). 117,119,229,284,286,293,308,319,321,326 We did not include data on APRI and NAFLD fibrosis scores from one study327 in the meta-analysis because of cut-offs that differed significantly from what is used in the literature. Only one study119 had low risk of bias in all of the domains of the QUADAS-2 tool; therefore, all our estimates may be biased. Studies that were judged as low risk of bias or unknown in the three most important QUADAS domains, namely patient selection, index test and reference standard, were 21% of the total number of studies (10 out of 48). 119,165,284,294,298,301,303,319,323 Quality assessment of included studies based on QUADAS-262 is shown in Table 13.

We explored potential sources of heterogeneity as outlined in the methods section. As all but one study were of low methodological quality, this potential source of heterogeneity could not be assessed. More specifically and according to source of heterogeneity, details of significant associations are:

• Histological score used (comparator is the Kleiner system): for NAFLD fibrosis score low cut-off in F2 and low and high cut-offs in F3, use of the Brunt scoring system resulted in significantly lower sensitivity and higher specificity.

• Full text versus abstract publication: for BARD score in F3, full-text publication was associated with significantly higher specificity.

• There was no heterogeneity identified in relation to transaminases levels.

Results: alcoholic liver disease

Data on patients with ALD were extracted from 12 studies. 114,273–283 Meta-analysis was performed separately for each non-invasive test assessed at each METAVIR stage (F1–F4). Summary sensitivity and specificity of non-invasive tests for each fibrosis stage are shown in Table 14. The median prevalence (minimum–maximum) of fibrosis stages F1–F4 in included studies was for F1 0.923 (single study), F2 0.633 (0.500–0.837), F3 0.509 (0.404–0.748) and F4 0.448 (0.145–0.971). Individual study characteristics, and forest plots and SROC plots of the different NILTs across fibrosis stages, are presented in Appendices 4, 5 and 6, respectively.

Overall, there were four different non-invasive tests reported for the diagnosis of ≥ F3, of which three (75%) were reported in single studies. 273,275,282 The most commonly evaluated non-invasive test was Fibroscan (four studies);273,277,278,281 however, the results did not converge with the bivariate random-effects model. There were one, five and five NILTs evaluated for F1, F2 and F4 fibrosis stages, respectively, none of which converged with the bivariate random-effects model. APRI (high and low cut-offs) in F2 were evaluated in two studies273,283 and Fibroscan in F4 was evaluated in six studies;114,273,274,276,278,281 all other tests were evaluated in single studies.

There were four different non-invasive tests reported for the diagnosis of F4, of which three (75%) were reported in single studies. 273,279,280 The most commonly evaluated non-invasive test was Fibroscan (six studies),114,273,274,276,278,281 however, with cut-offs that widely ranged from 11.4 to 25.8 kPa. The rest of the tests for F4 were PGAA [prothrombin time, gamma-glutamyl transpeptidase (GGT), apolipoprotein A1, α2-macroglobulin], Fibrotest (at a low and high cut-off) and APRI (results only reported at a high cut-off; no reason provided for not including a low cut-off).

Cut-offs for specific histological stages were predetermined in three studies (25%),276,277,282 whereas liver biopsy was of acceptable quality in one study (8%). 273 There was no study with low risk of bias in all the domains of the QUADAS-2 tool; therefore, all of our estimates may be biased. Studies that were judged as low risk of bias or unknown in the three most important QUADAS domains, namely patient selection, index test and reference standard, were 25% of the total number of studies (3 out of 12). 276,279,282 Quality assessment of included studies based on QUADAS-262 is shown in Table 15.

Although investigation of potential sources of heterogeneity was planned, it could not be performed due to the small number of studies.

Results: liver cirrhosis

Diagnostic accuracy of NILTs for cirrhosis was also analysed irrespective of aetiology of liver disease and used for the cirrhosis economic model. Summary sensitivity and specificity for cirrhosis are shown in Table 16. Median prevalence of cirrhosis in the evaluated studies was 0.18 (range 0–0.97). Fibroscan was by far the most commonly evaluated non-invasive test (65 studies). 28,29,75,76,86–88,91,95,98–100,102,105,106,110,114,116,119,130,137,141,147,153,155,159,161,164,170,172,173,194,199–201,211,223,224,225,230,241,242,246,247,250,251,263,264,272–274,276,278,281,288,296,298,308,323,324,236 We do not include a table on quality assessment of included studies to avoid repetition, as this was given separately according to disease aetiology.

Results: imaging modalities results that were used irrespective of liver disease aetiology

Data on diagnostic accuracy of imaging modalities for the non-invasive assessment of liver fibrosis were analysed irrespective of the aetiology of liver disease, with the exception of Fibroscan and ARFI. These two tests measure liver stiffness that is associated with the amount of fibrosis and there is evidence of different cut-offs according to disease aetiology. The summary sensitivity and specificity of these imaging modalities for METAVIR stages F1–F4 are shown in Table 17. 30,93,98,100,101,110,118,132,145,149,152,166,167,176,184,185,227,270,305,328–363

Certain diagnostic modalities that are not routinely used, such as MR spectroscopy,367 double-contrast material-enhanced MRI,368 maximal accumulative respiratory strain using ultrasound,234 liver enhancement ratio of gadoxetic acid enhanced MRI369 and single-photon emission CT parameters370 and real-time elastography,142 are not presented and were not used in the economic analysis, as they are not widely available and require further validation.

Data on ultrasound techniques using injection of contrast material (contrast-enhanced ultrasound)344 or measuring the splenic artery pulsatility index (ultrasound SAPI)152 were included in the analysis and in the economic modelling for HBV and HCV. These are based on ultrasound; however, they require well-trained and experienced operators and use signs that are not well validated. Moreover, all data come from specialised centres. Therefore, reported diagnostic accuracies are most probably overestimated and not reproducible in most centres. In most centres, ultrasound, CT and MRI can only diagnose cirrhosis with acceptable specificity but lower sensitivity, as early cirrhosis and lesser fibrosis stages are often missed. MR elastography is a promising tool, but is of limited use at the moment due to cost and availability.

Decision tree structure

A decision-tree model was constructed to evaluate the cost-effectiveness of the NILTs, liver biopsy, and the ‘treat all’ and ‘no treatment’ strategies. As per the schematic diagram depicting the flow of data in Chapter 3 (see Figure 1), the decision tree was populated with test sensitivity, specificity and average disease prevalence from the meta-analysis (see Chapter 4 for details), long-term costs and health outcomes from a series of Markov models, individual test costs sourced from published literature and hospital finance departments, and a measure of adverse effects associated with liver biopsy.

As discussed in Chapter 3, there are two stages to the decision tree analysis: the first stage where all tests are compared singly and the second stage where combinations of tests are compared using four different testing strategies. Schematic illustrations and descriptions of the sequential testing pathways are provided in Chapter 3. To estimate a cost and QALY for each testing strategy the long-term costs (and test costs) and QALY estimates (including disutility from liver biopsy if applicable) associated with each potential test outcome (true positive, false positive, true negative, false negative) were multiplied by the probability of each NILT returning a true positive, false positive, true negative or false negative result to give a cost and QALY estimate for each testing option.

Markov model structure

A series of Markov models were constructed to estimate the long-term costs and health outcomes associated with a correct (true positive and true negative) and incorrect (false positive and false negative) diagnosis for a hypothetical cohort of 1000 patients with HBV and suspected liver fibrosis or cirrhosis. An additional two Markov models were constructed to estimate the costs and outcomes associated with the ‘treat all’ and ‘no treatment’ approaches. Separate models were constructed for the HBeAg-positive and HBeAg-negative patient cohorts, as their natural history differs and, therefore, starting age, transition probabilities and relative risks (RRs) from treatment differed for both groups. The structural assumptions underlying the state transition models applied to both groups of patients. The models were evaluated over a lifetime period with a cycle length of 1 year. All costs were considered from the perspective of the NHS and health outcomes were measured in terms of QALYs. Costs and utilities were discounted using a rate of 3.5%. A threshold value for incremental cost-effectiveness was assumed to be £20,000–30,000 per additional QALY gained, based on UK guidelines. 66

Figure 3 shows a schematic representation of the patient pathway, which is a modified version of previously published models of liver fibrosis and cirrhosis in patients with HBV. 371,372 Fibrosis and cirrhosis health states were defined in the Markov model according to METAVIR score: mild fibrosis (F0–1), moderate fibrosis (F2–3) and compensated cirrhosis (F4). There were five other potential health states in the Markov model: decompensated cirrhosis, hepatocellular cancer (HCC), liver transplant, post liver transplant and a health state representing ‘death’, which the cohort could enter from all other health states at any time. The cohort progressed through the model as per the arrows in the illustration; the circular arrows leading back into the same health state indicate that patients could remain within that health state for longer than one cycle (cycle length set at 1 year) except for the liver transplant state, where patients could only progress to a post liver transplant health state or to death. The liver transplant state comprised two events (1 month’s duration for liver transplant and 11 months’ duration for post-liver transplantation care). Patients could progress to the HCC health state from the moderate, compensated cirrhosis and decompensated cirrhosis health states. Patients could not regress to an earlier health state. We note that although some recent studies show that fibrosis and even cirrhosis can regress with antiviral therapy,373 we did not allow for this in our model.

FIGURE 3.

Illustration of Markov model.

We assumed that a METAVIR test score of ≥ F2 equated to a positive test outcome (true positive and false positive) and treatment with antiviral agents would commence at this stage. Conversely, a METAVIR test score of < F2 indicated a negative test outcome. We incorporated treatment with peginterferon alfa-2a for 1 year for 10% of those who tested negative to reflect that a proportion of patients would receive treatment for necroinflammation. 45 The remaining 90% would undergo a policy of watchful waiting without immediate treatment.

The watchful waiting policy in the model incorporated a retest with a NILT every 2 years; the retest NILT was the same as the previous test used for the initial assessment. To allow for modelling, we assumed that the retest had perfect sensitivity and specificity and correctly diagnosed all patients. If a patient was diagnosed as positive, immediate treatment with antiviral agents would commence. We tested the robustness of this assumption in a sensitivity analysis by varying the retest sensitivity and specificity to mirror that of three commonly used NILTs: an indirect serum marker, APRI; a patented serum marker, Fibrotest; and an imaging modality, Fibroscan.

The initial starting health state for the population cohort in the models depended on the test outcome being modelled (true positive, false positive, true negative and false negative). For example, given a false positive or true negative test result, the population cohort entered the model in a F0–1 health state (no fibrosis), whereas for a true positive or false negative test result, the population cohort had an initial starting health state of F2–4, where the cohort (with fibrosis or cirrhosis) was then distributed among the health states F2–3 and F4 based on the prevalence data from the systematic review (56% and 44%, respectively).

Input parameters

The average disease prevalence used in the model (54%) was estimated as the proportion of patients with a METAVIR score ≥ F2, calculated from the meta-analysis results reported in Chapter 4. We used the sensitivity and specificity estimates for each NILT and the average prevalence estimate to calculate the probability of each test returning a true positive, false positive, true negative and false negative result. The probability of each NILT reporting a particular diagnostic test outcome is reported in Appendix 7.

Cohort data

We identified cohort characteristics (age and sex ratio) for use in the model from published NICE guidance on treatments for HBV: adefovir dipivoxil and pegylated interferon. After reviewing the sources of evidence for this guidance,374,375 we identified a Health Technology Assessment report for both treatments published in 2006 by Shepherd et al. 371 This report included an economic analysis. Shepherd et al. 371 sourced cohort data from a published study undertaken by Fattovich et al. ,376 which reported a median age and a male-to-female ratio based on 10 published studies for HBeAg-positive chronic HBV377–387 and four published studies for HBeAg-negative chronic HBV. 388–391 We employed the same assumptions as those used by Shepherd et al. :

• HBeAg-positive: starting age of 31 years and percentage of males set at 70%.

• HBeAg-negative: starting age of 40 years and percentage of males set at 90%.

Natural history: baseline transition probabilities used in model

The rate of disease progression in the models is regulated by transition probabilities. The systematic review and economic evaluation by Shepherd et al. 392 included a systematic review of epidemiological data for HBV. 371 We also reviewed a submission to NICE by Bristol Myers Squibb (BMS) for a single technology appraisal of entecavir treatment for HBV. Upon review, the Southampton Health Technology Assessment Centre (SHTAC) epidemiological search strategy was considered to be appropriate for our analysis. The papers used by the authors were reviewed and the search was rerun to carry out an updated search for natural history data. The submission report by BMS sourced data from the same literature sources as the SHTAC study.

Our updated epidemiological search strategy was undertaken to search from the year 2004 to 2012 (as SHTAC search was undertaken to 2004). The search was carried out in the MEDLINE database using the Ovid platform, and the search strategy is outlined in Appendix 2 (searched on 11 May 2012). The updated search located 597 papers, whose titles were reviewed to determine if they were relevant; 24 were retrieved for abstract review. None was relevant for our study.

As none of the papers located from the updated literature search was relevant, we reviewed the studies which informed the transition probabilities in the model constructed by Shepherd et al. 372,376,392–401

Some of the papers referenced by Shepherd et al. 371 were older published papers. 393,394 Others were published for other aetiologies of liver disease395 and other models did not provide separate transition probabilities for the mild and moderate health states. 399 The study by Wong et al. 393 was identified as relevant as the authors had reported separate transition probabilities for HBeAg-positive and HBeAg-negative. As the paper by Wong et al. 393 was published in 1995, we conducted a search for recently published studies which had cited this paper. This search located a 2010 paper by Dakin et al. ,372 which assessed the cost-effectiveness of various drug treatments for HBV.

The 2010 paper by Dakin et al. 372 sourced transition probability data from natural history studies, economic evaluations or the placebo arms of meta-analyses or RCTs. 393,395,396,398–400,402–410 The authors used a similar set of studies to those identified by Shepherd et al. 371 and many of the transition probabilities elicited were either identical or similar. The progression from compensated cirrhosis to decompensated cirrhosis was the same for both studies, as Dakin et al. 372 used a synthesis of the same papers396,398,399 used by Shepherd et al. 371

Dakin et al. 372 sourced aetiology-specific data for the probability of undergoing a liver transplant while in a decompensated cirrhosis or HCC health state, using data from the UK transplant registry, specifically for chronic HBV. This was felt to be a more relevant estimate to use in our model than the value used by Shepherd et al. ,371 who had used a value based on data from a 1997 study by Bennett et al. 395 This study had sourced a value from a paper on HCV411 and papers published in 1993412 and 1996. 413

We used the transition probabilities estimated by Dakin et al. 372 as the main source of transition probability data for the cirrhotic and post-cirrhotic health states (HCC, liver transplant and post liver transplant).

However, none of the studies reviewed elicited separate transition probability data for the precirrhotic health states (mild and moderate fibrosis). In the absence of transition probability data for these health states, we used data from a study in patients with mild chronic HCV. 414 This paper was identified in our review of data for our analysis of HCV but was also referenced as a source of cost data in the paper by Dakin et al. 372 We adopted this approach with regard to early transition probabilities and this was confirmed as appropriate with clinical colleagues.

Wright et al. 414 were unable to locate separate estimates of probabilities of progression from mild to moderate fibrosis disease stages health state from existing studies as they found that most data were based on retrospective natural history studies which do not use liver biopsy to stage fibrosis and progression. As the disease progression may not be linear, it may not be realistic to assume a constant rate of progression from mild fibrosis to cirrhosis, and so they estimated transition probabilities for mild and moderate health states using data from a trial of treatments for mild HCV undertaken in a number of London hospitals,415 during which the liver fibrosis stage was determined by liver biopsy. Transition probabilities are listed in Table 20.

Mortality data

An all-cause mortality rate was calculated using the Interim Life tables for England and Wales 2008–10. 416 The risk of death increased each year according to age and the rate was weighted to allow for sex mix. The all-cause mortality rate was added to an excess mortality value (identified from study by Dakin et al. 372) associated with the moderate fibrosis, compensated cirrhosis and HCC health states to provide a total risk of death per year. The all-cause mortality rate was not applied to the decompensated cirrhosis, liver transplant and post-liver-transplant health states; instead, a total mortality rate identified from the study by Dakin et al. 372 was applied. Mortality rates in the HBeAg-positive model ranged from 0.0007 to 0.337, and from 0.002 to 0.34 in the HBeAg-negative model. Excess mortality rates and total mortality rates employed within the models are listed in Table 20.

Antiviral treatment for hepatitis B: type and duration

The NICE website was reviewed to source national guidance on drug treatment for HBV.

We located guidelines for licensed drugs which included peginterferon alfa-2a (Pegasys^®, Roche), entecavir (Baraclude^®, BMS) and tenofovir disoproxil (Viread^®, Gilead). Entecavir and tenofovir have marketing authorisation within the UK for the treatment of chronic HBV infection in adults with compensated liver disease and evidence of active viral replication, persistently elevated serum ALT levels and histological evidence of active inflammation and/or fibrosis. Peginterferon alfa-2a has UK marketing authorisation for the treatment of HBeAg-positive and HBeAg-negative chronic HBV infection in adults with compensated liver disease and evidence of viral replication, increased ALT and histologically verified liver inflammation and/or fibrosis. Interferon antiviral agents are contraindicated in chronic hepatitis patients with decompensated cirrhosis. 375,417,418

The following treatment assumptions were employed in the model: only patients in the moderate fibrosis (F2–3), compensated cirrhosis (F4) and decompensated cirrhosis health states received treatment with antiviral agents; patients in the HCC, liver transplant and post liver transplant health states received usual standard of care.

Treatment with either entecavir or tenofovir (lifetime duration) was administered if patients tested positive (true positive or false positive); half of the eligible patients received entecavir and the other half received tenofovir.

Ten per cent of the patients who tested negative (true negative or false negative) received peginterferon alfa-2a for 1 year only, and if treatment was unsuccessful (we assumed that 30% of treated true negative patients would successfully respond to treatment and would no longer progress to any further health states except the death health state due to all-cause mortality), they would receive subsequent treatment with either entecavir or tenofovir for lifetime duration if diagnosed as positive (≥ F2) at retest. The remaining 90% of true negative and false negative patients would undergo ‘watchful waiting’, where they would then receive treatment with either entecavir or tenofovir for lifetime duration (if diagnosed as positive during a retest).

The dosage of treatment was based on national recommendations sourced from the British National Formulary (BNF) 64419 (see Table 24).

Treatment effectiveness

A meta-analysis by Woo et al. 420 which evaluated the relative efficacies of the first 12 months of treatment for chronic HBV was identified from a general search carried out using Google Scholar (http://scholar.google.com) and the search terms ‘treatment effectiveness’, and ‘tenofovir and entecavir’ (search date 12 May 2012). The paper evaluated a number of drugs – lamivudine, pegylated interferon, adefovir dipivoxil, entecavir, telbivudine and tenofovir – for use as monotherapy or in combination therapy in treatment-naive individuals.

Woo et al. 420 conducted a systematic literature review to locate studies of published RCTs of drugs used to treat chronic HBV as either monotherapies or combination therapies. They included studies that examined the impact of treatment in both HBeAg-positive and HBeAg-negative patients. The studies included in the review were required have examined the use of the drug using randomised, phase 3, controlled trials comparing new drug treatments with either a placebo or a licensed drug. The methodological quality of each paper was assessed independently by two reviewers using the Cochrane risk of bias tool.

The data were analysed using a Bayesian mixed-treatment comparison (MTC) analysis which allowed the authors to combine direct and indirect comparisons of the treatments. Our model used the treatment efficacy elicited from the Bayesian MTC for pegylated interferon, entecavir and tenofovir. The study did not provide separate efficacy data for peginterferon for HBeAg-negative chronic HBV, and we assumed the same RR as for HBeAg-positive. Treatment effectiveness represented by RRs and their associated CIs are displayed in Table 21.

We assumed that patients who tested false positive and who were in a mild state would receive the same treatment benefit from antiviral treatment as those in a moderate or cirrhotic health state. We also assumed that patients who test false negative and receive peginterferon alfa-2a for 1 year would not receive treatment benefit but would incur the costs and disutility associated with peginterferon alfa-2a treatment.

Cost data

To populate our model, we undertook a search for other relevant cost-effectiveness literature that would provide data on the costs associated with treating the different levels of fibrosis and cirrhosis in patients with HBV. To do this, we employed the search strategy devised by Shepherd et al. 371 and undertook an updated search using the MEDLINE database (Ovid platform, search date 10 May 2012, searched 2004–12). The search strategy is listed in Appendix 2. Additional search terms were added to locate papers related to other relevant treatment (entecavir, tenofovir disoproxil, and their brand names, baraclude and viread). Nine hundred and seventy-one papers were located and their titles were reviewed to determine if they were eligible. Twenty-three were retrieved for abstract review. Of these, 17 were excluded and six were retrieved for full review (Table 22). Chapter 3 contains further details of the literature review inclusion criteria used when reviewing and assessing the applicability of cost data literature.

Of the six papers reviewed, four papers included cost data from a study on mild HCV by Wright et al. 414 This study collected resource use and cost data alongside a RCT for mild HCV. 414,415 Detailed cost data were collected from three centres based in London, Newcastle and Southampton. Resource use information collected covered inpatient and outpatient care, investigations, procedures, drug use and other services including psychiatric services.

One of the studies, by Veenstra et al. ,426 sourced additional costs from a study by Bennett et al. ;395 however, this study collected costs for the health states using a US population which, due to the difference in the health-care systems, would not be transferable to a UK population.

The study by Brown et al. 421 collected data on resource use from specialists based in four countries including the UK. The authors used the data collected to identify resource use and associated costs for the management of fibrosis and cirrhosis.

The study by Takeda et al. 425 also used the 1995 Wong paper393 as a source; as this paper is an older published paper, we felt there would be more up-to-date costs available.

Three of the studies used data from the Cost-Effectiveness of Liver Transplantation (CELT) study424 to elicit a cost for the liver transplant health states. The CELT study collected costs on adult patients listed for an isolated liver transplant (aged 16 years and over) between December 1995 and December 1996. The costs were collected and split into phases, which were determined according to when the resource use took place and according to disease aetiology (HBV, HCV, alcoholic cirrhosis). Data were split into an assessment phase, candidacy phase, transplant phase and post-transplant phases. The assessment phase started at the date of admission for assessment of suitability for liver transplantation to the date of listing for transplantation (for patients who were not listed for transplantation, the discharge date was used as the end date). The candidacy phase started at the date of listing to the date the patient was admitted for the transplant operation. The transplant phase started at the date of admission for the transplant operation to the date of discharge following the operation and the post-transplant phase started at the date of discharge following the operation onwards for a period of 2 years. Resource use data on blood products used, number of dietitian sessions, drugs used, inpatient stay, nutritional support received, outpatient visits, physiotherapy sessions, tests, length of transplant operation and key treatments and investigations were collected.

Based on the review, we concluded that the data on costs for treating HCV from the study by Wright et al. 414 for the mild, moderate and cirrhotic health states would be comparable with the costs for treating patients with HBV in these health states, as the resource use identified and collected (inpatient, outpatient care, procedures) should be similar. The health state costs sourced from this study did not contain the cost of antiviral treatment; therefore, they were suitable to use and we added the cost of treatment separately in the model.

From the review, we decided to source information for the decompensated cirrhosis, HCC, liver transplant and post-liver-transplant health states from the CELT study. 424 Using the raw data collected in this study, we calculated the cost for all patients with HBV who had received a transplant (sample size 24). The average length of admission for the liver transplant operation was 28 days. We approximated this to 1 month and calculated the yearly cost of a liver transplant as 11 months of post-transplant care (estimated from the average monthly cost in the first year following transplantation) plus the month the transplant operation took place. We also calculated a cost for the post-transplant health state, estimated as the average monthly cost of the second year of post-transplant care (sample size 24).

To calculate a cost for the decompensated cirrhosis and HCC health states, we used the average cost for patients with HBV awaiting liver transplant (patients in the ‘assessment’ and ‘candidacy’ disease stages were used; sample size 25).

Costs were inflated to 2012 levels using NHS inflation indices. 67 Health state costs did not include the costs associated with antiviral treatment; these were included separately. Costs are shown in Table 23.

As we employed two different sources of data to populate our model, there is a sizable difference between the cost for the compensated and decompensated cirrhosis health states. We conducted a sensitivity analysis where we populated our model with the cost for the decompensated cirrhosis health state (£9121) and HCC health state (£8127) from the study by Wright et al. 414 The main component of the costs for these two health states in the Wright et al. 414 model was inpatient-days. We inflated the costs to 2012 prices using NHS inflation indices. 67

Test costs

Non-invasive liver tests using imaging modalities were sourced from published Department of Health reference costs. 427 Costs of direct and indirect serum markers were obtained from communication with finance departments based at the Royal Free Hospital. Costs for patented serum markers were sourced directly from manufacturers and via communication with finance departments based at the Royal Free Hospital (see Appendix 5, Table 68).

The two most commonly performed liver biopsy tests are percutaneous and transjugular liver biopsy. We assumed that, for our purposes, patients received a percutaneous liver biopsy, as this tends to be associated with less severe adverse effects. We estimated that a diagnostic liver biopsy would cost £956.61. 428 Where required, costs were inflated to 2012 prices using NHS inflation indices. 67 Tests costs and sources are listed in Appendix 9. All NILT tests costs are based on incremental costs and exclude the capital costs of the equipment.

Medication costs

Treatment costs and recommended dosages were sourced from the BNF 64419 and are listed in Table 24.

Utility data

An initial literature search for existing published utility data for HBV was undertaken using the MEDLINE database (Ovid platform, searched 11 May 2012, coverage 2004 to 2012). We supplemented this search using both the cost-effectiness analysis (CEA) Registry (searched 11 May 2012, all dates) and EuroQol website (searched 11 May 2012, all dates).

We searched the MEDLINE database using the search strategy for quality of life devised by Shepherd et al. 371 We updated this search to include papers from 2004 onwards. The full search strategy is outlined in Appendix 2. This search returned 121 papers.

We searched the EuroQol using the general search term ‘hepatitis B’. This search returned 16 papers. We also searched the CEA Registry using the search term ‘hepatitis B’, which returned 39 studies.

The title and abstract of each paper was reviewed and full papers were retrieved for review if they met our inclusion criteria. Chapter 3 outlines the inclusion criteria that applied when reviewing studies for quality of life data. After excluding three duplicate papers, eight were retrieved for full review. 372,414,425,429–432

Four of the studies retrieved372,423,425,429 used the same study by Levy et al. 430 as a source for utility values.

Levy et al. 430 employed the standard gamble technique to collect health-related utility data for six HBV-related health states (from both infected and uninfected respondents). Hypothetical health states were developed using expert opinion and the Liver Disease Quality of Life Instrument. Data were elicited from respondents from the USA, Canada, the UK, Spain, Hong Kong and China. The study analysed 1134 respondents, of whom 100 were from the uninfected population and 93 from the infected population were from the UK.

One study used clinical judgement to elicit health-related quality of life (HRQoL) values. 431 The authors noted that the weights were arbitrary and better sources could be used.

A 2008 study by McLernon et al. 432 conducted a systematic review of published literature. 414,433–436 They obtained values for the compensated cirrhosis, decompensated cirrhosis and HCC health states using these data; however, the studies had mainly been carried out in HCV and the search conducted looked for quality of life data for all aetiologies, not specifically for HBV.

The search also identified the 2006 study on treatment in mild HCV by Wright et al. 414 As none of the other studies reviewed identified separate utility values for mild and moderate health states, we adopted the same approach to sourcing data as that used for costs. We employed the same utility values used by Wright et al. 414 for the HCV mild, moderate and compensated cirrhosis health states. As both HBV and HCV lead to cirrhosis and related complications, it is assumed that cirrhosis would impact on the quality of life similarly for patients with HBV and HCV.

We decided to source utility values for the decompensated cirrhosis, decompensated cirrhosis, HCC, liver transplant and post liver transplant health states from the CELT study for HBV patients. 424 We used this source rather than one of the studies reviewed, as the only study reviewed which collected data directly used a standard gamble technique430 rather than a generic preference-based measure such as the EQ-5D. As we had data available for patients with HBV (sourced from CELT study), it was felt this would be a more accurate source of data to use rather than the data elicited from the study by Levy et al. 430 or the systematic review by McLernon et al. ,432 which were applicable to all liver disease aetiologies.

Additionally, we concluded that both the Wright et al. 414 and the CELT study424 would be suitable sources of data as they had both used the EQ-5D preference measure to elicit utility values within a UK population (see Chapter 3 for inclusion criteria for quality of life data).

During the mild HCV RCT,414 questionnaires were self-administered at baseline and treatment weeks 12, 24 and 48, and at follow-up weeks 12, 24 and 48. For the moderate and cirrhotic health states, 302 patients were sent an EQ-5D (of whom 60% of those with diagnosed as being in a moderate health state responded, and 54% of those diagnosed as having cirrhosis responded).

The CELT study424 collected HRQoL data using the EQ-5D questionnaire. We analysed the data collected for HBV patients at the time patients were placed on the waiting list for a transplant, and at 3 months, 6 months, 12 months and 24 months post transplant.

It was assumed that utility values for the decompensated cirrhosis and HCC health states would be the same (sample size 25). This was assumed to be equivalent to the average utility value at ‘listing’ stage.

A utility value for the liver transplant health state was estimated using an area under the curve approach and the average utility values collected at 3, 6 and 12 months post transplant (sample size 24). A utility value for the post liver transplant disease stage was estimated from the average utility for HBV patients collected at 24 months post transplant (sample size 24). A PSA was carried out using a utility decrement approach. A beta distribution is often employed for utility values; however, this may not be appropriate for states close to death, where values of less than one are possible. For the HBV model, we performed a simple transformation of the data (D = 1 – U, where D is the utility decrement and U is the utility value). The decrement was constrained on the interval ‘0 to positive infinity’ and a gamma distribution was then applied. 437 Table 25 lists the utility data used in the model.

As we sourced utility value data from two different studies, the utility values used in the study for the decompensated cirrhosis and HCC health states were slightly higher than those used for the compensated cirrhosis health state. We undertook a sensitivity analysis where we set the utility values for the decompensated cirrhosis and HCC health states to those used in the HCV model (see Chapter 6) – as these were lower values than those used in the HBV model – to test if this had any effect on the robustness of the results.

Adverse effects associated with antiviral treatment

As peginterferon alfa-2a has associated side effects (such as influenza-like symptoms, depression and anxiety), we allowed for the disutility associated with antiviral treatment to be reflected in the model. We modelled a disutility decrement that was applicable during treatment, using data identified from the Wright et al. 414 study. This study reported HRQoL data using the EQ-5D instrument for 144 patients. They used the data from weeks 12 and 24 from the baseline date to estimate utility decrement values of 0.11. The decrement was applied to all patients in the mild and moderate fibrosis and compensated cirrhosis health states as a multiplicative disutility for the duration of treatment with peginterferon alfa-2a. It was conservatively assumed that any side effects associated with treatment with entecavir or tenofovir would have no impact on quality of life.

Disutility from non-invasive liver treatment and liver biopsy

We have assumed that any adverse events associated with the NILTs would not have a significant impact on health-related utility. However, as liver biopsy is associated with morbidity and mortality risks and patient discomfort, we incorporated a utility decrement to represent expected adverse events. In the absence of identified data, this was modelled by applying a utility decrement of 0.2 based on data employed in a previous study,428 where the authors had conducted a literature review for data on liver biopsy associated adverse events and mortality. We undertook a number of sensitivity analyses to test the robustness of the results to changes in the utility decrement.

Estimated long-term costs and quality-adjusted life-years

The estimated long-term costs and QALYs for HBeAg-positive and HBeAg-negative chronic HBV (output from Markov models) are shown in Table 26.

Treatment benefit

Our base-case analysis assumed that patients who are in a mild health state (F0–1) receive the same treatment benefit (same effectiveness) as patients in a moderate or cirrhotic health state, despite being incorrectly diagnosed and treated. 373 We tested the robustness of this assumption in a sensitivity analysis by setting the treatment benefit for patients who incorrectly receive treatment while in a mild health state to zero.

Robust test accuracy data

In the base-case analysis, all tests were included despite there being limited data available for some tests. A sensitivity analysis including only tests where the standard bivariate random-effects model used for the meta-analysis63 converged was conducted. We conducted an analysis with the five tests where the bivariate model had converged – APRI high cut-off, Fibroscan, Fibrotest, FIB-4 low cut-off and APRI low cut-off – and compared with liver biopsy, ‘treat all’ and ‘treat no one’ strategies.

Utility value amendment

We carried out an analysis where we amended the utility values within the model for the decompensated cirrhosis, HCC, liver transplant and post liver transplant. The sensitivity analysis used the lower values from the HCV model (see Chapter 6 for values).

Change in liver biopsy decrement

We also carried out analyses using different utility decrement values to measure the adverse effects from liver biopsy. The base-case analysis set the utility decrement value to 0.2 based on a previous analysis which had conducted a literature search for data relating to adverse events and mortality associated with liver biopsy. 428 In the sensitivity analysis we set the utility decrement at 0 and 0.3 to test the impact on the robustness of the results.

Change to health state costs

We carried out an analysis where we changed the costs for the decompensated cirrhosis and HCC health states to the costs used in the Wright et al. 414 study.

Average disease prevalence

Prevalence within the model is based on studies that may have been carried out largely in tertiary care centres, and the prevalence of liver fibrosis in this population may be an overestimate. To test the impact of this we undertook a sensitivity analyses using the minimum prevalence (disease prevalence modelled as ≥ F2 at 27%), maximum prevalence (92%) and the 25th and 75th quartile values (43% and 65%, respectively), estimated from the meta-analysis of the systematic review data. In this sensitivity analysis, it is assumed that the sensitivity and specificity of the tests would be the same in populations with high and low prevalence as observed in the studies included in the review.

Sensitivity and specificity of retest

Our base-case analysis assumes that the retest (from the meta-analysis of the systematic review data in the watchful waiting strategy for patients with a negative test result) has perfect sensitivity and specificity. This is likely to overestimate the accuracy of the retest procedure. We tested the impact of this assumption by applying the sensitivity and specificity of three commonly used tests: APRI low cut-off (estimated sensitivity of 80% and specificity of 65%), Fibrotest (estimated sensitivity of 66% and specificity of 80%) and Fibroscan (estimated sensitivity of 71% and specificity of 84%).

Choice of tests for second stage of analysis

For the second stage of the analysis, the two most cost-effective tests when assessed within each specific test category singly (with and without a defined threshold) were used in the analysis of sequential testing. To test if changing the method used to choose tests for the second stage of the analysis had an effect on the overall result, we carried out an analysis where the most effective NILT from within each NILT category and the least costly NILT from each category were included in the analysis of sequential testing.

Change to hepatitis B e antigen-negative model (age and sex ratio)

We also undertook a sensitivity analysis, where we set the age-and-sex-mix data in the HBeAg-negative model equivalent to the age and sex mix in the HBeAg-positive model (base age set to 31 years and the proportion of cohort who were male set at 70%) to check if the results from both models would be similar. The only remaining differences between the HBeAg-positive and HBeAg-negative models related to the different efficacy of treatment and the increased probability of moving to a cirrhotic state from a moderate health state.

Reduction in length of time of successful response rate to peginterferon alfa-2a treatment

Our base-case analysis for true-negative patients assumes that 30% of those who receive treatment with peginterferon alfa-2a have a successful response and no longer progress to further health states in the model retaining a risk of all-cause mortality only. We tested this assumption in the model by assuming that the successful response lasts for 15 years only.

Change to non-invasive liver test costs

We carried out a sensitivity analysis where we changed the cost of the NILTs [we set the watchful waiting retest cost and all NILT costs within the model to the same cost; for comparison, we assumed an indirect serum marker test cost (we chose a commonly used test, APRI, as our comparator)]. By changing the cost of a NILT, we aimed to determine if changing the test cost (marginal cost) had an impact on the robustness of the results.

Hepatitis B e antigen-positive chronic hepatitis B

At a cost-effectiveness threshold of £20,000, the cost-effective strategy is to employ a sequential testing strategy where a direct serum marker, hyaluronic acid, is combined with an imaging modality, MR elastography, with an ICER of £19,612. With this testing strategy, positive test diagnoses are confirmed with a second NILT and if the results disagree, a liver biopsy is administered to confirm the result; negative test responses undergo a process of watchful waiting.

Using a higher cost-effectiveness threshold value of £30,000, an imaging modality, MR elastography, used singly, becomes the most cost-effective option, with an ICER of £28,585. When we assessed all tests singly in the first stage of the analysis, the most cost-effective test given a threshold value of £20,000 was GUCI, with an ICER of £19,716.

As the CEAF is more representative of the uncertainty around the optimal testing option, we present the CEAF (Figure 4) in the main analysis and the CEAC in Appendix 6.

FIGURE 4.

Cost-effectiveness acceptability frontier for HBeAg-positive second stage of the analysis. HA, hyaluronic acid; MRE, MR elastography.

The CEAF shows that the probability of hyaluronic acid combined with MR elastography (strategy 2) being the optimal testing option, given a cost-effectiveness threshold value of £20,000, is 4%. The CEAF also shows that MR elastography, used singly, has a 13% probability of being cost-effective for cost-effectiveness thresholds starting at £28,585, which reduces to a 3.5% probability when the cost-effectiveness threshold increases to £43,946. For cost-effectiveness threshold values greater than £43,946, the ‘treat all’ strategy has a 35% probability of being cost-effective.

What is noticeable about the CEAF is that two of the strategies (testing with MR elastography alone or a combination of hyaluronic acid and MR elastography) have a very low probability of being the optimal testing option, compared with a testing strategy of treating everyone.

For reasons of clarity, the CEAC presented in Appendix 10 displays only those testing options which had a ≥ 4% probability of being cost-effective. The CEAC also displays some tests that are not picked up by the CEAF, including the Forns (high cut-off) serum marker. The reasons for this include a skew in cost data and also the fact that those tests which have the highest probability of being cost-effective may not be the optimal testing option.

Liver biopsy was a comparator for both stages of the analysis; however, this testing option is dominated by other less costly but more effective options. Table 27 presents incremental results for the first stage of the analysis and Table 28 incremental results for the second stage where a number of combined tests are compared using a number of different sequential testing strategies (see Chapter 3 for details of testing strategies).