The clinical effectiveness and cost-effectiveness of STeroids or Pentoxifylline for Alcoholic Hepatitis (STOPAH): a 2x2 factorial, randomised controlled trial

Corticosteroids

Since 1971 there have been 13 randomised studies and four meta-analyses investigating the role of corticosteroid therapy for this condition. 11,12 Despite this apparent wealth of evidence, controversy persists. There remains deep division with regard to the use of corticosteroids. Advocates of the treatment cite significant improvement in the short- to medium-term mortality, while detractors cite the risks of sepsis and GI haemorrhage with corticosteroid therapy. Many of the published studies have been plagued by widely varying inclusion and exclusion criteria. The largest placebo-controlled study13 treated 90 patients and found no benefit with prednisolone compared with a similar placebo-treated group. This study was hampered by the inclusion of patients with both moderate and severe AH, as well as end-stage ALD. In the only study to require histological confirmation of AH in all patients, prednisolone was associated with a short-term improvement in mortality in patients, although this benefit was not apparent after 2 years. 14,15 On review of the published studies, none of these reached an adequate statistical power to make a statement with 80% confidence. The most recent meta-analysis of all of the available trials16 demonstrated that, although there was a trend of benefit with steroids, the results were not statistically significant (p = 0.2). However, a reanalysis of the three largest studies indicted a significant benefit from corticosteroids. 17 In this meta-analysis, patients with a DF ≥ 32 treated with prednisolone had 28-day mortality of 15%, compared with mortality of 35% among placebo-treated patients (p = 0.001).

Pentoxifylline

Pentoxifylline has also been studied in the treatment of AH. 18–20 It is believed to act, in part, by inhibiting the synthesis of the proinflammatory cytokine tumour necrosis factor alpha. 20 There has been one randomised controlled trial18 that showed significant benefit. In this study, 100 patients, all with a DF > 32, were enrolled. PTX was administered for 4 weeks, at a dose of 400 mg three times per day. In the PTX group, 12 of 49 patients (24.5%) died, compared with 24 of 52 (46.1%) in the placebo group during the index hospitalisation (p = 0.037). The principal benefit for the agent appeared to be a reduction in deaths attributed to hepatorenal syndrome. However, other studies found contrasting results and meta-analysis has failed to show any significant benefit of this drug. 19–21 Although published evidence for the use of PTX is inconclusive, the drug is widely used as an alternative to steroids, particularly in the USA, and further evaluation in a clinical trial is, therefore, clearly warranted. 22

Steroids and/or pentoxifylline

Combinations of steroids and PTX have been evaluated in clinical trials, using steroids alone as the control arm; however, the combination does not appear to have any benefit over steroids alone. 23,24 Three small studies25–27 have compared steroids directly with PTX but the results were inconsistent. One study28 explored the use of PTX as a rescue therapy once steroids had failed to improve liver function tests (LFTs) but this strategy did not influence survival.

Changes to original protocol

A summary of the changes made to the original protocol is given in Table 1.

Inclusion criteria

• Aged ≥ 18 years.

• Clinical AH:

  • serum bilirubin level of > 80 µmol/l

  • history of excess alcohol (> 80 g/day for males and > 60 g/day for females) to within 2 months of randomisation.

• Less than 4 weeks since admission to hospital.

• DF of ≥ 32.

• Informed consent.

Exclusion criteria

• Abstinence of > 2 months prior to randomisation.

• Duration of clinically apparent jaundice > 3 months.

• Other causes of liver disease including:

  • evidence of chronic viral hepatitis (hepatitis B or C)

  • biliary obstruction

  • hepatocellular carcinoma.

• Evidence of current malignancy (except non-melanotic skin cancer).

• Previous entry into the study, or use of either prednisolone or PTX within 6 weeks of admission.

• aspartate aminotransferase (AST) level of > 500 IU or alanine aminotransferase (ALT) level of > 300 IU (not compatible with AH).

• Patients with a serum creatinine level of > 500 µmol/l or requiring renal support.

• Patients dependent on inotropic support (adrenaline or noradrenaline). Terlipressin is allowed.

• Active GI bleeding.

• Untreated sepsis.

• Patients with known hypersensitivity to PTX, other methylxanthines or any of the excipients.

• Patients with cerebral haemorrhage, extensive retinal haemorrhage, acute myocardial infarction (within the last 6 weeks) or severe cardiac arrhythmias (not including atrial fibrillation).

• Note that patients with evidence of sepsis, GI bleeding or renal failure may be treated for these conditions for up to 7 days and, if stable, the patients can then be rescreened for eligibility. Treatment can continue for > 7 days if they are stable. Patients who are not stable after 7 days of treatment will not be eligible for the study.

Recruitment

Site staff assessed patients admitted into secondary care with an acute episode of AH for potential eligibility for the trial.

Informed consent

Potentially eligible patients (or their legal representatives) were informed about the trial by one of the study team, and then given a patient information sheet to review. Patients were given a minimum of 24 hours in which to consider the study and ask questions, after which they (or their legal representatives) were asked to give written informed consent to participate, on the trial informed consent form.

For relevant patients, consent given by a legal representative was in place until the patient recovered capacity, at which point the patient was informed about the trial and asked to decide whether or not they wanted to continue in the trial. Consent to continue was then sought from the patients themselves.

Randomisation and concealment

Site staff registered patients onto the trial via Trans European Network for clinical trials services (TENALEA), a web-based registration and randomisation system, after written informed consent was obtained from the patient (or their legal representative). Subsequently, screening assessments took place to ascertain eligibility. Non-eligible patients were deemed screening failures, while eligible patients were then randomised by site staff, via the TENALEA, to one of four trial treatment arms. Treatment allocation was blinded to site staff and the patient by providing each patient with a unique four-digit patient pack number. The treatment arm was also concealed to investigators and researchers. Only the study statisticians were unblinded and this was for analysis purposes only.

Randomisation was stratified and performed using the following two stratification factors:

1. geographic region (28 in total)

2. risk group – either ‘high’ or ‘intermediate’ risk (‘high’ risk was defined as either sepsis or history of GI bleeding in the previous 7 days or creatinine level of > 150 µmol/l, or any combination of the these; ‘intermediate’ risk was defined as no sepsis and no history of GI bleeding in the previous 7 days and creatinine ≤ 150 µmol/l).

Unblinding service

The Emergency Scientific and Medical Services at Guy’s and St Thomas’ NHS Foundation Trust provided a 24-hour unblinding service for the trial. The Emergency Scientific and Medical Services held a copy of the study randomisation list and were able to perform emergency unblinding if the enquirer considered the medical management of the patient to be dependent on knowing their treatment allocation.

Data collection and management

Sites entered trial-specific data, as specified in the protocol, onto paper case report forms (CRFs). Completed paper CRFs were sent to the Southampton Clinical Trials Unit (SCTU) that was responsible for the data management of the trial. Data were transcribed from paper CRFs into an InForm database (InForm version 5.0, ORACLE, Redwood Shoves, CA, USA) at the SCTU. A range of data validation checks was carried out within both InForm and SAS version 9.3 (SAS Institute Inc., Cary, NC, USA) to minimise incorrect or missing data.

Source data verification was undertaken during site monitoring visits, in accordance with the SCTU’s trial monitoring plan. For sites that recruited ≥ 15 patients a planned monitoring visit was scheduled. In total, 29 planned monitoring visits and two triggered monitoring visits took place.

Baseline

At baseline, a number of characteristics were collected to determine the patient’s prognosis at the beginning of the trial. The outcomes collected at this time point included the following: vital signs, World Health Organization performance status, concomitant medication, LFTs, prothrombin time (PT), full blood count (FBC), urea, creatinine, GI bleed or sepsis since screening and time since admission to hospital. Prognostic scores at baseline (DF, MELD score, GAHS and Lille score) were derived using baseline data with the exception of the Lille score, which also uses data from day 7. Demographic data were captured at screening.

Follow-up

Patients who were being treated as in-patients were followed-up every 7 days from the start of treatment until the end of treatment (day 28). If the patient was discharged from hospital during the treatment phase, then the 28-day follow-up visit was conducted by phone, when possible, in order to gather treatment information. All patients were scheduled for a follow-up hospital visit at 90 days and 1 year from start of treatment, at which time details of the patients’ alcohol consumption and attendance at alcohol counselling were collected. There were also repeated measurements of outcomes recorded at baseline (see Baseline) as well as if patients had experienced an occurrence of GI bleed or sepsis since their last visit. Finally, for the 1-year follow-up visit only, demographic data were collected on the patients’ marital status, employment status and housing status.

All patients were asked to consent to information about their health status being held and maintained by the Health and Social Care Information Centre and the NHS Central Register to enable long-term follow-up. In January 2013, the National Institute for Health Research Health Technology Assessment (funding body) awarded a no-cost extension to recruitment (from the end of December 2012 to the end of February 2014), but with follow-up reduced to a minimum of 28 days (to the end of March 2014), so that primary outcome data could be collected for all randomised patients.

Primary outcome

The primary outcome was 28-day mortality. This was chosen as this time point represents the end of the peak period of mortality for AH. Also, this is consistent with other AH trials, which allows a comparison to be made. Patient mortality at day 28 was treated as a binary outcome (1 = dead, 0 = alive).

Secondary outcomes

Analysis populations

All summaries and analyses were carried out on the intention-to-treat (ITT) population unless otherwise specified. The ITT population included all patients who were randomised regardless of treatment adherence. For analyses at a certain time point, the ITT population was modified to include only patients who were known to be alive at the specified time point or had died prior to the time point, that is, for primary analysis, this included all randomised patients for whom we knew that they were alive at day 28 or who died prior to or on day 28.

A per-protocol population was defined as:

• treatment adherence ≥ 75%

  • patients who reached day 28 needed to have had taken at least 75% of treatment (> 21 days) to be included in the per-protocol population

  • patients who withdrew or died before day 28 needed to have had a minimum of 7 days of treatment and taken at least 75% of treatment up until they withdrew/died

• patients who had no major protocol violations.

The primary end-point analysis was repeated for the per-protocol population.

Preliminary analyses

Summary statistics were produced for a selection of key clinical and demographic characteristics at baseline to assess baseline comparability between the four treatment arms. Any significant differences were reported.

Primary analyses

The primary outcome of 28-day mortality was analysed using logistic regression. Although the trial was not powered to detect a treatment interaction, this was tested using a secondary logistic regression model in order to test the underlying model assumptions surrounding factorial design.

Secondary analyses

The primary outcome analysis method was repeated for the 90-day and 1-year mortality or liver transplant outcomes. For OS, comparisons were made using Cox proportional hazards regression and Kaplan–Meier curves.

A number of analyses were carried out on the DF, GAHS, MELD and Lille prognostic scores. First, to assess the predictive power of the scores on 28-day, 90-day and 1-year mortality, a receiver operating characteristic (ROC) curve analysis was carried out. Graphical representation of the ROC curves was produced along with area under the curve statistics. Further ROC curve analysis was carried out on the Lille score comparing the two prednisolone-treated groups (not predefined). Univariate logistic regression analysis was also carried out to look at the relationship between the scores and outcome.

Univariate logistic regression analysis was carried out to assess the relationship between baseline prognostic factors and 28-day, 90-day and 1-year mortality. The prognostic factors included in this analysis were as follows: pyrexia, hypotension, tachycardia, alcohol intake, albumin, ALP, bilirubin, encephalopathy, PT ratio/INR, age, WBCs, urea and creatinine. Odds ratios with a 95% confidence interval (CI) and corresponding p-value were produced for each model fitted. Any factors that were significant in the univariate analysis were added to a multivariate model, which also included the treatment indicators for prednisolone and PTX. Backward elimination was then applied to produce adjusted odds ratios for the effects of prednisolone and PTX. Please note that the use of backward elimination was not predefined in the statistical analysis plan.

Descriptive summaries of the following were produced: causes of death, SAEs, alcohol behaviour and alcohol counselling attendance.

Finally, the effect of drinking habits on 1-year mortality was explored by fitting a logistic regression model with abstinence as the reference event. This was compared with the other three alcohol consumption statuses (reduced drinking to below government defined safety limits, reduced drinking but still above safety limits and not reduced). Please note that this analysis was not predefined.

For all regression models applied, the assumptions underlying these models were checked using appropriate methods; the Hosmer–Lemeshow goodness-of-fit test for assessing logistic regression modelling and the proportional hazards assumption for Cox proportional hazards modelling. Covariates used in logistic regression modelling were assessed for outliers. Subsequently, outliers were excluded during a sensitivity analysis to allow for any differences in results to be examined.

Sensitivity analyses

A sensitivity analysis of the primary end-point outcome was produced by excluding any patients who were randomised and followed-up but were later found to be ineligible through central monitoring of baseline characteristics. Patients were excluded if they had any of the following characteristics at baseline:

• < 18 years old

• bilirubin level of < 80 µmol/l

• maximum alcohol consumption in past 2 months < 70 units/week for a male

• maximum alcohol consumption in past 2 months < 52.5 units/week for a female

• time between initial admission and screening date > 4 weeks

• DF of < 32

• AST level of > 500 IU/l

• ALT level of > 300 U/l

• creatinine level of > 500 µmol/l

• unresolved sepsis

• unresolved GI bleed.

Please note that this sensitivity analysis was not predefined.

Subgroup analyses

Subgroup analyses were performed to further explore whether or not particular subsets of the patient population were at more benefit to treatment and prolongation of survival. The subgroups considered were the prognostic scores (DF, GAHS, MELD and Lille score) and risk at baseline. These analyses looked at 28-day, 90-day and 1-year mortality using logistic regression analysis methods as described in Statistical analysis. Please note that these subgroup analyses were not predefined.

Flow of participants through the trial

See Figure 2 for CONSORT flow diagram.

Unblinding of patients

Requests for unblinding were received for a small number of patients. In all cases there were reasonable clinical justifications for the unblinding (detailed in Table 3), which were judged to be acceptable by the trial chief investigator (ChI). In view of the hard trial end points and the small number of unblinding events these are not thought to have any impact on the trial result.

Day 90, year 1 and overall survival

Table 8 documents the secondary outcomes at day 90. At day 90, there had been 139 deaths out of 478 participants (29.1%) deaths in the PTX group compared with 146 deaths out of 490 participants (29.8%) deaths in the group not treated with PTX (odds ratio 0.97, 95% CI 0.73 to 1.28; p = 0.807). There had been 144 of 484 (29.8%) deaths in the prednisolone group compared with 141 of 484 (29.1%) deaths in the group not treated with prednisolone (odds ratio 1.02, 95% CI 0.77 to 1.35; p = 0.875) (Table 9).

At 1 year there had been 205 of 365 (56.2%) deaths in the PTX group, compared with 216 of 382 (56.5%) deaths in the group not treated with PTX (odds ratio 0.99, 95% CI 0.74 to 1.33; p = 0.972). There had been 210 of 371 (29.8%) deaths in the prednisolone group, compared with 211 of 376 (56.1%) deaths in the group not treated with prednisolone (odds ratio 1.01, 95% CI 0.76 to 1.35; p = 0.937) (Tables 10 and 11).

Neither prednisolone nor PTX appear to influence all-cause survival after 28 days (Figures 3–5). Analysis of the causes of mortality as recorded by the investigators indicate that most patients died from liver-related causes, and in many cases this included a contribution from sepsis. The 28-day mortality in this trial was appreciably less than that seen in other studies in this patient group. However, the overall mortality at 1 year is 56%, illustrating the extent to which ALD contributes to the alarming rise in liver disease mortality in the UK.

FIGURE 3.

Prednisolone vs. no prednisolone.

FIGURE 4.

Pentoxifylline vs. no PTX.

FIGURE 5.

Individual treatment arms.

Prognostic scores

Four scoring systems have been described for use in providing prognostic information in patients with AH. DF, GAHS and MELD are derived from clinical and laboratory parameters at baseline whereas the Lille score also incorporated the response to 1 week of prednisolone therapy, determined by the change in bilirubin level at 7 days. Although each of the scores were highly significantly associated with mortality (Table 12) at each time point, the diagnostic performance was not considered to be adequate at a threshold for the area under the ROC curve of 0.75 (Table 13 and Figure 6). Inevitably the prognostic value of the scores diminished with increasing duration of follow-up.

FIGURE 6.

Receiver operating characteristics curves for prognostic scores. (a) 28-day mortality; (b) 90-day mortality; and (c) 1-year mortality. AUC, area under the curve.

The impact of prognostic scores on the response to prednisolone was explored by stratifying patients by the magnitude of the prognostic score value or by their risk category (Table 14). These data suggest that the risk group (high vs. intermediate) has no impact on the effect of prednisolone on mortality. Based on the odds ratios, patients with a low DF and low GAHS appear to derive more benefit from prednisolone. However, in contrast, patients may benefit from prednisolone when they have a higher MELD score. These results appear contradictory and will need to be explored in further studies.

The Lille score was originally derived to predict mortality based on the response to prednisolone. When the analysis of Lille performance was restricted to only those patients who received prednisolone, the diagnostic performance improved (Table 15 and Figure 7). However, in the original description of the Lille score, a score ≥ 0.45 identified a population of patients who had an expected mortality of 75% by 6 months, whereas in our study a Lille score ≥ 0.45 was associated with around 50% survival at 6 months (Table 16 and Figure 8). At this survival rate the Lille score cannot be used to identify a population who may be suitable for liver transplantation as envisaged in the French pilot scheme.

FIGURE 7.

Receiver operating characteristics curves to compare the Lille score (prednisolone-treated group only). (a) 28-day mortality; (b) 90-day mortality; and (c) 1-year mortality. AUC, area under the curve.

FIGURE 8.

Kaplan–Meier graph for over survival by Lille score (< 0.45 and > 0.45) (prednisolone-treated group only).

Other predictive factors of mortality

Factors that influenced mortality at 28 days, 90 days and 1 year were analysed in univariate and multivariate analyses using baseline clinical and laboratory characteristics. Table 17 summarises the findings from the univariate analysis. As expected, the variables significantly associated with mortality were encephalopathy, bilirubin, PT ratio, age, WBC count, creatinine and urea.

Multivariate analysis was conducted by first including variables from the univariate analysis that were significant at the 5% level (Table 18). Variable selection was performed using backward elimination at the 5% significance level. The logistic regression analysis now demonstrated a clearly significant influence of prednisolone on 28-day mortality with an odds ratio of 0.609 (95% CI 0.409 to 0.909; p = 0.015). The odds ratio for 28-day mortality is adjusted for the following baseline factors: PT ratio or INR, bilirubin, age, WBC count, urea, creatinine and encephalopathy.

Further multivariate analysis was performed after the removal of patients who had outlying observations. A patient was classed as an outlier if they had a standardised residual with an absolute value of > 2 for the fitted variables. The results of this analysis (Table 19) show an important increase in the treatment effect of prednisolone (odds ratio 0.427, 95% CI 0.253 to 0.721); however, the treatment effect is confined to 28-day mortality with no impact on later time points.

Serious adverse events and deaths

As expected in this patient population, there were numerous adverse events and fatalities. As shown in Table 20 the vast majority of deaths were liver related and include all the common complications of acute-on-chronic liver failure.

Serious adverse events are summarised in Table 21. Over 75% of SAEs were at Common Toxicity Criteria For Adverse Events grade 3 or above. SAEs resulted in mortality in 31% of cases. Details of the SAEs are given in Table 22 and the incidence of infection is shown in Table 23.

Infections were frequent causes of SAEs and were reported more frequently in the prednisolone-treated group (see Table 21). Although infections led to mortality in approximately 35% of cases, the risk of mortality owing to an infection-related SAE was no greater in the prednisolone-treated group than in the controls, suggesting that corticosteroids increase the risk of infection but do not necessarily make the consequences of infection any worse (see Table 22). The most common site of infection was the lung and it is interesting to note that prednisolone had a more significant impact on pulmonary infections than any other site.

Alcohol behaviour

It is well recognised that return to heavy alcohol use is one of the most important determinants of medium- to long-term prognosis in patients admitted with AH or decompensated alcoholic cirrhosis. 32 We therefore sought to document the rates of recidivism in the STOPAH cohort. Accurate measures of alcohol consumption are difficult to ascertain, so we asked patients to report on if their alcohol consumption was:

1. completely abstinent

2. reduced to within government-defined safety limits

3. reduced but above safety limits (14 units/week for females, 21 units/week for males)

4. not reduced.

The alcohol use questions were asked at day 90 and 1-year follow-up.

Overall 45% of patients reported abstinence at 90 days and 36% reported abstinence at 1 year (Table 24). It is disappointing to record that only 22% of patients had attended one or more alcohol counselling sessions by the 90-day follow-up, which appears to have influenced alcohol recidivist behaviour. Inevitably there is a high proportion of missing data from these CRFs, reflecting a loss to follow-up, which in many cases is probably associated with recidivism.

The impact of abstinence (compared with other drinking patterns) at 90 days or mortality at 1 year was explored (see Table 25). These data indicate that even modest drinking, within government guidelines, is associated with increased mortality (odds ratio 2.17, 95% CI 1.07 to 4.39; p = 0.03) whereas resumption of previous levels of alcohol use results in approximately a threefold increased mortality compared with abstinence. These findings reinforce the need to promote and support abstinence in this patient group.

Methods

Quality of life was measured in a sample of patients with liver cirrhosis caused by alcohol abuse or viral hepatitis, using the SG technique. Utility values were obtained for patients in three health states that correspond to different stages of cirrhosis. All utility values are measured on a scale from 0 to 1, for which 0 represents death and 1 represents full health. Each participant was also asked to value their own health that day using both the SG and EQ-5D-3L. Standard clinical and laboratory variables were also collected from patient medical records and were used to calculate the Child–Pugh and MELD scores for each participant.

Standard gamble methodology

The SG is the gold standard technique for eliciting preferences for both chronic and temporary health states. 38 Figure 9 provides an example of a SG when respondents are presented with two alternatives.

FIGURE 9.

Example of a SG. p, probability.

In Figure 9, alternative A is a health state with certainty, and alternative B is a health state with uncertainty, that is, there is a gamble associated with this choice. Alternative B comprises two possible health states: the patient returns immediately to full health or dies immediately. Each of these states has probabilities attached to them. The respondent makes the choice between remaining in the chronic health state or taking the gamble with a probability that they will return to full health (p) and a probability that they would die (1 – p). The probabilities within alternative B continue to change in an iterative fashion until the participant becomes indifferent to alternatives A and B.

Results

A sample of 142 people with liver cirrhosis was recruited, 91 with alcoholic liver disease and 51 with viral hepatitis (see Table 27). The majority of participants had the mildest form of liver cirrhosis (58%). Also in Table 27 are the demographic and socioeconomic characteristics of the sample. The majority of respondents were male (75%), 39% were married, 44% were unable to work and 82% were white British. Thirty-one per cent of the sample stated that they had no formal qualifications. The average age was 57 years.

The mean values for each of the health states can be seen in Table 28. When health state X represents Child–Pugh A, health state Y represents Child–Pugh B and health state Z represents Child–Pugh C. Mean values follow the expected patterns, for which the most severe health states have the lowest health-state values. Paired sample t-tests were used to assess if valuations for state Child–Pugh A (X) were significantly different from Child–Pugh B (Y) and valuations for Child–Pugh B (Y) significantly different from Child–Pugh C (Z). It was expected that these differences would be statistically significant, reflecting the differences in severity between states and the different severities of the disease. These tests were performed for the full sample and for both subsamples. All were found to be statistically significant at the 5% level. Further t-tests were conducted to test whether or not there was concordance between own health, as valued by the SG, and own health as valued by the EQ-5D. As expected, no significant differences were found.

Discussion

This study measured QoL directly using the SG and indirectly via the EQ-5D-3L. The aim was to explore health-state values for various stages of disease, represented by different Child–Pugh classifications (A, B and C). However, the collection of SG data from participants with Child–Pugh C was complicated by the presence of encephalopathy in many of these participants, which interfered with the ability of participants to give informed consent and also to understand the concept. This problem accounts for the relatively small sample size in this category.

At the aggregate level, results of health-state valuations were in line with a priori expectations, when Child–Pugh A was preferred to Child–Pugh B and Child–Pugh B preferred to Child–Pugh C. As expected, at the aggregate level, statistically significant differences between the values of health states Child–Pugh A, B and C were found, adding credibility to the results. As disease severity increases, health-state valuations significantly decrease. Convergent validity was tested by examining valuations of an individual’s own health today, as measured by the SG and EQ-5D. These values were not significantly different, suggesting high levels of convergent validity between the two measures values. These results hold for the full sample analysis and subsample analyses.

At the individual level there were some inconsistencies in both the ranking of health states and the valuation exercise. Such results are not uncommon at the individual level. 45 Importantly, valuations at the aggregate level are consistent with a priori expectations.

Conclusion

As seen in the systematic review40 and a search of literature since 2008 there is little empirical research into health-state valuations in liver disease generally, and in alcoholic liver disease specifically, that can be used to derive QALYs for use in economic evaluations and decision-making. This study has measured QoL using the SG method to obtain utility values for three different classifications of disease: Child–Pugh A, B and C, and in two separate populations: those suffering from cirrhosis owing to alcoholic liver disease and those with cirrhosis owing to viral hepatitis. This research offers a unique opportunity to estimate QALYs on future trials for liver cirrhosis when QoL can be mapped to Child–Pugh classification of the trial participant.

Methods overview

This economic evaluation includes a within-trial analysis and a model-based analysis.

First, a cost-effectiveness analysis was conducted that allowed us to calculate the incremental cost per additional survivor at 28 days. Second, we planned to repeat the analysis for a 1-year time horizon; however, as estimates of the incremental cost per additional survivor may be difficult to interpret according to standard decision-making criteria within the NHS,46 a cost–utility analysis was also performed.

When we came to analyse the data there were few responses (n = 192) to the use of medical services questionnaire administered at 1 year across the four treatment arms. If we used these responses, even with imputations for the missing responses, there would be significant concerns that there would have selection biases given the nature and extent of the missing data. However, to extrapolate to 1 year and longer follow-up we used a Markov model. In a Markov model people move between discrete states of health over time. Within a Markov model, movement between states can take place only once every period of time, called a cycle length. In this model we have adopted a cycle length of 1 day, as this will allow us to cross-validate estimates with the within-trial analysis data and be sufficiently sensitive to changes over time. In this model the costs and utility values were collected over the treatment phase and 90 days to estimate the total average cost per day and average QoL utility per day for each treatment arm.

Within-trial analysis

Our main analysis was the relative cost-effectiveness of the study treatments at 28 days.

Sensitivity analysis

For the within-trial analysis, stochastic sensitivity analyses were performed. A stochastic sensitivity analysis was undertaken to allow presentation of the level of variance around outcome measures included in the cost–utility analysis. Uncertainty surrounding the cost-effectiveness ratio was addressed using the bootstrapping technique. The results of the bootstrapping simulation were presented on the ‘cost-effectiveness plane’, which highlights the preferred treatment option. If the results lie in the north-west or south-east quadrants the preferred treatment is clear, as one option dominates the other (i.e. is less costly and more effective). If the results lie in the north-east or south-west quadrants, the decision as to which is the preferred treatment is less clear (i.e. one option may be less costly but also less effective, or more effective but at greater cost); the ICER may aid this decision. The bootstrapping was also used to estimate CIs for both costs and effects from the four-arm and at-the-margins analyses. A cost-effectiveness acceptability curve was also used to present the probability of a treatment being cost-effective based on a range of values for society’s willingness to pay. In the four-arm comparison we compared the bootstrapped results of the mean costs and mean outcomes for each treatment option against every other treatment option.

An extreme sensitivity analysis was conducted on our base-case analysis removing the most costly 10% of the participants from each treatment arm. There was a number of participants across the treatment arms that had an extended period of admission and this appears to be a key cost driver in our analysis.

Results

There were 1103 participants recruited into the STOPAH trial. Four participants who withdrew from the trial and withdrew consent for their information to be used were excluded from the analysis, seven participants who were randomised incorrectly were also excluded; this was consistent with the statistical analysis. One participant was randomised after they died so they were also excluded from the health economics analysis. Of the 1091 participants left in our analysis, 223 were affected by the trial extension for recruitment and we have information only on their 28-day status. Nearly 400 participants died during the trial period and only 223 participants were alive and completed the 1-year trial follow-up.

The 1091 participants that were used in the health economic analyses were evenly dispersed across the four treatment arms; 272 receiving standard care (placebo), 274 receiving prednisolone and placebo, 273 receiving PTX and placebo and 272 receiving the dual treatment (prednisolone and PTX).

Costs were estimated during the initial admission and during the follow-up period. We initially planned on presenting the combined total average cost of follow-up for each treatment arm, however, owing to the reduction in the follow-up period because of the extended recruitment period we decided to present follow-up costs individually.

Tables 31 and 32 describe the average health-care resource use by the randomised arms at 90 days for both liver-related and non-liver-related use of services. There were 415 use of medical services questionnaires returned at day 90, and of these, five were completely blank. For our analysis we assumed that any participant who partially completed the questionnaire left their other responses blank, as they were not applicable to them. This assumption was explored in the sensitivity analysis. For participants who had died during the 90-day follow-up period, their resource use was automatically imputed as 0. This could cause an underestimation in our costs as they may have used some services within the data collection period but before they died. In our preliminary analysis of the health-care resource use data we discovered some unintuitive results reported by participants; an example is reporting 215 inpatient nights in the last 90 days. For our analysis we decided to cap these unintuitive responses at 90.

Tables 33 and 34 present similar data to Tables 31 and 32, but for the 1-year follow-up period.

Table 31 highlights that the prednisolone only treatment arm has a higher, on average, liver-related secondary care resource use. The number of inpatient nights for non-liver-related conditions for the prednisolone-only treatment arm is also the highest out of all treatment arms (see Table 32). The primary care resource use for the PTX/placebo treatment arm is relatively low compared with the other treatment arms, especially in relation to GP visits and office visits with a social worker or care advisor. The standard care treatment arm has the highest prescription use reported for both liver- and non-liver-related causes. For all other health-care resource use the pattern of resource use is similar for all treatment arms. A similar pattern of resource use is observed in Tables 33 and 34.

Utility data were collected at discharge, day 90 and 1 year but were used only in our 1-year analysis. Preliminary utility scores for each treatment arm suggest that, on average, the standard treatment arm has the highest utility values at discharge. At 90 days, the PTX-only arm has the highest average utility score but at 1 year, PTX only has the lowest average utility score. At 1 year, the standard treatment arm again has the highest average utility score (see Table 35). When interpreting Table 35 caution needs to be taken as the results are presented only for survivors who completed the EQ-5D-3L; this is because this table is also used as a data source for the model-based economic evaluation. The numbers with utility data in each treatment arm decrease by approximately 30% from discharge to the 90-day visit, with a further 40% decrease between the 90-day visit and 1-year visit. Imputations were made for missing data in the economic analysis at 1 year but they are not represented in Table 35. The assumptions made in the economic analysis at 1 year were explored in the sensitivity analysis and a multiple imputation method was adopted. For participants who partially completed the questionnaire we used the ‘worst-case scenario’ method of imputation. This assumption was explored and had little to no effect on the average utility scores for each treatment arms owing to the low number of participants it affected.

Primary analysis

The primary analysis conducted for the health economics focused on the costs and outcomes at the end of the treatment phase (day 28). The costs included in the 28-day analysis were the intervention costs, length of initial admission (including time spent in ICU), nutritional supplements, discharge procedures and adverse event investigations. Discharge procedures were considered only if a participant was discharged within the treatment phase. If a participant was discharged after the treatment phase, the length of their initial admission was capped at 28 days.

For participants whose discharge form included some information, but for whom complete data were not available, we assumed that what was recorded represented all procedures performed. If a participant was in hospital on any of the days during the treatment phase but had no laboratory test information, we assumed that the laboratory tests that were specified in the protocol were performed during and after the treatment phase. These laboratory tests included LFTs (AST, ALT, albumin, ALP, bilirubin and total protein), creatinine, urea, PT and FBC (haemoglobin, WBC, neutrophils and platelets). Other laboratory tests (e.g. glucose) were available and performed on other participants during the 28-day treatment phase, however, these additional tests were not specified in the trial protocol and hence not included in our imputation. This assumption was explored in the sensitivity analysis in our model-based analysis. This could result in a slight underestimation of laboratory costs for all treatment groups. The investigations performed as part of an adverse event were recorded in the adverse event form. We identified the time at which the investigation was performed based on each patient’s start date and the date of the investigation. Only investigations that were performed during the 28-day treatment phase were included in this analysis.

On average, standard treatment was the most costly treatment over 28 days. This was because of the longer initial admission and more frequent ICU admissions, on average, compared with the other treatment arms. The total cost on average for standard care was just under £4900 over the 28-day treatment phase. PTX alone was the only treatment to have a higher probability of death at 28 days than standard care and was, hence, on average, more costly and less effective (i.e. dominated) by prednisolone only and dual treatment. Both prednisolone only and dual treatment (prednisolone and PTX) dominated standard care and PTX only. At 28 days, prednisolone had on average a 0.023 higher probability of survival than standard care, while the dual treatment arm had an average 0.032 higher probability of survival (Table 36).

TABLE 36 - A 2 × 2 factorial design: within-the-table cost-effectiveness analysis

A, prednisolone; B, PTX; N/A, not applicable; O, placebo.

Patients whose status at 28 days was ‘unknown’ were excluded from this analysis.

Parameter	OO (n = 272)	AO (n = 274)	OB (n = 273)	AB (n = 272)
Mean costs, £ (SD)	4869 (8131)	3618 (4052)	4194 (2810)	3827 (2711)
Incremental cost/patient vs. OO, £	N/A	–1251	–675	–1042
Patients in each arm who died during the treatment phase, n (%)	45 (16.7)	38 (14.3)	50 (19.4)	35 (13.5)
Probability of mortality at 28 days (95% CI)	0.167 (0.1287 to 0.2169)	0.143 (0.1058 to 0.1878)	0.194 (0.1355 to 0.2344)	0.135 (0.0882 to 0.1654)
Probability of survival at 28 days	0.833	0.856	0.806	0.865
Incremental survival vs. OO	N/A	0.023	–0.027	0.032
Incremental cost/survival vs. OO	N/A	Dominant	Dominated	Dominant

An incremental analysis was performed across all treatment arms. If one treatment was not dominated by another treatment, then we calculated the ICER to compare these treatments. The least costly option was prednisolone. Only dual treatment was more effective than prednisolone alone but this was more costly and the incremental cost per additional survivor was over £26,125 (Table 37). This number is difficult to interpret but with other things remaining equal, every additional survivor at 28 days would need to gain almost 0.77 QALYs each over their remaining lifetime for the incremental cost per QALY to be £30,000 [note: PTX and prednisolone (dual treatment) has a 0.009 higher probability of an additional survivor at day 28]. Assuming the difference in cost remains the same then assuming the maximum willingness to pay per QALY is £30,000, then the QALY gain for each additional survivor would need to be 0.77 QALYs (209/30,000/0.009 = 0.77).

TABLE 37 - Incremental cost per additional survivor at 28 days

A, prednisolone; B, PTX; O, placebo; SD, standard deviation; WTP, willingness to pay.

Assumes no interaction between PTX and prednisolone.

Intervention	Cost (£)	Survival	Incremental cost per additional survivor (£)	Probability that intervention is cost-effective for different threshold values for society’s WTP for an additional survivor
				£0	£5000	£10,000
AO	3618	0.857	–	0.79	0.79	0.60
AB	3827	0.865	26,125	0.20	0.13	0.08
OB	4194	0.806	Dominated by AO and AB	0.01	0.08	0.30
OO	4869	0.833	Dominated by AO and AB	0.00	0.00	0.02

Figure 10 was produced by bootstrapping estimates of the mean costs and mean probability of dying across the four treatment arms. It is an illustrative example of the changes in the probability of each treatment being cost-effective at a range of willingness-to-pay values. It is analogous to a one-sided CI in that if it is judged that society is willing to pay an upper limit for the cost per survivor, then society will certainly pay a lower value.

FIGURE 10.

Cost-effectiveness acceptability curve: placebo vs. prednisolone vs. PTX vs. prednisolone and PTX.

At-the-margins analysis

The effects of both active treatments, prednisolone and PTX, were compared individually in two subsequent analyses [(prednisolone and placebo plus prednisolone and PTX vs. placebo and PTX plus placebo and placebo) and (prednisolone and PTX plus placebo and PTX vs. placebo and placebo plus prednisolone and placebo)]. Since prednisolone appeared to be the most cost-effective treatment at 28 days when compared with the other three treatments, our first analysis determined the cost-effectiveness of prednisolone vs. no prednisolone (see Table 38).

TABLE 38 - Cost-effectiveness analysis: prednisolone vs. no prednisolone

A, prednisolone; B, PTX; O, placebo; SD, standard deviation; WTP, willingness to pay.

Assumes no interaction between PTX and prednisolone.

Intervention	Cost, £ (SD)	Survival	Incremental cost per additional survivor	Probability that intervention is cost-effective for different threshold values for society’s WTP for an additional survivor
				£0	£5000	£10,000
Prednisolone (n = 546)	3722 (3448)	0.861	–	1.00	0.98	0.81
No prednisolone (n = 545)	4531 (6082)	0.800	Dominated	0.00	0.02	0.19

Table 38 shows the results of the at-the-margins analyses comparing prednisolone vs. no prednisolone. These results support the findings in the four-arm comparison, as prednisolone dominates no prednisolone (prednisolone and placebo plus prednisolone and PTX vs. placebo and PTX plus placebo and placebo). The results of this analysis were bootstrapped to show that the statistical imprecision surrounding the estimates of incremental cost and incremental survivors (see Figure 11) and probability of prednisolone being cost-effective compared with no prednisolone (see Figure 12). In Figure 11, caution needs to be taken when interpreting our results because we are looking at the incremental difference in mortality between both treatment arms. The majority of iterations produced from the bootstrapping analysis are in the south-west quadrant and this clearly illustrates that prednisolone is dominant because it has a lower cost on average and a lower probability of mortality compared with no prednisolone. Figure 12 highlights that as society’s willingness to pay for an additional survivor increases, the likelihood that prednisolone compared with no prednisolone would be considered cost-effective decreases, but it still remains the most likely to be considered cost-effective.

FIGURE 11.

Cost-effectiveness scatterplot: prednisolone vs. no prednisolone at 28 days.

FIGURE 12.

Cost-effectiveness acceptability curve: prednisolone vs. no prednisolone at 28 days.

The next at-the-margins analysis we conducted looked at PTX vs. no PTX. The incremental and bootstrapped results are presented in Table 39 and Figures 13 and 14.

TABLE 39 - Cost-effectiveness analysis: PTX vs. no PTX

A, prednisolone; B, PTX; O, placebo; SD, standard deviation; WTP, willingness to pay.

Assumes no interaction between PTX and prednisolone.

Intervention	Cost, £ (SD)	Survival	Incremental cost per additional survivor, £	Probability that intervention is cost-effective for different threshold values for society’s WTP for an additional survivor
				£0	£5000	£10,000
PTX (n = 545)	4012 (2765)	0.836	–	0.76	0.77	0.74
No PTX (n = 546)	4241 (6441)	0.845	25,444	0.24	0.23	0.26

FIGURE 13.

Cost-effectiveness scatterplot: PTX vs. no PTX at 28 days.

FIGURE 14.

Cost-effectiveness acceptability curve: PTX vs. no PTX at 28 days.

The PTX arm was slightly less effective but less costly than the no PTX arm. The incremental cost per additional survivor for no PTX compared with PTX was in excess of £25,000 (see Table 39). The plots of bootstrapped mean costs and differences in survival are shown in Figure 13. This figure illustrates the statistical imprecision surrounding estimates of survival but that it is highly likely that PTX is less costly than no PTX. The situation of the average iterations in the south-east quadrant suggests that despite the cost savings generated by PTX it has a very slightly higher probability of mortality compared with no PTX. The cost-effectiveness acceptability curve is illustrated in Figure 14. This figure illustrates that there is approximately a 75% chance that PTX would be considered cost-effective over the range of threshold values for society’s willingness to pay for an additional survivor.

Model-based analysis

Our model was used to estimate the cost-effectiveness of the treatment arms over a 1-year period and over a patient’s lifetime. The following section presents details on the methods we adopted, our results and our sensitivity analysis.

Methods

As previously mentioned (see Methods overview) we used a Markov model to estimate the relative cost-effectiveness of each of the four treatment arms over 1 year and over the patient’s lifetime time horizon. A daily cycle length was adopted as this patient group has a high mortality rate. The parameters used in the model were: cost of the treatment medications as our initial cost for each treatment arm; average total cost per day, if alive, of management; average utility per day if alive; and probability of mortality per day. By definition, both daily cost and utility were taken as 0 if the individual was dead.

The daily management cost per treatment arm incorporated: initial admission, ICU admission, discharge procedures, adverse events, liver biopsy, alcohol counselling, laboratory test, renal failure support, liver transplant, nutritional support and resource use at 90 days. The methods to determine these costs are described in Methods. After we estimated the total average resource use cost per treatment arm over 90 days we estimated the daily total average cost per treatment arm by dividing this figure by 90.

We then estimated the average utility per day by treatment arm for people alive on a given day. The utility score for someone who had died was taken as 0. To calculate the utility score by treatment arm per day we used the data reported in Table 35 and assumed that the baseline utility was –0.402. This value was chosen to represent the health state of patients with AH when they are admitted to hospital and enter the study. However, if a patient was discharged within 10 days we used the discharge utility score as the baseline score; this was to prevent us overestimating the treatment effect. This assumption was explored in a sensitivity analysis.

To estimate the utility score assigned to each day we first estimated the quality-adjusted life-days (QALDs) over the first 90 days of follow-up. This was accomplished using the trapezoid method.

⁽³⁾ QALDs = ( E Q 5 D b a s e l i n e × days to discharge ) + { [ ( E Q 5 D d i s c h a r g e − E Q 5 D b a s e l i n e ) / 2 ] × days to discharge } + [ ( E Q 5 D d i s c h a r g e × ( 90 − days to discharge ) ] + { [ ( EQ5D90 day − E Q 5 D d i s c h a r g e ) / 2 ] × ( 90 − days to discharge ) ] } .

To give an average utility score for the 90 days of follow-up, QALDs were then divided by 90. This value was used in the economic model as the utility score assigned for each day spent alive.

A small number of participants (< 1%) had their discharge utility score collected after 90 days, which was excluded from our analysis. This allowed us to maintain consistency with the assumptions made when estimating costs over 90 days. If a participant had a later discharge (i.e. later than 90 days) we assumed that this was ‘missing’ and used the multiple imputation method to estimate their utility score. We also reduced their length of discharge to 89 days for our QALD calculation so there was a difference of 1 day and preventing their utility score being multiplied by 0.

The probability of mortality on a daily cycle was estimated using the patients’ status at 1 year. We used information only on the participants that we had information for at 1 year. If the participant’s status was unknown (i.e. late randomisation or lost to follow-up), he/she was excluded from this calculation. We estimated the probability of mortality at 1 year on those who died, divided by the total number of patients with their status available at 1 year. This probability was converted into a rate, and this rate was converted into a daily probability of mortality for each treatment arm using the following method: [(1 + rate) ^{(1/365) + 1}]. We assumed the patients without a 1-year status would have the same probability of dying. We also included the annual probability of all-cause mortality and this was converted into a daily rate using the same methods as described above.

Figure 15 is an illustration of the model structure used to estimate 1-year results and extrapolate the trial results, and Table 41 reports the parameters used in our model.

FIGURE 15.

Economic model.

TABLE 41 - Parameters used in the economic model

A, prednisolone; B, PTX; O, placebo.

The range of the daily probability of all-cause mortality is estimated based on the assumption that the cohort has a starting age of 48 years and the daily rate changes on an annual basis until the cohort reach the age of 100 years or die.

Parameters	OO	AO	OB	AB
Daily probability of dying from AH	0.00162	0.00171	0.00168	0.00162
Daily probability of dying from all causesa	0.000007–0.000961	0.000007–0.000961	0.000007–0.000961	0.000007–0.000961
Initial cost (£)	0.00	10.48	18.10	28.58
Incremental cost per day spent alive, £ (SD)	95 (133)	80 (91)	78 (67)	80 (91)
Initial utility score	–0.402	–0.402	–0.402	–0.402
Incremental utility score for each day alive (SD)	0.3476 (0.3089)	0.3551 (0.2980)	0.2700 (0.3235)	0.3477 (0.2969)
Distribution form for costs	Gamma	Gamma	Gamma	Gamma
Distributions for cost/day – alpha	0.51020	0.77285	1.35531	1.46924
Distributions for cost/day – lambda	0.00537	0.00966	0.01737	0.01836
Distribution form for utilities	Beta	Beta	Beta	Beta
Distributions for utilities – alpha	0.47851	0.56062	0.23851	0.54692
Distributions for utilities – beta	0.89810	1.01814	0.64487	1.02604
Distribution form for daily probability of dying from AH	Beta	Beta	Beta	Beta
Distribution for daily probability of dying from AH – alpha	0.44064	0.46854	0.45864	0.44064
Distribution for daily probability of dying from AH – beta	272	274	273	272

The initial cost for each treatment arm is the cost of the initial medications. The distributions for incremental costs were estimated using a gamma distribution using the mean and variance to estimate the variables needed to define the gamma distribution. The incremental utility distributions were estimated using a beta distribution (again the mean and variance were used to estimate the parameters needed to define the shape of the beta distribution). The distribution of the daily probability of death owing to AH was estimated based on the number of people dying in each arm on a daily basis. Of note in this analysis is that the choice of parameter values is conservative for prednisolone alone with respect to mortality in particular.

Modelling results

The Markov model was run for 365 cycles in the first instance to present the results for the 1-year analysis. In this analysis, QALDs, as predicted by the model, have been converted into QALYs by dividing by 365.

The least costly and least effective option is PTX alone (Table 42). The next least costly option is prednisolone alone, which is associated with an incremental cost per QALY gained of £6924 compared with PTX alone. No treatment and PTX treatment are, on average, less effective and more costly than combination treatment (PTX and prednisolone) and unlikely to be cost-effectiveness compared with prednisolone alone.

Table 43 reports the undiscounted incremental cost-effectiveness over a 10-year horizon (which, given mortality rates, is equivalent to a lifetime time horizon). In this analysis, PTX alone is the least costly and least effective intervention, which is consistent with the lower daily cost assumed for this intervention. Dual treatment is dominated by prednisolone only as it is more costly on average and less effective. Prednisolone only has an ICER of £2867 so would be considered cost-effective compared with PTX only. No treatment is unlikely to be considered worthwhile. This analysis represents an extreme sensitivity analysis representing a scenario when prednisolone alone might be cost-effective. Table 44 reports a similar analysis but this time both costs and QALYs are discounted at the recommended 3.5% rate. The findings of this analysis are similar. In this case, standard treatment is dominated by dual treatment. This is because the standard treatment group have a marginally higher probability of survival but when the additional QALYs gained over a longer period of time are discounted, the average QALY gain for that treatment arm is lower than prednisolone only. Despite dual treatment not being dominated it would be highly unlikely to be considered cost-effective when compared with prednisolone only. What these two analyses illustrate is that relatively small differences in mortality, should they exist, could change conclusions about cost-effectiveness when a lifetime time horizon has been taken. Overall, prednisolone only appears to be the most favourable treatment option but caution needs to be taken interpreting these results.

Sensitivity analysis

As noted in the previous paragraph a PSA was conducted on our model parameters. The results of this analysis at both 1 year and a lifetime were very similar in that the likelihood of any one treatment being considered cost-effective was very similar (Figures 16 and 17). This reflects the considerable imprecision surrounding estimates used within the model and again represents a cautious interpretation of the evidence available. Taken together, the results of the deterministic and probabilistic economic evaluation would suggest that treatment with prednisolone alone is a promising treatment but there is insufficient economic evidence to be conclusive.

FIGURE 16.

One-year cycle PSA results.

FIGURE 17.

Lifetime (10-year) cycle PSA results.

We conducted a deterministic sensitivity analysis around some of the assumptions previously mentioned in the chapter that were used to estimate the average daily cost for each treatment option. In this analysis we assumed that all additional hospital admissions outside of the initial treatment phase (28 days) were in ICU. We used multiple imputations to estimate all missing utility values at discharge and 90 days. Table 45 presents the updated parameters that were used in our model and the results at 1 year after these changes were made. The key cost driver in this analysis was the cost of ICU admissions. This cost alone undermined any of the cost savings made by changing all of our other assumptions. In this analysis, despite producing similar results to our 1-year analysis, PTX alone would not be considered cost-effective because of the higher ICER (£85,427) associated with it, owing to the increase in costs. If patients with AH are treated in an ICU setting compared with a general ward, consideration needs to be taken on the effect this will have on overall cost-effectiveness of prednisolone.

TABLE 45 - Deterministic sensitivity analysis with 1-year Markov results

A, prednisolone; B, PTX; N/A, not applicable; O, placebo.

Parameter	OB	AO	AB	OO
Average daily cost, £ (SD)	349 (308)	360 (344)	373 (345)	407 (372)
Average daily utility score (SD)	0.2911 (0.2925)	0.3283 (0.2830)	0.3278 (0.2713)	0.3177 (0.2991)
Total average costs at 1 year, £	94,897	97,400	102,435	111,741
Incremental cost difference at 1 year, £	N/A	2503	5034	9306
Total average QALY score at 1 year	0.2129	0.2422	0.2455	0.2379
Incremental QALY difference at 1 year	N/A	0.0293	0.0033	–0.0077
Incremental cost/QALY at 1 year, £	–	85,427	1,525,455	Dominated by AO and AB

Therapeutic effects

Although the trial results show that prednisolone at a dose of 40 mg daily for 28 days reduces the mortality at 1 month of severe AH by approximately 30% (odds ratio 0.72, 95% CI 0.52 to 1.01; p = 0.056), this fails to reach the conventional threshold for statistical significance. In multivariate analyses, which allow small variations in baseline variables to be taken into account, the beneficial effects of prednisolone are much clearer (odds ratio 0.609, 95% CI 0.409 to 0.909; p = 0.015). Although the sample size is not powered for a multivariate logistic regression analysis, the adjusted odds ratio provides a more precise estimate of the treatment effect. Taken together with previous studies, summarised in Mathurin’s meta-analysis,17 it is reasonable to conclude that prednisolone does have a therapeutic benefit up to 28 days. However, after 28 days the survival curves for prednisolone-treated and prednisolone-untreated patients converge, and at 90 days and 1 year there are no survival benefits to receiving steroids. Survival beyond 28 days may be influenced by adverse events related to steroid use (see Adverse events) or, alternatively, may be influenced by other factors and, in particular, recidivism. We were able to show a major effect of recidivism at 90 days or mortality at 1 year, but do not have the data to demonstrate an effect at earlier time points.

Pentoxifylline has no impact on the mortality at 28 days, 90 days or 1 year. In previous studies,18 PTX has conferred survival benefit through protection of renal function because of acute kidney injury. In this trial the rate of acute kidney injury was low, so although there was numerically fewer acute kidney injuries in the PTX-treated patients this failed to reach statistical significance. The combination of prednisolone with PTX had no effect on patients’ survival at 28 days, 90 days or 1 year. Without being able to demonstrate an effect on survival or acute kidney injury it is difficult to claim any therapeutic benefit for PTX.

Adverse events

Prednisolone, as expected, was associated with an increased risk of infection, with 13% of patients experiencing an infection-related SAE in the prednisolone group compared with 7% in the control group. Approximately 35% of infection-related SAEs resulted in mortality but this proportion did not differ between treatment groups. The higher risk of an infection-related SAE continued in the prednisolone group after the 28 days but by 90 days the risk was equivalent, irrespective of the treatment arm. Infection is therefore likely to have accounted for the excess deaths after 28 days in the prednisolone-treated group that caused the convergence of the mortality curves. However, it is difficult to be sure that infection is the key cause of death in this group of patients, when mortality is usually associated with multiorgan failure. Drawing on previous studies12 as well as results from the STOPAH trial, it is possible to conclude that patients who fail to respond to steroids after 7 days, defined by a Lille score of > 0.45, may not benefit from further treatment and, therefore, steroids should be stopped to avoid further risk of infection. This is coherent with current European Association for Study of the Liver (EASL) guidelines. 53

Although PTX has previously been reported to protect against hepatorenal syndrome or acute kidney injury in patients with AH, this benefit could not be confirmed in the STOPAH trial as only 1.5% of patients in the PTX-treated group and 2.5% in the controls experienced acute kidney injury.

Prognostic scores

The three prognostic scores based on baseline clinical and laboratory variables did not perform well for the prediction of mortality at 28 days, with deteriorating performance for mortality at 90 days and 1 year. None of the scores gave area under the ROC curve > 0.75. The scores do not inform physicians on the likely response to prednisolone and, therefore, how to select patients who would benefit from the intervention. At present, no scoring system assesses the risk of infection, which would be helpful in initial treatment selection.

The 2011 trial of early liver transplantation for patients with AH relied on the Lille score to identify prospective candidates. 54 However, the data in the STOPAH trial suggest that patients with an adverse Lille score (> 0.45) have an expected survival of around 50% compared with the 25% in the original study. This rate of survival without transplantation does not justify the use of scarce donor organs, meaning that a new system for selecting transplant candidates should be considered.

Health economics

The health economic component of this study comprised two elements: a preference elicitation exercise using a SG experiment and an economic evaluation.

The preference elicitation component measured QoL directly using the SG, and indirectly using the EQ-5D-3L. The aim was to collect information on the health state valuation of Child–Pugh classification A, B and C. The results of the SG showed good face validity, with the Child–Pugh classification A health state being preferred to Child–Pugh classification B health state, and the Child–Pugh classification B health state being preferred the Child–Pugh classification C. The analysis also suggested high levels of convergent validity with the EQ-5D-3L responses.

These data would be useful for future studies exploring the cost-effectiveness of other treatments for alcoholic cirrhosis.

With respect to the economic evaluation, both prednisolone and PTX are very low-cost medications but the costs of managing AH within the UK NHS are high. Furthermore, the trial addressed the possibility that one or other, or indeed both, treatments may help alleviate some of the considerable morbidity and early morbidity associated with the condition. Given the profile of the condition it was plausible, however, that a more effective treatment might not be cost-saving simply because those who survive longer may need considerable further costly care. The economic evaluation was expressly designed to explore these trade-offs.

At 28 days, the total average cost of treatment varied between £3618 per patient (prednisolone only) to £4869 per patient (placebo patients). On average, in a within-the-table analysis, which compares all four arms, prednisolone only was the least costly treatment and slightly less effective than dual treatment. However, prednisolone was most likely to be considered cost-effective over the range of values considered for society’s willingness to pay for an extra survivor. Much stronger results in favour of prednisolone were found in the at-the-margins analysis, when prednisolone was compared with no prednisolone. PTX was unlikely to be considered cost-effective, except in the at-the-margins analysis comparing PTX vs. no PTX, when PTX was most likely to be considered cost-effective, primarily because the estimates of effects were slightly in its favour (see Figure 14).

In terms of the incremental cost per QALY, the results tended again to favour the use of prednisolone. However, in this case any conclusions would be tentative because the assumptions made in the model around survival worked slightly against prednisolone and also there were very limited data available up to 1 year.