Stratifying risk in the prevention of recurrent variceal hemorrhage: Results of an individual patient meta-analysis

ELIGIBILITY CRITERIA

Studies that included patients with cirrhosis undergoing prevention of recurrent esophageal variceal hemorrhage were included in this analysis if they were published as full articles, if patients were randomly allocated to receive the combination of EVL plus BB versus either EVL or BB alone, if information regarding overall rebleeding and death was available, and if the original data sets of the trials were available.

IDENTIFYING STUDIES, INFORMATION SOURCES, AND SEARCH

We performed a literature search up to December 30, 2015, in Medline, Embase, the Cochrane Central Register of Controlled Trials, and the Cochrane Database of Systematic Reviews, using the terms “rebleeding,” “recurrent hemorrhage,” “esophageal varices,” “variceal obliteration,” “beta-blockers,” “endoscopy,” “band ligation,” and “cirrhosis.” We also reviewed publications in personal reference lists and citation sections of the recovered articles; manually searched abstracts presented at meetings of the American Gastroenterological Association, the American Association for the Study of Liver Diseases, the European Association for the Study of the Liver, the American Society of Gastrointestinal Endoscopy, and the British Society of Gastroenterology; and looked at authors of the abstracts to identify whether these abstracts had been published in full. Finally, we also searched clinical trials at www.controlled-trials.com and www.clinicaltrials.gov.

STUDY SELECTION AND DATA-COLLECTION PROCESSES

Two reviewers independently selected the studies by screening the titles and abstracts to identify those that fulfilled the inclusion criteria (J.M., A.A.). Principal investigators of the trials that met eligibility criteria were contacted to obtain individual patient data for all participants in their trials.

The original data sets were checked for completeness and internal consistency and amended through correspondence with the principal investigators. To build the final data set of IPD, we excluded studies with insufficient information and those that could not be obtained upon request.

DATA ITEMS

Patient-level data were analyzed by two authors (J.M. and D.A.). The following data were extracted from the data sets: patient characteristics (age, sex, etiology of cirrhosis, serum bilirubin, serum albumin, prothrombin time, Child-Turcotte-Pugh [Child] class),11 presence of ascites or encephalopathy, outcomes (all-source rebleeding, variceal rebleeding, and mortality), and follow-up (time-to-event for each outcome). Careful initial evaluation was performed to ensure completeness of data and to check consistency of the results of the primary analyses for each trial with published reports.

OUTCOMES

Primary endpoints were all-source rebleeding and mortality, and the secondary endpoint was variceal rebleeding. As described by the principal investigators of the included trials, all-source and variceal rebleeding were defined as the presence of melena or hematemesis with an upper gastrointestinal or an esophageal variceal source, respectively, identified by emergency endoscopy after randomization. All-source rebleeding was chosen as a primary outcome because EVL may reduce the risk for variceal hemorrhage but can also cause bleeding from esophageal ulcers, leaving the number of bleeding events per patient unchanged. Mortality was defined as death from any cause. The rate of adverse events was also explored.

STATISTICAL ANALYSIS

The main analysis was a one-stage IPD meta-analysis comparing time-to-event data. This uses the raw data from each study and allows for adjustments for confounders and tests for interactions. Because traditional meta-analysis models are not suited to analyze individual time-to-event patient data, a multilevel Poisson mixed-effects model with random intercepts was used to model time-to-event data that are clustered within studies. In the multilevel Poisson mixed-effects model, the effect of interventions is estimated as a hazard ratio approximated by incidence rate ratios (IRRs). This approach models survival with a fixed treatment effect and a different baseline hazard function for each study. So, the assumption of proportional hazards for all studies is relaxed, while still assuming proportional hazards between treatment groups within each study.12 The effect of intervention was stratified by the severity of cirrhosis, as assessed by Child class (A versus B/C). We specifically checked for the modifier effect of Child on the interventions by including interaction terms in the models. A significant interaction would indicate that the treatment effect was different between Child A and Child B/C patients.

Within each stratum, the model was further adjusted by including the following confounders: etiology (alcohol versus viral versus others), ascites (presence versus absence), encephalopathy (presence versus absence), and the continuous variables age, bilirubin, and albumin. Prothrombin time was not included because of the variability in expressing the result of this variable among the trials (prothrombin time in some trials, prothrombin activity in others). Inclusion of confounders allows fine-tuning of the model within each stratum compared with the coarse adjustment using Child as the sole category. A covariate was considered a confounder factor if the difference between adjusted and unadjusted model coefficients for the intervention variables varied >10%. In such a case, the IRR was shown along with the list of confounders used for adjustment. Otherwise, we reported the model without adjustment for confounders. Heterogeneity among trials was measured with the random effect parameters of the model, transformed into median IRRs (MIRRs). The MIRR shows the extent to which the probability of the event is determined by the trial in which the patients were recruited. The closer the MIRR is to 1, the lower the variability of the individual risk of an event among trials.13

We performed two meta-analyses: the first explored trials that compared combination therapy versus BB alone (control group), while the second explored trials that compared combination therapy versus EVL alone (control group).

We carried out a sensitivity analysis to verify the robustness of our results on all-source and variceal rebleeding. First, we considered death as a competing event for rebleeding, and we performed a competing risk analysis including study as a covariate. Second, we tested, using logistic multilevel regression, the impact of including all the patients into the analysis of rebleeding, even those patients excluded from the main analysis because they lacked information on time-to-event. Finally, we performed a post hoc exploratory analysis of the effect of the association of BB with isosorbide-5-mononitrate (ISMN) in those four trials that compared pharmacological therapy and EVL versus EVL alone, two of which used BB alone as pharmacological therapy, while the other two used BB with ISMN.

All analyses were done with Stata v13 (Stata Statistical Software, Release 13; StataCorp LP, College Station, TX) and the “xtmepoisson” command.

STUDY SELECTION

Our search strategy identified 798 potential references. We excluded 792 references: 771 because they were clearly ineligible by reading the title and the abstract and 17 because of duplicity (Fig. 1). Four additional studies were also excluded: two of them because they were in abstract form only and data could not be obtained from the authors14, 15; another because its primary endpoint was the incidence and natural history of portal hypertensive gastropathy in patients with previous variceal bleeding treated with EVL or BB16; and the last one because it used the hepatic venous pressure gradient response to BB to allocate patients to combination therapy or BB alone, instead of randomly allocating consecutive patients.17 This left six studies for inclusion in the meta-analysis.

STUDY CHARACTERISTICS

Table 1 shows the general characteristics of the six studies that met inclusion criteria. All the trials, except the one by Ahmad et al.,18 had two study groups that compared the combination of EVL and BB versus BB (three trials) or versus EVL (four trials). The trial by Ahmad et al. compared four groups, BB plus ISMN combined with EVL, BB alone, BB plus ISMN, and EVL alone. The specific BB used was propranolol in two trials18, 19 and nadolol in four trials,20-23 at a dose adjusted to reduce resting heart rate by 25% or to 55 bpm18, 20-22 or to the maximum tolerated dose without reducing heart rate below 55 bpm.19, 23 In all studies, BB were started within 5 days of the index bleed. BB were associated with ISMN (20 mg twice daily) in the three trials that compared combination therapies with BB alone18, 22, 23 and in two of the four studies that compared combination therapy with EVL alone.18, 19 In one of the other two trials, BB were administered together with sucralfate.22 EVL sessions were performed with multiband devices in all the trials and repeated at intervals that varied between 10 days and 4 weeks. The mean follow-up of the trials varied from 14 to 23 months.

The Child score was used to assess the severity of cirrhosis in all trials, and the publication by Pugh et al.11 was only referenced in the trial by Lo et al.21 In all trials, rebleeding was considered as such if significant (Table 1). Significant rebleeding was defined in all studies, except the one by De la Peña et al.,20 by the need to transfuse more than two blood units or by a drop in hemoglobin >2 g/dL. De la Peña et al. required a drop in hemoglobin to consider rebleeding, but they did not describe the amount.20

Supporting Table S1 shows the methodological quality indices of the included trials.24 All trials had an adequate generation of allocation sequence and adequate allocation concealment. The baseline characteristics of treatment and control groups were balanced in all trials. No trial had blinded investigators, and only the trial by Kumar et al. had a blinded endoscopist.19 The reasons for withdrawal were clear in all trials.

IPD OBTAINED

The final IPD data set was built after excluding patients with insufficient information in the original data sets and those for whom data could not be obtained upon request to the principal investigators. Supporting Table S2 shows the reasons for and the number of patients excluded from the original data sets: 2 lacked Child class,21 1 lacked any clinical information,20 26 had noncirrhotic etiology of portal hypertension,19 and 10 were lost to follow-up immediately after randomization and discharge.19 In addition, 150 patients from the study of Ahmad et al.18 and 9 patients from the study of Lo et al.21 were excluded from the main analysis because they lacked information on time-to-event but were included in the analysis of rebleeding using logistic multilevel regression models. The data set from Lo et al.'s study lacked information regarding the source of rebleeding; thus, 120 patients from this study were not included in the analysis of variceal rebleeding.22 All patients included in the final data set had information regarding death. Because the studies lacked a uniform standard definition of adverse events and the information reported in the databases was insufficient, we could not perform a safety analysis.

PATIENT CHARACTERISTICS

Characteristics of patients allocated to each treatment group in the six included trials are shown in Supporting Table S3. The final IPD base was built with 768 patients: 350 patients allocated to combination of EVL and BB, 206 to EVL alone, and 212 to BB alone. Patient characteristics were similar between the combination therapy arms of both sets of trials and between the combination and the monotherapy arms of each set of trials (Table 2).

PRIMARY OUTCOMES IN TRIALS COMPARING COMBINATION THERAPY (EVL PLUS BB) VERSUS BB ALONE

All-Source Rebleeding

We used 278 patients from two trials to assess overall rebleeding (Fig. 2; Supporting Table S2).22, 23 Fifty-four out of 140 (39%) patients rebled in the combination therapy group, and 58 out of 138 (42%) patients did in the BB group. The IRR (adjusted for etiology, bilirubin, and encephalopathy) for overall rebleeding was 1.00 (95% confidence interval [CI], 0.68-1.47; P = 0.996); that is, there were no differences in rebleeding between treatment groups in the overall group.

Sixty-eight of the 278 patients were Child A, and 27 of them rebled (40%), 9 on combination therapy and 18 on BB. Eighty-five out of 210 Child B/C patients rebled (40%), 45 on combination therapy and 40 on BB. The combination of EVL plus BB was more effective than BB alone in preventing overall rebleeding in Child A patients (IRR, 0.40; 95% CI, 0.18-0.89; P = 0.025) but not in Child B/C patients (IRR, 1.36; 95% CI, 0.87-2.14; P = 0.180) (adjusted for etiology, bilirubin, and encephalopathy) (Fig. 2). This interaction (i.e., difference in response to therapy by Child strata) was statistically significant (P = 0.009). There was no statistically significant heterogeneity among trials (MIRR, 1.25; 95% CI, 1.06-2.50; P = 0.063).

PRIMARY OUTCOMES IN TRIALS COMPARING COMBINATION THERAPY (EVL PLUS BB) VERSUS EVL ALONE

Death

A total of 416 patients from four trials contributed to the analysis of mortality (Fig. 2; Supporting Table S2).18-21 Seventeen out of the 210 (8%) patients on combination therapy died compared to 33 out of the 206 (16%) patients on EVL, the IRR for mortality being 0.50 (95% CI, 0.28-0.89; P = 0.019). Heterogeneity among trials was present for this analysis (MIRR, 2.16; P < 0.001).

Four of the 104 Child A patients (4%) died, 2 on combination therapy and 2 on EVL, whereas 46 out of 312 Child B/C patients died (15%), 15 on combination therapy and 31 on EVL. Therefore, the combination of BB and EVL did not significantly influence mortality in the low-risk Child A patients (IRR, 1.11; 95% CI, 0.16-7.93; P = 0.915). However, in Child B/C patients combination therapy did result in a significantly lower mortality compared to treatment with EVL alone (IRR, 0.46; 95% CI, 0.25-0.85; P = 0.013). Figure 3 shows that the cumulative probability of survival was significantly (P < 0.01) higher in Child B/C patients on combination therapy but not in Child A patients.

SECONDARY OUTCOME

Rebleeding from a variceal source was assessed in 278 patients from the trials that compared combination therapy versus BB alone.22, 23 Forty-one of 140 patients (29%) rebled from varices in the combination therapy group versus 51 of 138 patients in the BB group (37%). The overall IRR (adjusted for bilirubin and encephalopathy) for variceal rebleeding was 0.81 (95% CI, 0.53-1.23; P = 0.320). When stratified by Child, 68 of the 278 patients were Child A, and 25 of them rebled (37%), 7 on combination therapy and 18 on BB therapy. Sixty-seven of 210 Child B/C patients rebled (32%), 34 on combination therapy and 33 on BB. Combination therapy (BB plus EVL) was more effective than BB alone in preventing variceal rebleeding in Child A patients (IRR, 0.26; 95% CI, 0.11-0.61; P = 0.002) but not in Child B/C patients (IRR, 1.24; 95% CI, 0.75-2.05; P = 0.410), showing a significant interaction (P = 0.002). No heterogeneity was detected among trials (MIRR, 1.23; 95% CI, 1.05-2.70; P = 0.109).

Variceal rebleeding was assessed in 220 patients from the trials that compared combination therapy versus EVL alone.19, 20 Twelve of 115 (10%) patients rebled from varices in the combination therapy group versus 18 of 105 (17%) patients in the EVL group. The overall IRR for variceal rebleeding was 0.52 (95% CI, 0.25-1.11; P = 0.091). When stratified by Child, 71 of the 220 patients were Child A, and 4 of them rebled (6%), 1 on combination therapy (25%) and 3 on EVL therapy (75%). Twenty-six of 149 Child B/C patients rebled (17%), 11 on combination therapy and 15 on EVL. Therefore, the combination of BB and EVL was not more effective than EVL in preventing variceal rebleeding either in Child A patients (IRR, 0.22; 95% CI, 0.02-2.13; P = 0.191) or in Child B/C patients (IRR, 0.63; 95% CI, 0.28-1.41; P = 0.264). There was no significant heterogeneity in this analysis (P = 0.497).

SENSITIVITY ANALYSIS

We performed two sensitivity analyses for overall and variceal rebleeding, one considering death as a competing event and the other including all the patients in the analysis, even those without information on time-to-event.

Competing risk regression confirmed the estimates of the main analysis (Table 3).

In the second sensitivity analysis, we tested whether exclusion of 159 patients from the studies by Ahmad et al.18 and Lo et al.21 who lacked time-to-event information could have biased the results. Logistic multilevel regression analyses of all patients with information on Child class and treatment outcome independently of time-to-event also confirmed the estimates of the model using time-to-event data. The only different result was the P value for interaction in the response to EVL by Child strata, which in contrast to the main analysis was marginally significant in this analysis (P = 0.067) (Table 3).

Additionally, the post hoc exploratory analysis in the trials that compared combination therapy versus EVL alone showed that compared to EVL alone trials that used BB without ISMN20, 21 showed a statistically significantly (P = 0.007) larger impact on reducing all-source rebleeding than those trials associating ISMN to BB,18, 19 whereas the effect on mortality was not significant (Supporting Table S4).