Introduction
Inflammatory bowel disease (IBD) comprising Crohn’s disease (CD) and ulcerative colitis (UC) affect an estimated 1.5 million Americans, 2.2 million individuals in Europe, and several thousands more worldwide.1, 2 The therapeutic goal has evolved from relief of IBD‐related symptoms to the more ambitious goal of mucosal healing, though considerable debate exists about the optimal definition for this endpoint.3-8 Supporting this is the recognition that symptom‐based disease activity scores, long widely used as markers of efficacy, correlate poorly with endoscopic inflammation.9, 10 In addition, emerging longitudinal data suggests that individuals who are able to attain mucosal healing have a superior long‐term prognosis including a lower risk of surgery, hospitalisations, and need for systemic steroids.4, 6, 11-14
The past two decades have witnessed a significant expansion in the number of treatment options available for CD and UC, initially with monoclonal antibodies against tumour necrosis factor α (anti‐TNF) (infliximab, adalimumab, certolizumab pegol, golimumab), and subsequently anti‐integrin (natalizumab, vedolizumab) therapies. In addition, there is recognition that use of these biologics in combination with a conventional immunosuppressive (azathioprine, mercaptopurine, methotrexate) may yield superior outcomes and improve durability of therapy.15, 16 Thus, with this increase in the number of individual drugs with distinct mechanisms of action as well as possibility of combination therapy, it is increasingly important that appropriate positioning of therapies within the treatment algorithm must be informed by comparative effectiveness and safety. While a few recent studies have provided such comparisons,17-21 many examined only clinical outcomes of remission and response. The few examining mucosal healing as an endpoint were restricted to UC and focused on biological therapies, and did not examine the utility of combination with a conventional immunosuppressive.17, 22-24
With this evolution of therapeutic goals, there is an important need for comparing the effectiveness of these agents in achieving the preferred endpoint of mucosal healing. Recent reviews have thus far been restricted to a single therapeutic class and failed to provide pairwise comparisons.25 To address this gap in the literature, we performed this systematic review and network meta‐analysis of relevant randomised controlled trials to (i) examine the efficacy of each therapeutic class in inducing and maintaining mucosal healing in moderate‐to‐severe CD and UC; and (ii) to perform network meta‐analysis utilising direct and indirect evidence from clinical trials to derive comparative efficacy of various therapies in achieving mucosal healing.
Methods
Data sources and study selection
A full overview of the search strategy is presented in Data S1. In brief, a systematic electronic search was performed on MEDLINE/PUBMED and EMBASE to identify relevant full‐text articles published between 1980 and 2015 examining mucosal healing as an outcome in randomised controlled trials in moderate‐to‐severe CD or UC.
Inclusion and exclusion criteria
To be eligible for inclusion, we required the study to be (i) a randomised controlled trial in adult patients with moderate‐to‐severe CD or UC including either a placebo arm or two active treatment arms; (ii) examined medications currently approved by the United States Food and Drug Administration (FDA) for treatment of CD or UC (azathioprine, methotrexate, mercaptopurine, infliximab, adalimumab, certolizumab, golimumab, natalizumab, vedolizumab); (iii) examine mucosal healing as an endpoint and (iv) were available in full text in the English language. Our exclusion criteria included studies of medications used only in mild‐to‐moderate CD or UC (budesonide, 5‐aminosalicylates), and those that were observational, nonrandomised or open label.
Data collection
The decision for inclusion of the study was made independently by two authors (AC, ANA) and disagreements settled through consensus. Data were extracted on the number of cases receiving active treatment(s) or placebo, the type, dose and interval of active treatment(s), and duration of follow‐up. For each included trial, the scoring system used to assess endoscopic endpoints was extracted, and mucosal healing defined as per the criteria used for each trial. For trials where there was more than one dose arm of an active treatment, the pooled estimates were included if both doses were used commonly in clinical practice (e.g. infliximab 5 mg/kg and 10 mg/kg). On the other hand, if the dosing in one arm was not commonly used in clinical practice (adalimumab 80/40 mg or golimumab 400/200 mg), then only data from the United States Food and Drug Administration approved dosing arm was used.
Study quality was assessed using the Cochrane collaboration tool assessing risk of bias in several domains – blinding of investigators, blinding of participants, personnel, and outcome assessors, completeness of outcome data, reporting of selected outcomes, random sequence generation, allocation concealment and other sources of bias.26 Studies judged to be at low risk in all domains were considered to be at a low risk of bias.
Statistical analysis – direct comparisons
Pooled absolute rates and odds ratios (OR) of mucosal healing for all treatments compared to placebo using direct comparisons was calculated using a random‐effect model using the DerSimonian‐Laird weights.27 A sensitivity analysis was performed using fixed‐effect models, and results are reported separately if they differed from the random‐effect model. Heterogeneity was assessed using the Cochrane’s Q and i2 statistic with values > 50% and P < 0.10 indicating significant heterogeneity. Analyses were performed separately for CD and UC, and induction and maintenance trials respectively. Trials that assessed mucosal healing at <24 weeks were considered induction trials while those that assessed it beyond this were considered maintenance studies. If a study offered endpoints at two different time points within the same time category (e.g. at week 30 and week 52), the later time point was used in the analysis. We assessed for publication bias using the Begg’s and Egger’s tests and constructed funnel plots to determine asymmetry. A P < 0.05 indicated statistical significance for all tests. Direct comparisons were performed using stata 13.2 (StataCorp, College Station, TX, USA).
Pairwise comparisons – network meta‐analysis
We conducted a random‐effect Bayesian network meta‐analysis to inform pairwise comparative efficacy of different treatments in the induction and maintenance of mucosal healing.28, 29 Similar to a traditional meta‐analysis, this model assumes that each trial for a pair of treatments estimates a mean treatment, which varied around a common mean with a shared between‐study variance. However, with a network meta‐analysis, each mean treatment effect is broken down into basic “parameters” that are unique for each intervention. By setting the treatment effect for an arbitrary reference treatment to zero (we chose placebo), we then compared these basic parameters to calculate the treatment effect (log odds ratio) and 95% credible interval (CrI) and the probability of superiority/inferiority between every pair of treatments. A probability of superiority of ≥97.5% corresponds to a credible interval that does not overlap 1 and was considered statistically significant. For comparisons where direct and indirect comparisons were combined, potential inconsistency was assessed.
As mucosal healing is a rare outcome and the networks included a relatively small number of studies, it is difficult to obtain a reliable estimate of the between‐study heterogeneity in a random‐effect network meta‐analysis. We relied on an advantage of Bayesian analyses, which require the specification of a prior distribution for each model parameter. For our primary analysis, we used the published prior of Turner for the between‐study variability in a semi‐objective outcome in trials of drug vs. placebo.30 We conducted sensitivity analyses using an uninformative prior in a random‐effect model and a fixed‐effect model (which assumes there is no between‐study variability). All chains were run with 10 000 burn‐in iterations followed by 10 000 monitoring iterations. Convergence was assessed by running three chains, inspecting the sampling history plots, and calculating Gelman–Rubin–Brooke statistics.
