Systematic Review
An SLR was performed in accordance with PRISMA guidelines [
29] to identify trials of GLP-1 RAs in patients with T2D who are inadequately controlled on 1–2 OADs. Searches of MEDLINE
®, Embase, and the Cochrane Library were performed via Ovid on April 5, 2016, with updates occurring on October 3, 2016 and August 16, 2017 (Table S1 in the Electronic supplementary material, ESM). Searches of conference proceedings were also carried out for the EASD (2014–2016), the International Society for Pharmacoeconomics and Outcomes Research (ISPOR; 2014–2017), the International Diabetes Federation (IDF; 2013 and 2015), and the ADA Scientific Sessions (2014–2017).
Following a study screening hierarchy for exclusion, all titles and abstracts identified through the literature searches were screened by two reviewers to assess whether they met the PICOS (population, interventions, comparators, outcomes, study design) selection criteria (Table S2 in the ESM). It should be noted that the PICOS criteria were slightly broader than what was required for the current analysis and included additional populations of interest. This is because the SLR was designed to support another NMA assessing the efficacy and safety of once-weekly semaglutide and other GLP-1 RAs in patients who are inadequately controlled on basal insulin. Once title and abstract screening were completed, the reviewers reconciled any existing discrepancies between their selections of studies. The same two reviewers independently screened full-text articles for all studies identified as included at the title and abstract screening phase. When a consensus could not be reached between the two reviewers during reconciliation processes, a senior reviewer provided arbitration. In addition, data from digital curves were extracted using digital extraction tools. Any discrepancies observed between the data extracted by the two analysts were adjudicated by a third reviewer.
NMA Methodology
A NMA was performed, in accordance with guidance from the National Institute for Health and Care Excellence (NICE), ISPOR, and the Cochrane Institute [
30‐
34], to compare the efficacy and safety of GLP-1 RAs in patients with T2D. In the analysis, the primary intervention of interest was once-weekly semaglutide (0.5 mg and 1.0 mg), and the primary comparators of interest were all other licensed doses of GLP-1 RAs—liraglutide once-daily (QD), dulaglutide QW, exenatide twice-daily (BID), exenatide QW, lixisenatide QD, and albiglutide QW. While albiglutide is soon to be withdrawn from the market, it remains a comparator of interest as the reason for withdrawal was not related to the safety of the medicine [
35]. GLP-1 RAs were often taken with other background anti-diabetic medications in the trials. In order to reduce variability between populations across the different trials, the definition of the add-on to 1–2 OADs population was aligned as closely as possible to the relevant SUSTAIN trials of once-weekly semaglutide (the primary intervention of interest). Therefore, studies assessing GLP-1 RAs as an add-on to one OAD (defined as > 90% of patients inadequately controlled on metformin monotherapy, i.e., sufficiently similar to the patient population in SUSTAIN 2 or 7) or as an add-on to 1–2 OADs (defined as < 100% of patients inadequately controlled on two OADs, i.e., sufficiently similar to the patient population in SUSTAIN 3 and 4) were considered for inclusion. Trials involving patients inadequately controlled on one OAD that was not metformin were also excluded in order to reflect SoC and align with international guidelines [
3].
All trials identified in the SLR were examined for data on at least one outcome of interest, and the ability to form a best-case connected network was assessed. The feasibility of generating evidence networks for each of the 20 outcomes of interest (Table S2 in the ESM) was examined next; the outcomes of interest included glycemic control outcomes (e.g., change from baseline in HbA1c, proportion of patients achieving HbA1c < 7% [53 mmol/mol] or ≤ 6.5% [48 mmol/mol]), weight outcomes (e.g., change from baseline in weight, body mass index [BMI], proportion of patients achieving weight loss of ≥ 5 or ≥ 10%), SBP, fasting plasma glucose (FPG), postprandial plasma glucose, proportion of patients achieving HbA1c < 7% without weight gain and without hypoglycemia, and safety outcomes (including the incidence of discontinuations due to adverse events [AEs], nausea, vomiting, diarrhea, pancreatitis, and hypoglycemia [overall, severe, non-severe, nocturnal]).
All analyses of continuous outcomes were performed using a normal likelihood, identity link, shared parameter model, to account for both arm-level and trial-level data reported within the included studies. For the analysis of dichotomous outcomes, a binomial likelihood (assuming a normal distribution), logit link model was used. For each outcome, both fixed effects (FE) and random effects (RE) models were run, and the model with the better fit (based on the deviance information criterion [DIC] and average posterior residual deviance) was used. Exploratory meta-regression analyses were also performed to determine whether a covariate-adjusted model would provide a significantly better model fit. The NMA models were implemented using WinBUGS software (MRC Biostatistics Unit, Cambridge, UK) [
36] and employed a Bayesian framework with the use of uninformative prior distributions. Three Markov Monte Carlo chains were used, starting from different initial values of selected unknown parameters. Convergence for all models was assessed using standard diagnostic methods for evaluating convergence [
37]. In addition, autocorrelation plots were assessed to detect the presence of autocorrelation in the chains. Following this, model convergence inferences were made from data obtained by sampling for a further 20,000 iterations using all the samples. If models failed to converge, the feasibility of a Bucher indirect comparison was considered. Bucher indirect comparisons were calculated in STATA 13 (StataCorp., College Station, TX, USA), using the indirect command [
38].
The results of the NMA are presented as mean treatment differences or odds ratios (ORs) and an associated 95% credible interval (CrI). For continuous outcomes of interest, a treatment which offers a greater mean reduction from baseline is favored—for example a reduction in HbA1c (%), SBP (mmHg), and weight (kg). For dichotomous outcomes, a treatment which offers an increase in the OR is favored—for example, higher odds for achieving a HbA1c level < 7%. Where reducing the probability of achieving an outcome is favorable (e.g., a reduction in discontinuations due to AEs), a treatment which offers a reduction in the OR is favored. In Bayesian statistics, it is considered that differences exist only where the CrI does not include the null value for treatment differences or one for ORs. In certain cases, once-weekly semaglutide may achieve a numerical reduction/increase against a comparator, but unless the CrI excludes the null value (for treatment differences) or one (for ORs), it is assumed that there is no difference.
In addition to the mean treatment difference or OR, the median ranks of each treatment (and associated 95% CrI) are provided, where a treatment with a median rank of 1 is considered the best. If two drugs share a ranking, then both are assigned one value lower (e.g., if two drugs are both ranked as the second highest, they will both be given the median rank score of 3). An additional ranking outcome is also presented—the surface under the cumulative ranking curve (SUCRA). The SUCRA is a numerical summary statistic of cumulative ranking probability plots (the probability a treatment is among the top
n treatments [between the first and
nth rank]) [
39]. A higher SUCRA value indicates an increased possibility that a treatment is in the top rank. A treatment which is ‘certain to be the best’ will have a SUCRA value of 1 (i.e., 100%), and a treatment ‘certain to be the worst’ will have a value of 0 (i.e., 0%) [
39]. This simplifies information about the effect of each treatment into a single value, allowing the complex results from NMA networks to be expressed with relatively few numbers. When the median rank and SUCRA values are in accordance (e.g., if a treatment has the highest median rank and the highest SUCRA score), this adds further weight to the interpretation.
NMAs combine all available evidence from clinical trials to estimate treatment effects. As this involves combining direct and indirect measures of effect, it is important to examine whether or not these two ‘sources’ of evidence are consistent with one another. Accordingly, all NMAs were formally assessed for inconsistency using Bucher’s method (as outlined by the NICE technical support document [TSD] 4) [
33]. Informal assessments were also performed by comparing the results of the NMA with the data reported across the studies included in the evidence networks. In the event of evidence of a substantial inconsistency that may change the conclusions based on the analysis selected (direct, simple indirect, or NMA), the following three steps were taken [
33]: (1) data were verified for accuracy; (2) if data were found to be correct, meta-regression or restricted analyses were performed to address the imbalance driving the issue; (3) if the second step failed, or was infeasible, further analyses were considered, within the limitations of avoiding bias.
Finally, this article does not contain any new studies with human subjects or animals performed by any of the authors.