Selection of studies

The target population was defined as adult (aged ≥18 years) T2DM patients, who despite receiving a stable dose of SU monotherapy for at least 8 weeks (at half the maximum dose or maximum tolerable dose), had inadequate glycaemic control (Table 1). The target population were required to receive a stable dose of SU monotherapy to ensure that the maximal effect of this medication was achieved and to ensure that HbA1c was stabilised at baseline[41]. Patients had received first-line SU therapy because metformin was considered inappropriate as per current clinical treatment guidelines[13]. Eligible studies were prospective randomised-controlled trials (RCTs), conducted in the target population, that evaluated therapies in combination with an SU as part of a dual therapy regimen. The interventions of interest were therapies (and specifically doses of the therapies) licensed in the EU for use in combination with a SU, and which comprise the treatment options at this stage of the treatment pathway. The specific comparators in this analysis therefore included: dapagliflozin (the first SGLT2 inhibitor), DPP-4 inhibitors, GLP-1 analogues and TZDs. Metformin and insulin were not included as comparators in this analysis since they are not relevant comparators at this stage of the treatment pathway in clinical practice[13]. Additionally, alpha-glucosidase inhibitors were not considered as they are not commonly used[42, 43].

For inclusion in the meta-analysis, studies were required to report at least one of the primary endpoints of interest at 24 +/- 6 weeks. The outcomes of interest were pre-defined in the study protocol and selected based on clinical priorities: efficacy measures selected were mean change in HbA1c from baseline, mean change in weight from baseline, mean change in systolic blood pressure from baseline and proportion (number) of patients experiencing at least one episode of hypoglycaemia. The risk of hypoglycaemia, while being a safety outcome, was included as an outcome of interest because it is of primary importance to the management of T2DM. Current EMA guidelines state that therapeutic confirmatory studies should typically be of 6 months duration[41]. Since not all studies report endpoints at 24 weeks, a study window of 24 weeks +/- 6 weeks was selected; all included studies would therefore have a minimum of an 18 week follow-up period, corresponding to the EMA guidelines for minimum recommended maintenance and titration periods (16 weeks plus 2 weeks, respectively)[41]. The pre-defined study protocol also permitted an expansion of the study window to 24 +/- 8 weeks as a sensitivity analysis for borderline studies.

A number of study exclusion criteria were also applied (Table 1). These included results from open label extensions of RCTs and studies of SU as part of a triple therapy regimen.

Systematic review methodology

A systematic review was conducted, using a structured search (Additional file1) via the OVID platform of the following electronic databases: CENTRAL (March 2013), Medline and Medline in Process (1946 to 9th April 2013), and Embase (1980 to 2013 week 14)[44]. The search strategy employed was set up to be sensitive, to ensure that all relevant publications would be retrieved by the search, although this strategy would also retrieve a high proportion of irrelevant abstracts which would subsequently need to be excluded during the screening process. In addition, the 2012 and 2013 (where available) conference proceedings from the following organisations were searched: American College of Cardiology (ACC), American Diabetes Association (ADA), American Heart Association (AHA), European Association for the Study of Diabetes (EASD), The Obesity Society and The International Diabetes Federation. The clinical trials registry (ClinicalTrials.gov) was hand-searched for unpublished trials. Citations were included in the systematic review if they met the pre-defined inclusion criteria (Table 1). The initial screening for relevant studies was based on the citation abstract and title. For the second stage, the full-text publication was retrieved and the citations were re-screened to ensure their eligibility for inclusion in the systematic review. Disputes as to eligibility were referred to a third reviewer (PF). A quality assessment of the included studies was conducted using the Cochrane Collaboration’s tool for assessing risk of bias[45].

Meta-analysis

All meta-analyses were conducted on a modified intent-to-treat population, which is defined as the set of patients who were randomised and received at least 1 dose of study medication. For some studies the efficacy analysis set was further restricted to those with baseline and ≥1 follow-up efficacy result. Due to the limited number of studies available, study arms were pooled by drug class to improve accuracy by increasing the amount of data available for each class-level comparison. The pooled summary measure for continuous endpoints is weighted mean difference (WMD) and for binomial outcomes is the odds ratio (OR). ORs were utilised in order to facilitate indirect comparisons using the Bucher method and the logit (log odds) model within the NMA[46, 47]. For continuous endpoints, both the mean and standard error were required for the meta-analysis. If the standard error was not reported then this was calculated from the confidence interval (CI) or the standard deviation (SD), or imputed from the observed standard errors using the method described in the Cochrane Handbook for Systematic Reviews of Interventions[48].

Direct meta-analysis and Bucher indirect comparisons

Fixed and random-effects direct meta-analyses were conducted in Stata IC version 12.1 using the metan package SJ9_2: sbe24_3[49, 50]. The random-effects model uses the method of DerSimonian & Laird, with the estimate of heterogeneity being taken from the fixed-effect Mantel-Haenszel or inverse variance model. The inverse variance approach calculates study weight based on the assumption that variance is indirectly proportional to the study importance, whereas the Mantel-Haenszel approach uses a similar methodology whilst assuming a fixed treatment effect. In most instances, the between-study heterogeneity based on the I² statistic from the direct random-effects meta-analysis could not be assessed due to the limited scope of the analysis.

Simple indirect comparisons were made using the Bucher method, in order to provide an assessment of consistency for the NMA[46, 51].

Network meta-analysis

In an NMA, treatment effects are calculated for all treatments using all available evidence in one simultaneous analysis[36–38]. NMA methods build on the principles of indirect comparisons and preserve the randomised comparisons within each trial[46, 51]. The NMA methodology was based on the National Institute for Health and Care Excellence (NICE) Decision Support Unit recommendations for random and fixed effect Bayesian network meta-analysis: for mean change in HbA1c from baseline and mean change in weight from baseline, a normal NMA model with identity link was used; for count of patients with hypoglycaemia, a binomial model with logit link was used[34].

All models were fitted to the data via Bayesian Markov Chain Monte Carlo methods (specifically Gibbs sampling) using WinBUGS[52]. The WinBUGS models were run for a minimum of 100,000 iterations to ensure model convergence. An estimate of how well the values predicted by the model fitted the observed dataset was provided by the mean residual deviances (total residual deviance divided by number of datapoints) as well as the deviance information criteria (DIC) output by WinBUGs[34]. The model with the lowest DIC output by WinBUGs was deemed to best predict a replicate dataset of the same structure of that observed, provided differences were more than 5[53, 54].

Both random-effects and fixed-effect models were tested. Random-effects NMA allows the true treatment effect (eg. OR between two treatments) to vary between studies due to heterogeneity. In these random-effects models, a uniform (uninformative) prior was used for the between-studies SD (as per Hasselblad[55] and Gelman[56]). However, due to the limited number of studies involved in this analysis, fixed-effect models were preferred in all instances.

The absolute change for each treatment was calculated using the relative effects and the baseline risk/absolute change for the reference arm[57]. The reference treatment was the arm with the most data available, ie. placebo-control.

Previous meta-analysis have shown that there is a correlation between baseline HbA1c and change in HbA1c over follow-up[58]. Furthermore it is known that glycaemic control is harder to achieve in overweight or obese patients[59]. To account for the potential differences in baseline values across RCTs, and given the potential effect modification that can be attributed to baseline HbA1c and baseline weight, covariate analyses using baseline values were performed for both HbA1c and weight primary endpoints using the methodology recommended in Evidence Synthesis, Technical Support Document 3[60].

Systematic review search results

The OVID database search retrieved 2,923 citations of which 1,848 were unique. In addition, 52 abstracts were retrieved from the conference proceedings search and 1 unpublished study was identified from the clinical trial registry. After reviewing the title/abstract, 1,870 citations were excluded, with a further 24 excluded following full-text review (Additional file2). Seven citations (representing 7 different studies) were eligible for inclusion in the systematic review, with 5 studies fulfilling the eligibility criteria for inclusion in the meta-analysis (Figure 1, Table 2).

Summary of included and excluded studies

Three classes of anti-diabetes treatment were covered by the 5 studies fulfilling the inclusion criteria (Figure 2): DPP-4 inhibitors (3 studies), GLP-1 analogues (1 study) and SGLT2 inhibitors (1 study). All studies identified were placebo-controlled. Study duration ranged from 18 weeks to 30 weeks, including 3 studies that reported endpoints at the 24 week benchmark (Table 2)[31, 61–64]. Overall the included studies were comparable in terms of HbA1c, age and body mass index (BMI) patient entry criteria, and the baseline characteristics were similar across the studies (Additional file3). The quality assessment of the included studies indicated a low risk of bias overall (Additional file4).

Of the 24 studies excluded following review of the full-text, 6 studies were excluded based on the dose of SU received at randomisation (Additional file2). A stable dose of SU was critical not only to ensure that SU treatment had reached maximum efficacy but also to avoid potentially confounding differences in HbA1c at baseline. An additional 7 studies were excluded because they did not evaluate a dual therapy comparison of interest (Additional file2).

From the studies eligible for inclusion in the meta-analysis, sufficient data were reported for 3 of the key outcomes at the desired follow-up: change in HbA1c, change in weight and number of patients reporting hypoglycaemia. Only 2 studies reported mean change in systolic blood pressure and neither reported the corresponding standard errors. As a result this outcome was not analysed due to insufficient data.

Direct meta-analysis

Based on the fixed-effect direct meta-analysis, all classes of anti-diabetes treatments resulted in a statistically significant greater decrease in HbA1c at follow-up compared to placebo (p < 0.01) (Table 3). Dapagliflozin treatment resulted in a statistically significant larger decrease in weight at follow-up compared to placebo (-1.54 kg [95% CrI: -2.16, -0.92]; p < 0.01), whereas DPP-4 inhibitor therapy resulted in a statistically significant increase in weight at follow-up compared to placebo (0.57 kg [95% CrI: 0.09, 1.06]; p < 0.02). GLP-1 analogue treatment did not lead to a statistically significant change in weight but it did result in a statistically significant higher odds of hypoglycaemia at follow-up compared to placebo (10.89 [95% CrI: 4.24, 38.28]; p < 0.01). In contrast, the odds of hypoglycaemia for dapagliflozin and DPP-4 inhibitors were not significantly different to placebo (p > 0.05) (Table 3, Figure 3).

Comparison	HbA1c		Weight (kg)		Hypoglycaemia
	NMA (95% CrI)	Direct/Bucher indirect (95% CI)	NMA (95% CrI)	Direct/Bucher indirect (95% CI)	NMA (95% CrI)	Direct/Bucher indirect (95% CI)
	WMD v placebo	Direct comparison	WMD v placebo	Direct comparison	OR v placebo	Direct comparison
DPP-4 v placebo	-0.56 (-0.70, -0.41)†	-0.56 (-0.70, -0.41)*	0.57 (0.09, 1.06)†	0.57 (0.09, 1.05)*	1.87 (0.82, 4.72)	1.8 (0.78, 4.17)
GLP-1 v placebo	-0.80 (-1.04, -0.56)†	-0.79 (-1.02, -0.55)*	-0.65 (-1.37, 0.07)	-0.65 (-1.37, 0.07)	10.89 (4.24, 38.28)†	10.05 (3.60, 28.06)*
SGLT2 v placebo	-0.69 (-0.86, -0.52)†	-0.69 (-0.86, -0.52)*	-1.54 (-2.16, -0.92)†	-1.54 (-2.16, -0.92)*	1.75 (0.67, 4.89)	1.71 (0.66, 4.48)
	WMD head-to-head	Bucher indirect comparison	WMD head-to-head	Bucher indirect comparison	OR head-to-head	Bucher indirect comparison
GLP-1 v DPP-4	-0.24 (-0.52, 0.04)	-0.23 (-0.50, 0.04)	-1.23 (-2.09, -0.36)†	-1.22 (-2.09, -0.35)*	5.89 (1.56, 26.06)†	5.58 (1.49, 20.99)*
SGLT2 v DPP-4	-0.13 (-0.35, 0.09)	-0.13 (-0.35, 0.09)	-2.11 (-2.90, -1.33)†	-2.11(-2.90, -1.33)*	0.94 (0.25, 3.50)	0.95 (0.27, 3.37)
SGLT2 v GLP-1	0.11 (-0.18, 0.40)	0.10 (-0.19, 0.39)	-0.89 (-1.84, 0.07)	-0.89 (-1.84, 0.06)	0.16 (0.03, 0.65)†	0.17 (0.04, 0.69)*
	WMD v baseline		WMD v baseline		Probability of hypoglycaemia
placebo	0.05 (-0.10, 0.20)	N/A	-0.48 (-0.78, -0.18)†	N/A	3.99% (2.30%, 6.85%)	N/A
DPP-4	-0.51 (-0.72, -0.30)†	N/A	0.09 (-0.47, 0.66)	N/A	7.23% (2.75%, 18.67%)	N/A
GLP-1	-0.75 (-1.03, -0.47)†	N/A	-1.13 (-1.91, -0.35)†	N/A	31.38% (13.00%, 64.10%)	N/A
SGLT2	-0.64 (-0.87, -0.42)†	N/A	-2.02 (-2.71, -1.33)†	N/A	6.81% (2.33%, 19.04%)	N/A
	Model parameter results
DIC (WinBUGs)	-10.36	N/A	13.07	N/A	72.301	N/A
Average resdev	1.37	N/A	1.14	N/A	2.25	N/A

*Statistically significant result (p < 0.05); †statistically significant based on 95% CrI; CI, confidence interval; CrI, credible interval; DIC, deviance information criterion; DPP-4, dipeptidyl peptidase-4 inhibitors; GLP-1, glucagon-like peptide-1 analogues; OR, odds ratio; NMA, network meta-analysis; resdev, residual deviance; SGLT2, sodium glucose co-transporter 2 inhibitors; WMD, weighted mean difference.

Between-study heterogeneity, based on the I² statistic from the direct random-effects meta-analysis, was assessed for DPP-4 inhibitors only, since no other class of therapy had more than one trial. No significant heterogeneity between the 3 DPP-4 inhibitor trials was observed for any of the outcomes: HbA1c I² = 0%, p = 0.61; weight I² = 0%, p = 0.41; hypoglycaemic events I² = 0%, p = 0.64 (Figure 3).

Bucher indirect comparisons

Pairwise Bucher indirect comparisons revealed a statistically significant difference in the weighted mean difference in weight between dapagliflozin and DPP-4 inhibitors (-2.11 kg [95% CI: -2.90, -1.33; p < 0.01]) and between GLP-1 analogues and DPP-4 inhibitors (-1.22 kg [95% CI: -2.09, -0.35; p = 0.01]). It also demonstrated that the odds of hypoglycaemia with dapagliflozin treatment or DPP-4 inhibitor treatment are significantly lower than GLP-1 analogue treatment (0.17 [95% CI: 0.04, 0.69; p = 0.01] and 0.18 [95% CI: 0.05, 0.67; p = 0.01] respectively). All other Bucher indirect comparisons showed no statistically significant differences in outcomes between treatments (Table 3).

NMA

For the HbA1c data, the random-effects model without covariates provided the lowest DIC value. However, due to the low number of studies in the network, the random-effects model overestimates the uncertainty, which is evident from the high between study SD and large credible intervals (CrI) which includes an unrealistic range of values (Figure 4). The most reliable estimates in terms of CrI were therefore provided by the fixed-effect NMA, which demonstrated that all 3 classes of treatment resulted in statistically significant larger decrease in HbA1c at follow-up compared to placebo (based on the 95% CrI) (Table 3). The coefficients estimating the impact of baseline HbA1c on treatment effects were not significant for either the fixed or random-effects models (Figure 4). It was noted, however, that the inclusion of the baseline HbA1c covariate had a larger impact on dapagliflozin than the DPP-4 inhibitor and GLP-1 analogue classes. This may be explained by the observation that for both DPP-4 inhibitors and GLP-1 analogue study arms, baseline HbA1c was close to the average baseline, while the baseline HbA1c for the dapagliflozin trial was lower than the study arm average.

The fixed-effect model provided the best overall fit for the change in weight data (Table 3). As with the HbA1c outcome, both random-effects models had high between study SD and large CrI, indicating substantial uncertainty in the estimates from these models (Figure 4). Based on the fixed-effect NMA, dapagliflozin treatment resulted in a significantly larger decrease in weight at follow-up compared to placebo (-1.54 kg [95% CrI: -2.16, -0.92]), which is in contrast to treatment with GLP-1 analogues (-0.65 kg [95% CrI: -1.37, 0.07]) and DPP-4 inhibitors (0.57 kg [95% CrI: 0.09, 1.06]). The coefficients estimating the impact of baseline weight on treatment effects were not significant for either the fixed or random-effects covariate models.

Two models were evaluated as part of the NMA for the hypoglycaemia endpoint; fixed-effect and random-effects models. The random-effects model provided the best overall fit for the hypoglycaemia data, but both models produced large CrI (Figure 4), and the between-study SD was high. Given the lack of studies, the fixed-effect NMA provides the most robust estimates (Table 3), and indicates that the odds of hypoglycaemia for placebo were similar to dapagliflozin and DPP-4 inhibitor add-on treatment, but significantly greater than placebo for GLP-1 analogue add-on treatment (10.89 [95% CrI: 4.24, 38.28]).

Results of pairwise comparisons estimated from the NMA were comparable to the Bucher indirect estimates (Table 3). No statistically significant differences in weighted mean difference for HbA1c were seen between drug treatments. Both dapagliflozin and GLP-1 analogues resulted in a larger decrease in weight compared to DPP-4 inhibitors, with the weighted mean difference being larger for dapagliflozin vs DPP-4 inhibitors than for GLP-1 analogues vs DPP-4 inhibitors (-2.11 kg [95% CrI: -2.90, -1.33] and -1.23 kg [95% CrI: -2.09, -0.36] respectively). Although dapagliflozin resulted in a larger decrease in weight compared to GLP-1 analogues, this was not statistically significant. Both dapagliflozin and the DPP-4 inhibitors demonstrated statistically significant lower odds of hypoglycaemia compared to GLP-1 analogues (0.16 [95% CrI: 0.03, 0.65] and 0.17 [95% CrI: 0.04, 0.64] respectively). No other statistically significant differences between treatments were seen.

Strengths and limitations of the analysis

A general limitation of meta-analysis is the underlying assumption that trials and outcomes are sufficiently similar to allow for data to be pooled. In order to ensure that this assumption held in this analysis, the systematic review was designed to be highly specific with strict inclusion criteria. All exclusions were made based on the pre-defined criteria, and primarily designed to ensure that the analysis remained focused on a specific target population and clinically relevant comparators. This focus to the review and NMA also makes the results meaningful to clinical decision making and appropriate for cost-effectiveness analysis[74].

There will always be differences between studies, and some degree of heterogeneity in the analysis may provide a measure of uncertainty in effects due to differences in clinical practice. One potential source of heterogeneity is differences in follow-up times across studies. A 24 week timepoint was chosen since most registration studies are of this duration. The follow-up was allowed to vary by 24 +/- 6 weeks to allow for a pooled analysis of short-term endpoints, without impacting on weight and hypoglycaemia endpoints, which would accumulate over time. The impact of differences in SU background treatment was not considered to be a major limitation in this analysis since trial arms were balanced at baseline in terms of SU treatment and the SU dose was stable during the trial. Of the five studies eligible for inclusion, three studies used glimepiride as the background SU during follow-up, and while two studies did not specify the SUs used as background treatment, prior SU treatment duration was at least 8 weeks pre-enrolment for all studies (Table 2).

The limited availability of studies has hampered our assessment of heterogeneity by limiting the scope of the sensitivity analysis that can be conducted. As a result, there are insufficient studies with which to estimate the between-study variance with precision; therefore the CrI in the random-effects models may reflect uncertainty due to a lack of data rather than the true variance in treatment effects. Based on the size of network on which the analyses were performed therefore, the fixed-effect analyses are viewed to be the most robust estimates of treatment effects.

Results obtained from the NMA were consistent with results from the direct meta-analysis and Bucher indirect comparisons. This is due to the ‘star’ shaped network whereby all studies are connected via the placebo-control. In addition, the analyses presented here have been performed on a small number of studies. One advantage of NMA over the Bucher method is that it uses all the available placebo data in one analysis, while the Bucher method incorporates only some of the placebo data in each pairwise comparison. Inclusion of all the available data for NMA typically increases the power of the analysis. However this was not the case in this analysis given the small number of studies.