A population‐adjusted indirect comparison of cardiovascular benefits of once‐weekly subcutaneous semaglutide and dulaglutide in the treatment of patients with type 2 diabetes, with or without established cardiovascular disease

Abstract Introduction Cardiovascular (CV) effects of once‐weekly subcutaneous (s.c.) semaglutide 0.5 and 1 mg and dulaglutide 1.5 mg are reported in their respective placebo‐controlled cardiovascular outcome trials (CVOTs), SUSTAIN 6 and REWIND. There is no head‐to‐head CVOT comparing these treatments and heterogeneity between their CVOTs renders conventional indirect comparison inappropriate. Therefore, a matching‐adjusted indirect comparison (MAIC) was performed to compare the effects of s.c. semaglutide and dulaglutide on major adverse cardiovascular events (MACE) in patients with and without established cardiovascular disease (CVD). Methods Individual patient data from SUSTAIN 6 were matched with aggregate data from REWIND, using a propensity score method to balance baseline effect‐modifying patient characteristics. Hazard ratios (HRs) for three‐point (3P) MACE (CV death, non‐fatal myocardial infarction, non‐fatal stroke), anchored via placebo, were then indirectly compared between balanced populations. Sensitivity analyses were performed to test the robustness of the main analysis. Results After matching, included effect modifiers were balanced. In the main analysis, s.c. semaglutide was associated with a statistically significant 35% reduction in 3P MACE versus placebo (HR, 0.65 [95% confidence interval [CI]; 0.48, 0.87]) and nonsignificantly greater reduction (26%) versus dulaglutide (HR, 0.74 [95% CI; 0.54, 1.01]). Results were supported by all sensitivity analyses. Conclusions This study demonstrated a statistically significant lower risk of 3P MACE for s.c. semaglutide versus placebo, in a population with lower prevalence of pre‐existing CVD than that in the pre‐specified primary analysis in SUSTAIN 6. Reduction in 3P MACE with s.c. semaglutide was greater than with dulaglutide, although not statistically significant.


| INTRODUC TI ON
Type 2 diabetes (T2D) is a chronic and progressive metabolic disorder associated with an elevated risk of microvascular and macrovascular complications, including cardiovascular disease (CVD), which can result in considerable morbidity and mortality. [1][2][3][4] Previous studies have shown that, while the effect of intensive blood glucose control decreases the risk of microvascular complications after a median of 5 years of follow-up, 5 its effect on macrovascular complications is only observed in the longer term for some cardiovascular (CV) outcomes. 6 However, more recently, some glucose-lowering medication classes have demonstrated significant CV benefit versus placebo in far shorter timeframes in their cardiovascular outcomes trials (CVOTs). These include glucagon-like peptide receptor agonists (GLP-1 RAs) and sodium-glucose co-transporter-2 inhibitors (SGLT-2is). 7 For patients with T2D who have established CVD or indicators of high risk of CVD, GLP-1 RAs and SGLT-2is are recommended by the American Diabetes Association (ADA) and European Association for the Study of Diabetes (EASD), European Society of Cardiology and American College of Cardiology. 3,8,9 However, based on findings of CVOTs, the ADA and EASD recommend GLP-1 RAs as the preferred option when atherosclerotic CVD predominates and SGLT-2is as the preferred option when heart failure (HF) or chronic kidney disease predominates. 8 As well as differences between treatment classes, previous analyses suggest that CV benefit may vary within treatment class. 7,10 There are currently no head-to-head randomized controlled trials (RCTs) comparing CV benefit within treatment classes, and, in the absence of Food and Drug Administration (FDA) guidance on a standardized approach to the design of CVOTs, differences in study design between some CVOTs can make indirect comparison challenging. Robust, within-class comparison could help to guide decisions on which product in a treatment class should be used to treat individual patients with T2D with CVD or CV risk factors.
Guidelines from the ADA and EASD specify that the GLP-1 RA products used to treat patients with T2D and established CVD or at high risk of CVD should have proven CVD benefit, defined as having a label indication of reducing CVD events. 8 In early 2020, two GLP-1 RAs with once-weekly dosing regimens, subcutaneous (s.c.) semaglutide and dulaglutide, were both approved by the FDA in this indication. 11,12 The CV effects of s.c. semaglutide were assessed in SUSTAIN 6, which demonstrated a statistically significant 26% reduction in the risk of three-point (3P) major adverse cardiovascular events (MACE) (CV death, non-fatal myocardial infarction [MI] or non-fatal stroke) versus placebo in patients with T2D with established CVD and/or CV risk factors. 13 The CV effects of dulaglutide were assessed in REWIND, which reported a statistically significant 12% reduction in the risk of 3P MACE with dulaglutide versus placebo with established CVD and/or CV risk factors. 14 In the absence of head-to-head data comparing s.c. semaglutide with dulaglutide, an indirect comparison of these treatments based on their respective CVOTs could help to determine the most suitable GLP-1 RA for patients with T2D at high CV risk.
Network meta-analysis (NMA) is a well-established method for conducting indirect treatment comparisons in the absence of headto-head trials between treatments. Recently published NMAs have compared CVOTs to assess the effect of glucose-lowering drugs on CV outcomes. 10,[15][16][17] However, NMA adopts assumptions of homogeneity and similarity to provide unbiased estimates of treatment effects, and there must be no relevant heterogeneity between trials, which must have similar study designs, patient populations and outcome measures, and must be comparable on effect modifiers. 18,19 When significant heterogeneity exists between trials, NMA is rendered inappropriate. Alternative indirect comparative methods are available that seek to overcome heterogeneity, and the choice of an appropriate method will be determined by the type of evidence available, as described by Lingvay et al, 2020. 20 Matching-adjusted indirect comparison (MAIC) is an alternative method that can be used when individual patient data (IPD) are available for a treatment of interest and only published aggregate data (collated by treatment arm) are available for the comparator. MAIC addresses differences in patient populations using a propensity score-based approach, which can provide a less biased estimate by weighting the IPD for an index treatment to match the aggregate baseline characteristics for a comparator. [21][22][23] In an unpublished NMA feasibility analysis for comparison of GLP-1 RAs, substantial heterogeneity was identified between CVOTs for s.c. semaglutide and dulaglutide in terms of patient baseline characteristics. Patients enrolled in SUSTAIN 6 were more likely to have experienced a prior CV event than those enrolled in REWIND, with approximately twice the proportion of patients experiencing prior ischaemic stroke (11.6% vs. 5.3%, respectively) and/ or prior MI (32.5% vs. 16.2%, respectively). As such, an NMA was deemed unsuitable for comparing these CVOTs. Therefore, with the availability of IPD from SUSTAIN 6 and aggregate data from REWIND, a MAIC was performed. The objective was to assess CV outcomes with s.c. semaglutide in the population with fewer prior CV events as assessed in REWIND and to indirectly compare the relative effects of s.c. semaglutide versus dulaglutide on rates of 3P MACE for patients with T2D with or without established CVD.

| Overview of the MAIC methodology
The MAIC approach was first published in 2010 23 and has subse- Technical Support Document (TSD) 18. 21 The NICE guidance was accompanied by published code for use with the statistical package R, 24 to enable MAIC to be carried out according to the recommendations set out in TSD 18 (Appendix D of the publication). Within the therapeutic area of diabetes, the MAIC approach has previously been used to compare the efficacy of two treatments within the same treatment class (dipeptidyl peptidase-4 inhibitors) in a specific patient population. 25 Further details of the MAIC methodology are provided in the Supporting Information.

| Empirical approach and model specification
The methods used in the current study align with the NICE guidance. 21 A systematic literature review (SLR) for interventions studied in CVOTs was conducted, alongside the unpublished NMA feasibility assessment, with a particular focus on GLP-1 RA comparators. RCTs identified as relevant for the key treatments of interest were SUSTAIN 6, 13 for which IPD were available, and REWIND 14 with aggregate data. The 0.5 and 1 mg doses of s.c.
semaglutide from the SUSTAIN 6 trial were pooled, as were the matching placebo arms, as the interest was in outcomes associated with s.c. semaglutide, not the specific doses. This increased the potential pool of patient data for s.c. semaglutide and was consistent with the results presented in the SUSTAIN 6 publication. These trials had a common comparator in placebo and thus an anchored MAIC could be conducted. PIONEER 6 26 was also identified as a CVOT for which a different formulation of semaglutide (for once-daily oral administration) was reported. Although not included in the main analysis, PIONEER 6 was included in a sensitivity analysis to compare the CV effects of the semaglutide molecule with dulaglutide.
Patient populations in SUSTAIN 6 and REWIND were similar in terms of age, gender and race (Table 1). However, patients in SUSTAIN 6 had a longer duration of diabetes than those in REWIND and a higher baseline HbA 1c , with higher proportions of patients receiving insulin therapy. In addition, SUSTAIN 6 included higher proportions of patients with a history of CVD than REWIND, along with higher proportions of patients with some CV risk factors (estimated glomerular filtration rate [eGFR] <60 ml/min/1.73 m 2 and albuminuria [urinary albumin-to-creatinine ratio (UACR) ≥3.39 mg/mmol]).
Effect-modifying variables are not well established for CV outcomes in patients with T2D. Therefore, to enable matching in the analysis, potential effect modifiers were identified using input from clinical experts. For the purpose of this study, potential effect modifiers were identified as prior HF, prior MI, prior stroke or transient ischaemic attack (TIA), peripheral arterial disease (PAD), eGFR and albuminuria. Of these, prior MI, prior stroke or TIA and existing albuminuria were also described as 'clinically relevant baseline characteristics for the primary endpoint' in the REWIND trial publication. 14 The six potential effect modifiers were considered to allow evaluation of the effect of GLP-1 RAs across the spectrum of CV risk. In SUSTAIN 6 and REWIND, all variables were specified as dichotomous variables, that is, as a proportion of the patients in the trial with or without the condition at baseline. The exceptions were eGFR and UACR. In the primary REWIND publication, for eGFR, patients' renal function was categorized as normal/mild (eGFR ≥60 ml/min/1.73 m 2 ) and moderate/severe (eGFR <60 ml/min/1.73 m 2 ) (dichotomous data), as well as mean and standard deviation (SD) measurements (continuous data) from the exploratory renal analysis publication. 27 Similarly, for UACR, both the proportion of patients with albuminuria (dichotomous data) and the median and interquartile range (IQR) were available from the REWIND publications. Other risk factors, including diabetes duration, HbA 1c at baseline and smoking status, were not adjusted for in the model as these were considered to be prognostic and not effect-modifying factors. In an anchored MAIC, provided prognostic factors are balanced between study arms, it is recommended not to adjust for them in the model since this can lead to loss of precision in the estimate of the relative treatment effect. 21 In the main analysis, summarized in Figure 1, IPD from SUSTAIN 6 were matched to the aggregate effect modifier baseline data from REWIND, based on matching all identified potential effect modifiers, with eGFR and albuminuria categorized as dichotomous variables (<60 vs. ≥60 ml/min/1.73 m 2 and <3.39 vs. ≥3.39 mg/mmol, respectively). The matching for potential effect modifiers was achieved by running the relevant portion of the published NICE code in R (version 3.5.3). Thus, the s.c. semaglutide IPD were weighted to match the dulaglutide baseline characteristics using a form of propensity score model. 21,22 The weighting was calculated from the relevant baseline characteristic covariates only and was therefore independent of the outcome.
Details of the form of the propensity score modelling and weighting calculations can be found in the Supporting Information.

| Outcomes of interest
Three-point MACE was chosen as the primary outcome of interest as this was the primary endpoint in the identified trials. It was defined as first occurrence of death from CV causes (including undetermined death), non-fatal MI or non-fatal stroke. The published NICE code was modified to be suitable for use with time-to-event analyses such that, following matching, an adjusted hazard ratio (HR) for s.c. semaglutide versus placebo for 3P MACE was estimated in the target REWIND population. This was achieved by applying the calculated s.c. semaglutide patient weightings to the corresponding 3P MACE patient outcomes in a weighted Cox regression. The standard errors for the estimates were calculated using a robust sandwich estimator.
The relative treatment effect of s.c. semaglutide and dulaglutide in the REWIND population could then be indirectly calculated using the HR for s.c. semaglutide versus placebo calculated in the first step, along with the HR reported in the REWIND publication for dulaglutide versus placebo. 14

| Sensitivity analyses
Three sensitivity analyses were performed to test the robustness of the results. The motivation for these was to explore the impact TA B L E 1 Baseline characteristics of patients enrolled in SUSTAIN 6 and REWIND  to the REWIND trial. This was in accordance with the recommended approach in the NICE guidance, which proposes matching trials separately as a better approach than simply pooling the IPD and treating it as one large population in a single MAIC calculation. 21 As in the main analysis, this pooled HR could then be used along with the HR from REWIND to indirectly compare the semaglutide molecule with dulaglutide.
Another sensitivity analysis (sensitivity analysis 2), which used IPD from SUSTAIN 6 only as in the main analysis, explored the choice of potential effect modifiers with re-classification of eGFR and UACR data to balance the mean value and SD for eGFR as a continuous variable and the mean and SD of the log UACR. The log scale was chosen for UACR because only median and IQR were available from the published REWIND data, suggesting some skew in the data.
Consequently, an estimation of mean and SD was first calculated on the natural scale from the median and IQR using the methods proposed by Wan et al, 2014 28 and then the log values calculated for matching.
The final sensitivity analysis (sensitivity analysis 3) also used IPD from SUSTAIN 6 only and considered exclusion of baseline kidney function factors (eGFR and albuminuria) entirely. A comparison of the adjusted potential effect-modifying baseline patient characteristics in the SUSTAIN 6 and REWIND trials before and after matching is presented in Table 2. After matching, the characteristics were exactly balanced between the trials, with the effective sample sizes of the population in SUSTAIN 6, a measure of the patient overlap between trials, being approximately 20% smaller than the original trial data ( Table 2). Results of the matching of PIONEER 6 baseline data for the sensitivity analysis are presented in the Supporting Information.

MACE-results of the main analysis
In the main analysis, following the re-weighting of the observed 3P

| Sensitivity analyses results
In sensitivity analysis 1, when IPD from the SUSTAIN  In sensitivity analysis 2, which explored the impact of reclassifying eGFR and UACR data as continuous rather than dichotomous data, and sensitivity analysis 3, which excluded eGFR and albuminuria data entirely, the mean HR values for s.c. semaglutide versus placebo and versus dulaglutide were comparable with those estimated in the main analysis (Table 4). A forest plot showing all the relative treatment effects after matching IPD from semaglutide trials with REWIND aggregate data is presented in Figure 2.

| DISCUSS ION
This study compared the relative effect of two GLP-1 RAs (s.c. semaglutide vs. dulaglutide) on rates of 3P MACE (CV death, non-fatal MI, or non-fatal stroke) in patients with T2D with or without established CVD, using a MAIC approach.
The main analysis showed that, compared with placebo, s.c. semaglutide was associated with a statistically significant 35% reduction in 3P MACE in a population with a lower prevalence of pre-existing CVD than that enrolled in the SUSTAIN 6 trial. This adjusted reduction was greater than that demonstrated in the pre-specified primary analysis from SUSTAIN 6 in which s.c. semaglutide was associated with a 26% Baseline characteristics Abbreviations: DULA, dulaglutide; eGFR, estimated glomerular filtration rate; ESS, effective sample size; HF, heart failure; MI, myocardial infarction; NYHA, New York Heart Association; PAD, peripheral arterial disease; PBO, placebo; s.c., subcutaneous; SEMA, semaglutide; TIA, transient ischaemic attack; UACR, urinary albumin-to-creatinine ratio. a NYHA stage unclear at time the analysis was conducted; stage II-III was chosen for matching from SUSTAIN 6 as this was considered a more conservative approach, that is the difference in proportions of patients with prior HF between trials with I-III would have been even wider than the 17.9% versus 8.6%. b Includes 7 patients with >50% stenosis in peripheral arteries on angiography or imaging, but no prior diagnosis of PAD. c At baseline, compared with values at screening reported in (Marso et al. 13 ), n = 939.

TA B L E 2
Comparison of effectmodifying baseline patient characteristics from SUSTAIN 6 and REWIND trials before and after matching is that it uses a common comparator to balance prognostic variables within studies.
The main limitation of the current study is likely to be the selection of potential effect modifiers, which are not well established for CV outcomes in T2D. To overcome this, we selected potential effect modifiers based on clinical expertise and aimed to select those that would provide estimates of the effects of GLP-1 RAs across the spectrum of CV risk. However, the choice of potential effect modifiers is, to some extent, subjective and it is anticipated that choosing different variables to match may alter the findings of the analyses.
Furthermore, results could only be adjusted for reported aggregate data and corresponding data available from IPD records. Therefore, bias associated with differences in baseline characteristics between trials may remain. An additional limitation is the difference in follow-up periods between the studies, with a median follow-up time The findings of this study may help to improve understanding of the clinical differences between products within the GLP-1 RA treatment class for patients with T2D at high CV risk.

ACK N OWLED G EM ENTS
The authors would like to thank Søren Rasmussen, John Breddy and Thomas Hansen, from Novo Nordisk A/S, for providing support with statistical analyses, Andrei-Mircea Catarig from Novo Nordisk A/S for providing medical input to the analysis, and Sophie Doran from DRG Abacus (part of Clarivate), for providing medical writing support, which was funded by Novo Nordisk A/S.