Comparative Effectiveness of Second-line Antihyperglycemic Agents for Cardiovascular Outcomes: A Multinational, Federated Analysis of LEGEND-T2DM

Abstract

Background:

Sodium-glucose co-transporter-2 inhibitors (SGLT2is) and glucagon-like peptide-1 receptor agonist (GLP1-RAs) reduce major adverse cardiovascular events (MACE) in patients with type 2 diabetes mellitus (T2DM). However, their effectiveness relative to each other and other second-line antihyperglycemic agents is unknown, without any major ongoing head-to-head trials.

Objectives:

To compare the cardiovascular effectiveness of SGLT2is, GLP1-RAs, dipeptidyl peptidase-4 inhibitors (DPP4is), and sulfonylureas (SUs) as second-line antihyperglycemic agents in T2DM.

Methods:

Across the LEGEND-T2DM network, we included ten federated international data sources, spanning 1992–2021. We identified 1,492,855 patients with T2DM and cardiovascular disease (CVD) on metformin monotherapy who initiated one of four second-line agents (SGLT2is, GLP1-RAs, DPP4is, or SUs). We used large-scale propensity score models to conduct an active comparator, target trial emulation for pairwise comparisons. After evaluating empirical equipoise and population generalizability, we fit on-treatment Cox proportional hazard models for 3-point MACE (myocardial infarction, stroke, death) and 4-point MACE (3-point MACE + heart failure hospitalization) risk, and combined hazard ratio (HR) estimates using random-effects meta-analysis.

Results:

Over 5.2 million patient-years of follow-up and 489 million patient-days of time at-risk, there were 25,982 3-point MACE and 41,447 4-point MACE events. SGLT2is and GLP1-RAs were associated with a lower 3-point MACE risk than DPP4is (HR 0.89 [95% CI, 0.79–1.00] and 0.83 [0.70–0.98]), and SUs (HR 0.76 [0.65–0.89] and 0.72 [0.58–0.88]). DPP4is were associated with a lower 3-point MACE risk versus SUs (HR 0.87 [0.79–0.95]). The pattern was consistent for 4-point MACE for the comparisons above. There were no significant differences between SGLT2is and GLP1-RAs for 3-point or 4-point MACE (HR 1.06 [0.96–1.17] and 1.05 [0.97–1.13]).

Conclusions:

In patients with T2DM and CVD, we found comparable cardiovascular risk reduction with SGLT2is and GLP1-RAs, with both agents more effective than DPP4is, which in turn were more effective than SUs. These findings suggest that the use of GLP1-RAs and SGLT2is should be prioritized as second-line agents in those with established CVD.

CONDENSED ABSTRACT

In this largest study of 1.5 million patients with type 2 diabetes mellitus (T2DM) and cardiovascular disease (CVD) across ten international cohorts, SGLT2 inhibitors (SGLT2is) and GLP-1 receptor agonists (GLP1-RAs) were associated with an 11% and 17% lower risk of cardiovascular events compared with dipeptidyl peptidase-4 inhibitors (DPP4is), respectively. All three agents were associated with lower cardiovascular events than sulfonylureas, with 24% and 28% lower risk with SGLT2is and GLP1-RAs, and 13% lower risk with DPP4is. Our study defines strong empirical evidence in support of exclusive second-line use of SGLT2is and GLP1-RAs among those with T2DM and CVD.

INTRODUCTION

Over the past decade, the therapeutic options for type 2 diabetes mellitus (T2DM) have undergone a significant transformation.¹ Sodium-glucose co-transporter-2 inhibitors (SGLT2is) and glucagon-like peptide-1 receptor agonists (GLP1-RAs) have expanded the role of antihyperglycemic agents from managing high blood glucose to addressing the elevated cardiovascular risk in patients with T2DM.^2,3 In several large randomized clinical trials (RCTs), SGLT2is and GLP1-RAs reduced major adverse cardiovascular events (MACE), such as myocardial infarction (MI), hospitalization for heart failure (HF), and cardiovascular mortality.^4–11 While older antihyperglycemic agents, like sulfonylureas (SUs), have not undergone similarly comprehensive trials to evaluate their cardiovascular efficacy or safety, some studies suggested their neutral or even harmful effect on cardiovascular outcomes, with increased risk of hypoglycemia.^12–15 Furthermore, direct comparisons of SGLT2is and GLP1-RAs with dipeptidyl peptidase-4 inhibitors (DPP4is), which are antihyperglycemic agents with neutral effects on MACE, have not been conducted and no major trials are in progress. Nevertheless, DPP4is and SUs continue to be used in clinical practice and are recommended as second-line T2DM agents in national clinical practice guidelines.^16–19

Despite the availability of evidence from SGLT2i and GLP1-RA trials, several gaps in evidence challenge the development of treatment recommendations in T2DM.^16,20,21 Specifically, trials of SGLT2is and GLP1-RAs were not designed as head-to-head comparisons with older agents but rather as additive treatments on top of commonly used T2DM agents.^5,6,22 As a result, the relative cardiovascular efficacy of newer versus older agents is not known. Moreover, the comparative cardiovascular efficacy of SGLT2is and GLP1-RAs have not been evaluated. Thus far, comparative effectiveness assessments have been based on indirect estimates from clinical trials or comparative effectiveness drawn from a limited number of data sources.^23–25 The findings from observational studies using single data sources are challenged by our observations that the uptake of these agents, and therefore, the selective pressures on the use of these agents, varies considerably across sites,¹⁸ with the potential to introduce bias in the evaluation of comparative effectiveness. These evidence gaps pose a significant challenge in designing treatment algorithms that rely on the comparative effectiveness and safety of drugs.²¹ As a result, there is a large variation in clinical practice guidelines and clinical practice with regard to these medications, with many patients initiated on the newer therapies and many others treated with older regimens.^18,26

To address this, we conducted the Large-scale Evidence Generation and Evaluation across a Network of Databases for Type 2 Diabetes Mellitus (LEGEND-T2DM) initiative,²⁷ a large-scale, systematic, federated evaluation of 4.7 million patients with T2DM across multiple international observational data sources, where we leveraged robust state-of-the-art methodological and analytic strategies to minimize residual confounding, publication bias, and p-hacking. Here, we compared the cardiovascular effectiveness of four second-line antihyperglycemic agents — SGLT2is, GLP1-RAs, DPP4is, or SUs — when initiated on the background of metformin therapy in T2DM.

METHODS

Data Sources

In this study, we included ten real-world data sources from the LEGEND-T2DM network, including six administrative claims and four EHR databases across four countries during 1992–2021 (Central Illustration). These represent data from six national-level and four health-system datasets from the US, as well as data sources from Germany, Spain, and the United Kingdom. The vast majority of patient records span from the mid-2000s to today, covering two decades of T2DM treatment as well as the introduction of many second-line antihyperglycemic agents. All LEGEND-T2DM data sources were previously standardized to the Observational Health Data Sciences and Informatics (OHDSI)’s Observational Medical Outcomes Partnership (OMOP) common data model (CDM) version 5,²⁸ which mapped international coding systems into standard vocabulary concepts. The use of the OMOP CDM allows a federated analysis of data, without patient-level data sharing, using consistent cohort definitions and study design. All data partners received institutional approval or exemption for their participation. Details of data sources are presented in eTables 1–2.

Central Illustration: Study Design and Analytical Methodology to Evaluate the Comparative Effectiveness of SGLT2is, GLP1-RAs, DPP4is, and SUs for Cardiovascular Outcomes.

We leveraged ten federated data sources from four different countries comprising more than 4.7 million individuals with T2DM on metformin monotherapy initiating one of the four second-line agents. We included a subset of 1.5 million patients with established CVD to assess the comparative cardiovascular effectiveness of these agents. We mapped the data sources to the OMOP CDM and leveraged robust state-of-the-art methodological and analytic strategies to minimize residual confounding, publication bias, and p-hacking. Abbreviations: CCAE, IBM MarketScan^® Commercial Claims and Encounters Data; CVD, cardiovascular disease; GDA, Germany Disease Analyser; IMRD, UK-IQVIA Medical Research Data; MACE, major adverse cardiovascular events; MDCD, IBM Health MarketScan^® Multi-State Medicaid Database; MDCR, IBM Health MarketScan^® Medicare Supplemental and Coordination of Benefits Database; OCEDM, Optum© Clinformatics Extended Data Mart - Date of Death; OEHR, Optum© de-identified Electronic Health Record Dataset; OMOP-CDM, Observational Medical Outcomes Partnership-common data model; SIDIAP, Information System for Research in Primary Care; T2DM, type 2 diabetes mellitus; USOC, United States Open Claims; VA, Department of Veterans Affairs Healthcare System.

Study Design

We identified patients with T2DM and established cardiovascular disease (CVD) on metformin monotherapy who initiated on any of the drug ingredients within one of the SGLT2i, GLP1-RA, DPP4i, or SU drug classes (eMethods, eTable 3).¹⁸ Each exposure cohort thus consists of new users of each drug class. We require patients with T2DM and CVD to have at least one year of prior observation in the database, with at least 3 months of prior metformin use before initiating a second-line agent, and no prior exposure to a comparator second-line or other antihyperglycemic agent or more than 30 days of insulin exposure (eMethods). To ensure statistical power, we executed analyses for comparisons and data sources with at least 1,000 patients in each arm. To evaluate the comparative effectiveness of antihyperglycemic agents, we constructed target–comparator–database combinations, where we compared one antihyperglycemic agent (target) with another agent (comparator) across data sources. For this, we employed a new-user cohort design for target and comparator agents within each data source in order to emulate the hypothetical target trial.^29,30 Methodological principles in the study design have been carefully constructed based on the evidence by experts and were leveraged previously to minimize bias and improve reproducibility.^27,31

Study Outcomes

Across all data sources and pairwise exposure cohorts, we assess the relative risks of two primary and four secondary cardiovascular outcomes. The two primary outcomes of interest are (1) 3-point MACE, including acute MI, stroke, and sudden cardiac death (SCD), and (2) 4-point MACE, which additionally includes hospitalization for HF. Secondary outcomes of interest include the four individual MACE components. We construct outcome cohorts based on previously developed phenotypes validated and tested in prior work (eMethods).^31–36

For each outcome cohort, we included patients with no events prior to treatment initiation and defined continuous drug exposure as consecutive drug prescriptions with less than 30-day prescription gaps. We considered an on-treatment time-at-risk (TAR) definition that follows a patient from treatment initiation to treatment discontinuation, which captures direct treatment effects while allowing for escalation with additional T2DM agents.

Statistical Analysis

We employed a systematic federated analytic framework to address residual confounding, publication bias, and p-hacking.^27,37 This framework uses data-driven, large-scale propensity score (PS) adjustment for measured confounding,³⁸ a large set of negative control outcome experiments to address unmeasured confounding and systematic bias,^39–41 study diagnostics to ensure validity and generalizability,^41,42 and a principled meta-analysis approach to aggregate evidence across data sources. We used standardized vocabularies to construct consistent computable definitions of all study cohorts, covariates, and outcomes. We provide full disclosure of all hypotheses investigated and pre-specify and report all analytical procedures in the published protocol of LEGEND-T2DM.²⁷ To promote open science and avoid publication bias, we have disseminated all results in a publicly available R ShinyApp,⁴³ and all analytic code is publicly available on GitHub.⁴⁴

Within each data source, we estimated the relative risks of all six outcomes between each pair of new-user cohorts, taking one exposure cohort as the target and the other as a comparator. For each pairwise comparison and each data source, we adjusted for measured confounding and improved balance between cohorts by matching and stratifying on PS.⁴⁵ We estimated the PS by a data-driven approach that adjusts for a broad range of predefined baseline patient characteristics through regularized regression.³⁸ These characteristics included demographics, comorbidities, concomitant medication use, and healthcare utilization in the period before the initiation of the second-line antihyperglycemic agent. The choice of PS stratification vs. matching was based on the approach that achieves a standardized mean difference (SMD) <0.15 across all covariates.⁴⁶ When both stratification and matching provide sufficient balance, we preferred stratification over matching and thus reported results based on stratification when available, as the former improves patient inclusion and, therefore, generalizability. We then used Cox proportional hazards models to estimate hazard ratios (HRs) of each outcome for each comparison, conditional on PS stratification or variable-ratio patient matching.

We also sought to address residual bias, which may persist in observational studies even after PS adjustment that controls for measured confounding.^39,40 For this, we conducted negative control (falsification) outcome experiments for each comparison and outcome, where the null hypothesis of no differential effect (i.e., HR = 1) is believed to be true for each outcome. We selected 100 negative controls through a data-driven algorithm that identifies OMOP condition concept occurrences with similar prevalence to the outcomes of interest that lack evidence of association with exposures in published literature, drug–product labels, and spontaneous reports, which we then confirmed by expert review (eMethods).⁴⁷ We list these negative controls in eTable 4. From these negative control experiments, we learn an empirical null distribution that informs residual study bias, i.e., a deviation from the empirical null across all outcomes represents a quantitative surrogate for the residual bias. We calibrate each original HR estimate to compute a calibrated HR estimate and 95% confidence interval (CI).³⁹ We declare an HR as significantly different from the null if the calibrated p-value is <0.05 without considering multiple testing corrections.

We assessed, while blinded to the results, study diagnostics to ensure reliability and generalizability for all comparisons and only report estimates that pass the diagnostics.^41,42 These study diagnostics included (1) minimum detectable risk ratio (MDRR) as a metric for statistical power, (2) preference score distributions between the target and comparator cohorts to evaluate empirical equipoise and population generalizability, and (3) cohort balance before and after propensity score (PS) adjustment, defined by the absolute SMDs on extensive patient characteristics. A study passed diagnostics if MDRR was less than 4, and >25% of patients had a preference score between 0.3 and 0.7 on both arms and maximum SMD <0.15 after PS adjustment. Additional diagnostics for visual examination included (4) calibration plots on negative control outcomes to examine residual bias, and (5) Kaplan-Meier plots, all publicly available on the R ShinyApp,⁴³ to visually inspect proportionality assumptions for the Cox models.

We reported all HR estimates, their 95% CIs, and p-values post-calibration for studies that passed diagnostics. To aggregate evidence across non-overlapping data sources, we combined all calibrated HR estimates for each comparison using a random effects meta-analysis approach.^49,50 Further, for each comparison, we performed leave-one-out meta-analyses by calculating the calibrated risk estimates while systematically removing one data source at a time.

RESULTS

Cohort Characteristics

Across ten federated longitudinal data sources from four countries, we identified 4,730,887 patients with T2DM. This included 1,492,855 patients with T2DM and established CVD on metformin monotherapy who initiated one of the four second-line antihyperglycemic agents and had no prior use of any other antihyperglycemic agents (Central Illustration). Among these patients, 244,694 (16.4%) initiated SGLT2is, 123,991 (8.3%) GLP1-RAs, 413,236 (27.7%) DPP4is, and 710,934 (47.6%) SUs (Table 1). US Open Claims (USOC) with 1,163,413 patients with T2DM and established CVD, followed by Optum Clinformatics Extended Data Mart - Date of Death (OCEDM) (N=80,159) and Optum de-identified Electronic Health Record Dataset (OEHR) (N=66,417) contributed the most patients to the study population. Median on-treatment time-at-risk for patients varied by drug class and database between 2.0 and 23.2 months. At least 25% of the patients were exposed to their first drug class for more than 12 months across a majority of databases (Table 1).

Table 1:

Population Size and Follow-up Time for Initiators of SGLT2is, GLP1-RAs, DPP4is, and SUs Across Databases.

Drug/Data Source	Patients, No.	On-treatment Follow-up Time (days), Median [IQR]
SGLT2i
CCAE	10,135	161 [62 – 537]
GDA	4,337	181 [97 – 537]
MDCD	1,314	114 [30 – 383]
MDCR	2,013	112 [34 – 541]
OCEDM	11,597	118 [55 – 423]
OEHR	8,996	71 [29 – 181]
SIDIAP	2,281	322 [103 – 895]
USOC	196,710	143 [61 – 525]
VA	7,311	135 [71 – 443]
Total	244,694
GLP1-RA
CCAE	7,446	116 [45 – 441]
MDCR	1,120	94 [29 – 443]
OCEDM	6,546	98 [36 – 368]
OEHR	4,525	60 [29 – 173]
USOC	104,354	112 [52 – 432]
Total	123,991
DPP4i
CCAE	22,801	159 [63 – 597]
GDA	10,428	221 [97 – 741]
IMRD	4,024	354 [122 – 1,174]
MDCD	4,449	142 [56 – 555]
MDCR	10,541	187 [81 – 713]
OCEDM	23,918	150 [59 – 569]
OEHR	18,145	95 [89 – 312]
SIDIAP	8,126	695 [204 – 1,916]
USOC	299,835	157 [59 – 657]
VA	10,969	220 [90 – 707]
Total	413,236
SU
CCAE	18,093	148 [59 – 559]
GDA	1,235	216 [119 – 684]
IMRD	846	350 [106 – 1312]
MDCD	6,758	124 [33 – 503]
MDCR	8,887	159 [59 – 632]
OCEDM	38,098	177 [85 – 652]
OEHR	34,751	110 [89 – 386]
SIDIAP	619	600 [158 – 1,809]
USOC	562,514	183 [89 – 752]
VA	39,133	186 [90 – 689]
Total	710,934

Abbreviations: CCAE, IBM MarketScan^® Commercial Claims and Encounters Data; DPP4i, dipeptidyl peptidase 4 inhibitor; GDA, Germany Disease Analyzer; GLP1-RA, glucagon-like peptide-1 receptor agonist; IMRD, UK-IQVIA Medical Research Data; MDCD, IBM Health MarketScan^® Multi-State Medicaid Database; IQR, interquartile range; MDCR, IBM Health MarketScan^® Medicare Supplemental and Coordination of Benefits Database; OCEDM, Optum© Clinformatics Extended Data Mart - Date of Death; OEHR, Optum© de-identified Electronic Health Record Dataset; SGLT2i, sodium-glucose co-transporter-2 inhibitor; SIDIAP, Information System for Research in Primary Care; SU, Sulfonylurea; USOC, United States Open Claims; VA, Department of Veterans Affairs Healthcare System.

Addressing Confounding Across Target-Comparator-Database Combinations

PS adjustment achieved pre-specified covariate balance in patient baseline characteristics for pairwise class comparisons across all databases (Figure 1, eTables 5–47). The maximum SMD across target–comparator–database combinations consistently decreased after PS stratification (Figure 1). For example, in comparison of patients initiating SGLT2is (target) with patients initiating DPP4is (comparator) in the IBM MarketScan Commercial Claims and Encounters (CCAE) database, before PS adjustment, patients initiating SGLT2is were more frequently men and had obesity and HF relative to patients initiating DPP4is. However, after PS adjustment, the SGLT2is or DPP4is populations were well balanced on all demographic and clinical patient characteristics (eTable 10).

Figure 1: Maximum Standardized Mean Difference Before and After Propensity Score Stratification for All Covariates Across Target-Comparator-Database Combinations.

This figure shows to what extent the propensity score stratification reduced the maximum SMD for all covariates across six target-comparator pairs and ten data sources. Abbreviations: CCAE, IBM MarketScan^® Commercial Claims and Encounters Data; DPP4i, dipeptidyl peptidase 4 inhibitor; GDA, Germany Disease Analyzer; GLP1-RA, glucagon-like peptide-1 receptor agonist; IMRD, UK-IQVIA Medical Research Data; MDCD, IBM Health MarketScan^® Multi-State Medicaid Database; IQR, interquartile range; MDCR, IBM Health MarketScan^® Medicare Supplemental and Coordination of Benefits Database; OCEDM, Optum© Clinformatics Extended Data Mart - Date of Death; OEHR, Optum© de-identified Electronic Health Record Dataset; PS, propensity score; SGLT2i, sodium-glucose co-transporter-2 inhibitor; SIDIAP, Information System for Research in Primary Care; SMD, standardized mean difference; SU, Sulfonylurea; USOC, United States Open Claims; VA, Department of Veterans Affairs Healthcare System.

Empirical Equipoise Across Target-Comparator-Database Combinations

For most data sources, all executed class comparisons were in empirical equipoise (>25% of patients had preference scores between 0.3 and 0.7 on both arms) (eFigures 1–10). GDA, SIDIAP, and VA databases showed less equipoise for comparisons involving SGLT2is and DPP4is. However, in general, PS adjustment achieved sufficient covariate balance in terms of preference score distribution to reduce concerns that measured the estimated effects of baseline confounding biases (eFigures 11–30). Furthermore, the study had limited residual systematic error, where calibration of effect estimates using negative control outcomes resulted in an increase in the proportion of nominal 95% CIs that included 1 for control outcomes across a majority of comparisons and databases (eFigures 31–50). For example, for SGLT2is vs DPP4is comparison in SIDIAP using PS matching, before calibration, the nominal CIs covered 75.0% of control estimates; after calibration, they covered 91.7% (eFigure 48).

Comparative Effectiveness for Primary Endpoints

During 1,337,809 patient-years of follow-up, there were 25,982 3-point MACE and 41,447 4-point MACE events. SGLT2is and GLP1-RAs were associated with a lower hazard of 3-point MACE compared with DPP4is (HR 0.89 [95% CI, 0.79–1.00] and 0.83 [0.70–0.98], respectively), and with even lower HRs of 0.76 (95% CI, 0.65–0.89) and 0.72 (95% CI, 0.58–0.88) versus SUs, respectively. DPP4is were associated with a lower risk of 3-point MACE risk versus SUs (0.87 [0.79–0.95]). A consistent pattern was observed for 4-point MACE across the above-mentioned comparisons. In direct comparisons of SGLT2is and GLP1-RAs, there were no differences in either 3-point or 4-point MACE (1.06 [0.96–1.17], and 1.05 [0.97–1.13], respectively) (eTables 48–60, Figure 2).

Figure 2: Meta-analytic Calibrated Hazard Ratio Estimates for Comparative Effectiveness of SGLT2is, GLP1-RAs, DPP4is, and SUs for Cardiovascular Outcomes.

The figure shows the comparative effectiveness of SGLT2is, GLP1-RAs, DPP4is, and SUs for composite and individual cardiovascular outcomes across six target-comparator pairs with the y-axis representing target and the x-axis showing comparators. Points on the left side of the dashed line represent HR that favors the target (vertical agent) and points on the right side of the dashed line represent HR that favors the comparator (horizontal agent). Abbreviations: DPP4i, dipeptidyl peptidase 4 inhibitor; GLP1-RA, glucagon-like peptide-1 receptor agonist; HF, heart failure; HR, hazard ratio; MACE, major adverse cardiovascular events; MI, myocardial infarction; SGLT2i, sodium-glucose co-transporter-2 inhibitor; SU, Sulfonylurea.

Comparative Effectiveness for Secondary Endpoints

Acute MI

There were 13,536 episodes of acute MI across data sources during the follow-up. SGLT2is, GLP1-RAs, and DPP4is were associated with a lower hazard of acute MI compared with SUs (0.81 [0.69–0.95], 0.70 [0.56–0.87], and 0.86 [0.77–0.95], respectively). While GLP1-RAs were associated with a lower risk of acute MI compared with DPP4is (0.85 [0.73–0.99]), the hazard of acute MI was comparable for SGLT2is vs DPP4is (0.95 [0.83–1.08]). Compared with GLP1-RAs, SGLT2is were associated with a higher hazard of acute MI (1.19 [1.05–1.35]) (eTable 48, eTables 61–66, 122–127, Figure 2).

Stroke

There were 13,999 episodes of stroke events across data sources during the follow-up. SGLT2is, GLP1-RAs, and DPP4is were associated with a lower hazard of stroke compared with SUs (0.77 [0.65–0.92], 0.79 [0.63–0.98], and 0.90 [0.81–0.99], respectively). Compared with DPP4is, SGLT2is and GLP1-RAs had a comparable hazard of stroke (0.89 [0.74–1.07] and 0.89 [0.77–1.03], respectively). There was no significant difference between SGLT2is and GLP1-RAs (0.93 [0.82–1.05]) regarding the hazard of stroke (eTable 48, eTables 67–72, Figure 2).

SCD

Over follow-up, 3,789 SCD occurred across data sources. SGLT2is and GLP1-RAs were associated with a lower hazard of SCD compared with DPP4is (0.62 (0.51–0.75), and 0.65 [0.44–0.96], respectively) and SUs (0.60 [0.48–0.73], and 0.57 [0.42–0.78], respectively). DPP4is versus SUs and SGLT2is versus GLP1-RAs had comparable hazards of SCD (0.96 [0.81–1.13], and 1.11 [0.84–1.45], respectively) (eTable 48, eTables 73–78, Figure 2).

HF Hospitalization

During the follow-up, 30,743 HF hospitalizations were recorded across all databases. SGLT2is and GLP1-RAs were associated with a lower risk of HF hospitalization compared with DPP4is (0.80 [0.68–0.95] and 0.76 [0.67–0.86], respectively) and SUs (0.68 [0.58–0.80] and 0.64 [0.52–0.78], respectively). DPP4is were associated with a lower hazard of HF hospitalization compared with SUs (0.90 [0.82–0.99]). Compared with GLP1-RAs, SGLT2is had a comparable hazard for HF hospitalization (1.02 [0.93–1.12]) (eTable 48, eTables 79–84, Figure 2).

Sensitivity Analyses

Across all outcomes, 98.4% (690/701) of calibrated relative risk estimates remained within the confidence intervals of the primary meta-analysis when systematically removing each data source from the leave-one-out meta-analysis. Within the primary endpoints of 3-point MACE and 4-point MACE and secondary endpoints of acute MI and hospitalization for HF, none of the leave-one-out analyses were outside of the confidence intervals of the meta-analysis (Figure 3, eFigures 51–56).

Figure 3: Swarm Plot of Calibrated Hazard Ratio Estimates of Major Outcome Meta-analysis and Leave-one-out Meta-analysis.

Circles depict the calibrated relative risk of each leave-one-out study in which one original data source is removed from the meta-analysis. Diamonds depict the original meta-analysis effect estimate with all data sources. Points are color-coded by major outcomes, the y-axis represents six target-comparator pairs, and the x-axis on log-scale measures the calibrated relative risk of each outcome.

Abbreviations: DPP4i, dipeptidyl peptidase 4 inhibitor; GLP1-RA, glucagon-like peptide-1 receptor agonist; HF, heart failure; MACE, major adverse cardiovascular events; MI, myocardial infarction; SGLT2i, sodium-glucose co-transporter-2 inhibitor; SU, Sulfonylurea.

DISCUSSION

In this largest, multinational, federated study of 1.5 million patients with T2DM and established CVD across ten cohorts initiating a second-line antihyperglycemic agent after metformin monotherapy, SGLT2is and GLP1-RAs were associated with an 11% and 17% lower risk of cardiovascular events compared with DPP4is, respectively, the class of drugs approved before the US FDA-mandated cardiovascular outcomes trials for antihyperglycemic agents. All three agents were associated with lower cardiovascular events than SUs, with 24% and 28% lower risk with SGLT2is and GLP1-RAs and 13% lower risk with DPP4is. These patterns were consistent across both ischemic events, such as MI and stroke, as well as for HF. Between SGLT2is and GLP1-RAs, there was no difference in composite cardiovascular effectiveness outcomes – both 3-point and 4-point MACE, but there was a lower risk of acute MI with GLP1-RAs compared with SGLT2is. The study accounted for multiple null control outcomes and calibrated all estimates for residual bias on null control outcomes, with multiple sensitivity analyses confirming the findings of the primary analysis.

This builds on the evidence of comparative cardiovascular effectiveness of currently used antihyperglycemic therapies, especially in the sparse landscape of head-to-head RCTs of any of these agents. The CAROLINA trial is the only RCT that pursued a head-to-head evaluation of the cardiovascular safety of a DDP4i (linagliptin) and an SU (glimepiride) and found no difference between the two drugs.¹⁴ In contrast, we found a lower risk of cardiovascular adverse outcomes with DPP4i use compared with SU. Class-level comparisons and assessments across large and diverse populations likely contribute to these observed differences. Moreover, this does not suggest an absolute benefit from DDP4i but may even reflect a relative harm from SU, which has been suggested in prior studies.^52–56 There are no ongoing or future trials expected to fill this knowledge gap. Moreover, a few observational studies have used single data sources to evaluate the cardiovascular outcomes associated with these drugs.⁵⁷

Some associations observed in the present study, such as those with DPP4i, contrast with others where the smaller study population may have necessitated studying a combination of drugs directly against SU but challenges isolating an effect.¹⁸ These analyses are limited by local patterns of treatment use and challenges with exposure and outcome ascertainment in individual sources that may confound the observed association of these agents with outcomes, further supporting the need for comparative effectiveness evaluations at scale. Our recent work shows that the use of antihyperglycemic agents varies substantially in populations across the world.¹⁸ In the present study, we developed the target trial for head-to-head of these agents, focusing explicitly on the same disease stage, i.e. when escalation to a second-line agent is being done. We conducted the same study across ten different data sources, ensuring that the effects of the intervention were not limited to the selective pressures or outcome differences in a single population. We ensured the empirical equipoise by comparing SMDs before and after PS stratification. Finally, we ensured that unmeasured confounding, potentially manifesting as directional effects across several negative control outcomes, were used to calibrate the PS stratified effect estimates.

The findings have important implications for clinical practice. Our observations support the recent change in clinical practice guidelines recommending first-line use of SGLT2is and GLP1-RAs among individuals with established CVD.^59–61 However, our recent work shows a substantial proportion of patients with T2DM are often initiated on DPP4is and SUs despite their limited cardiovascular effectiveness or known superiority on other glycemic or other diabetes-related outcomes.^59–61 There is also evidence that suggests a potentially larger reduction in risk of acute MI with GLP1-RAs compared with SGLT2is, which were not replicated for stroke. We found no difference between SGLT2is, GLP1-RAs, and DPP4is in terms of stroke outcome. However, SGLT2is, GLP1-RAs, and DPP4is were all associated with a lower risk of stroke compared with SUs. In the absence of head-to-head comparisons of these agents for the stroke outcome, our findings are in line with the collective evidence from cardiovascular outcome trials, wherein these agents were compared against placebo.^15,62

Moreover, we did not observe an exclusive role of decreased HF risk with SGLT2i, as suggested in current guidelines, as there was no observed difference in HF risk among those receiving SGLT2i or GLP1-RA. The original recommendations for HF were based on inference from RCTs of individual agents and emerging evidence. In previous trials of GLP1-RAs, there was a lack of protocolized assessment of left ventricular ejection fraction, natriuretic peptides, and HF status, which limits inference on the effect on HF outcomes. However, the STEP-HFpEF trial, which carefully selected participants with confirmed HFpEF at baseline and closely monitored HF outcomes, demonstrated improvements in HF outcomes with semaglutide. Moreover, there was no interaction between the recorded ejection fraction and HF outcomes in STEP-HFpEF, suggesting the potential broader role of GLP1-RAs on HF-related adverse events..^65,66 Our study favors the continued evaluation of GLP1-RA as a therapy to reduce HF risk and improve outcomes of those with HF.^67–69

There are limitations that merit consideration. First, to enable the ability to capture information across data sources spanning administrative and EHR data sources necessitated the use of administrative definitions of both exposures and outcomes. However, to ensure the information is accurately captured, we used definitions that have been validated in prior studies.^70,71 Second, we expect the identification of the exposure and outcomes to vary across data sources. We conducted a random effects meta-analysis to account for this expected variability. Findings of the leave-one-out sensitivity analyses further ensured the consistency of the effect estimates across data sources. Although we did not explicitly report heterogeneity statistics, the consistency of effect estimates indirectly supports the assumption that our findings are not overly sensitive to any specific data source and further confirms the validity of the observations.

Third, we adopted an alpha criterion of 0.05 without adjusting for multiple testing. Of note, the study did not seek to test any key hypotheses, especially with the paucity of any quality evidence in this clinically important domain. Instead, we focused on rigorous methods and thoughtful interpretation that qualitatively assessed the directional consistency of effects across outcomes. Furthermore, we reported 95% CI of the effect estimates and did not report any p-values to draw attention to the effect size rather than significance testing. Fourth, capturing outcomes for EHR-based datasets, such as for recurrent hospitalizations and deaths, differs from administrative sources that track events across institutions and capture out-of-hospital death events.^16,17 However, we find that sources that capture cross-institution out-of-hospital information did not differ and had similar outcomes as those that focused on institutionally captured death events, further supporting the robustness of the drug association with outcomes. Fifth, the study does not explicitly address the potential for bias introduced by adherence differences between agents. Nevertheless, pooling of data across multiple data sources with expected variation in adherence would likely address the issues with differential adherence in a specific group. Moreover, the consistency of effects in analyses that excluded individual datasets also is reassuring. Sixth, the relatively short time-at-risk for antihyperglycemic agents may limit inferences on long-term outcomes. The shorter follow-up time is a function of the rigorous on-treatment design of the study, wherein patients exit the cohort in case of treatment discontinuation or escalation, which is essential for minimizing the misattribution of treatment effects. Moreover, while a shorter follow-up with a fewer number of events can bias the effect estimates toward the null, we found differential comparative effectiveness for the included agents.

Seventh, inference on all outcomes can still be subject to residual confounding. However, our use of calibrated estimates explicitly addresses the identification of residual confounding across a series of null control outcomes. Eighth, the study did not account for the competing risk of death. This was an a priori decision given the population not presumably at marked total mortality risk and the computational efficiency of Cox proportional hazards models. The consistency of the composite MACE that included SCD data with the individual outcomes suggests that mortality effects did not meaningfully differ across groups. Finally, given the evolving indications for these drug classes for the management of T2DM, it may be essential to evaluate their comparative effectiveness as first-line agents without metformin background therapy. Additionally, while the study’s scope was limited to class-level analyses, there is an unmet need for further evidence generation on drug-level comparative effectiveness. These evaluations represent the goals of the group’s future work. Furthermore, future comparative effectiveness studies could incorporate outcomes such as hemoglobin A1c and weight to better understand potential mediating effects.

CONCLUSIONS

Our large, rigorous multinational network study of patients with T2DM and established CVD found cardiovascular risk reduction with SGLT2is and GLP1-RAs, with both agents more effective than DPP4is, which in turn were more effective than SUs. These findings suggest that the use of GLP1-RAs and SGLT2is should be prioritized as second-line agents in those with established CVD.

PERSPECTIVES.

Competency in Medical Knowledge

The Large-scale Evidence Generation and Evaluation across a Network of Databases for Type 2 Diabetes Mellitus (LEGEND-T2DM) initiative is a large-scale, systematic, federated evaluation of individuals with T2DM across multiple international observational data sources, where we leveraged robust state-of-the-art methodological and analytic strategies to minimize residual confounding, publication bias, and p-hacking. In this largest study of 1·5 million patients with T2DM and established cardiovascular disease (CVD) across ten cohorts from four countries initiating a second-line antihyperglycemic agent — sodium-glucose co-transporter-2 inhibitors (SGLT2is), glucagon-like peptide-1 receptor agonists (GLP1-RAs), dipeptidyl peptidase-4 inhibitors (DPP4is), or sulfonylureas (SUs) — after metformin monotherapy, SGLT2is and GLP1-RAs were associated with an 11% and 17% lower risk of cardiovascular events compared with DPP4is, respectively. All three agents were associated with lower cardiovascular events than SUs, with 24% and 28% lower risk with SGLT2is and GLP1-RAs, and 12% lower risk with DPP4is. These patterns were consistent across both ischemic events, such as myocardial infarction and stroke, as well as for HF.

Competency in Patient Care

Our study defines strong empirical evidence in support of contemporary clinical practice guidelines that suggest SGLT2is and GLP1-RAs are used for those with T2DM and CVD. Our observations also provide support for exclusive second-line use of SGLT2is and GLP1-RAs among those with CVD.

ABBREVIATIONS

CVD

cardiovascular disease

DPP4i

dipeptidyl peptidase-4 inhibitor

GLP1-RA

glucagon-like peptide-1 receptor agonist

HF

heart failure

MACE

major adverse cardiovascular events

MI

myocardial infarction

RCT

randomized clinical trial

SGLT2i

sodium-glucose co-transporter-2 inhibitor

SU

sulfonylurea

T2DM

type 2 diabetes mellitus