Heart and kidney benefits of empagliflozin compared to GLP-1 receptor drugs: final year results from the EMPRISE study

The study population included patients with T2D aged 18 years or older (65 years or older in Medicare), who initiated empagliflozin or cardioprotective GLP-1RA (liraglutide, albiglutide, dulaglutide, injection semaglutide) from 8/2014 to 9/2019 (Additional file: sFigure). Exenatide and lixisenatide were not considered in the GLP-1RA group due to the lack of demonstrated cardiovascular benefits, and oral semaglutide was not yet approved during the timeframe of this study. Cohort entry was on the date of initiation of empagliflozin or a comparator without history of any SGLT2i or GLP-1RA prescriptions for 12 months, which was defined as the baseline period. We required patients to have continuous coverage for insurance plans and a recorded diagnosis of T2D during this baseline period. We excluded patients with recorded diagnoses of type 1 or secondary diabetes, malignancy, ESKD or kidney replacement therapy, human immunodeficiency virus, solid organ transplant, or a nursing home admission during baseline (Additional file: sTable). Patients who initiated both empagliflozin and GLP-1RA on the cohort entry date were excluded. 1 1 1 1

To identify study outcomes, patients were followed from one day after cohort entry until the earliest of: discontinuation of the index drug, switching to the comparator drug class, switching from an initial to an alternative agent within the same class, gap in insurance coverage (> 30 days), death, or end of the study (September 30, 2019). We considered patients to be on medications until 60 days after the end of the last prescription’s supply.

Primary outcomes were (i) a composite of myocardial infarction (MI) or stroke, (ii) hospitalization for heart failure (HHF), defined as an HF diagnosis in the primary discharge position, (iii) MACE, defined as a composite of MI, stroke, or cardiovascular mortality, and (iv) a composite of cardiovascular mortality or HHF. Secondary outcomes included the composite of all primary outcomes (MACE or HHF) (Medicare data only), the composite of MI, stroke, or HHF, HHF defined as diagnosis in all discharge positions (HHF-broad), individual components of MACE, unstable angina, coronary revascularization, all-cause mortality, and ESKD. To allow sufficient time for patients to progress to ESKD, we restricted the population to patients with chronic kidney disease (CKD) stages 3–4, defined using a validated claims-based algorithm [23]. We report detailed outcome definitions in Additional file 1: sTable 2.

We defined primary outcomes using validated claims-based algorithms, with high specificity (93–98%) and positive predictive value (PPV > 98%) [24–26]. Date of death was ascertained from the Vital Status files, with linkage to the Social Security Administration (SSA) data, which has been validated and captures > 95% of deaths in older adults aged > 65 years in the US [27, 28]. Cause of death was ascertained through linkage with the National Death Index data considering only diagnoses in the primary position, and was available only in Medicare data [29].

We identified 143 covariates a priori based on literature review and clinical knowledge: demographics, census region, calendar time of cohort entry, modified Charlson/Elixhauser combined comorbidity score [30], validated claims-based frailty index [31], diabetes complications, glucose-lowering medication use on cohort entry and during baseline, cardiovascular diseases, systemic comorbidities, chronic disease medications, and measures of healthcare utilization across different healthcare settings as a proxy for the general health status and the intensity of care. Patient characteristics were measured during the baseline period using administrative enrollment data, diagnosis or procedure codes, and pharmacy National Drug Codes. Data on laboratory results were available in Clinformatics® (~ 45%) and MarketScan (~ 5–10%). We provided the complete list of covariates in Additional file 1: sTable 3.

To improve comparability of the two treatment groups, we 1:1 matched patients initiating empagliflozin with those initiating GLP-1RA based on the estimated propensity score (PS). We estimated PS as the predicted probability of initiating empagliflozin relative to cardioprotective GLP-1RA, separately within each database, after controlling for 143 baseline characteristics using multivariable logistic regression [32]. Since laboratory test results were only available in a subset of the population, they were not included in the PS model. We matched using the nearest neighbor approach without replacement [33], and the maximum allowed difference (caliper) in PS between empagliflozin and GLP-1RA was 0.01 [33]. Balance in covariates, including that of laboratory results, was assessed using absolute standardized mean differences (SMD) (lower values indicate better balance) [34] and the post-matched c-statistic of the model predicting the exposure conditional on baseline covariates (values closer to 0.5 indicate better balance) [35].

To allow tight control of baseline CVD, and risk factors related to evolving treatment indications over time, we conducted PS estimation and matching separately within each baseline CVD subgroup [ASCVD or HF (yes/no)] and each calendar time block (before and after 2018), across a total of 4 strata, within each database. The year 2018 was chosen to better capture the period before and after the shift in treatment guideline recommendations for SGLT2i and GLP-1RA [36–38]. After matching, we pooled all the PS-matched databases, and estimated treatment effects in the final pooled database using stratified likelihood [39]. We estimated hazard ratios (HR) using Cox proportional hazards models, and rate differences (RD) using Mantel-Haenszel methods [39]. We present the cumulative risk of outcomes over the follow-up period using Cumulative Incidence Function (CIF) plots using the Kaplan-Meier method.

We estimated treatment effects within the following subgroups: (i) age (≥ vs. <65 years), (ii) sex (male vs. female), (iii) history of baseline ASCVD (defined as a diagnosis for any major ASCVD, including MI, angina, coronary atherosclerosis or other forms of chronic ischemic heart disease, coronary procedure, ischemic stroke, peripheral arterial disease or surgery, or lower extremity amputation), and (iv) baseline HF. Within each subgroup, the PS was re-estimated, and patients were re-matched on the newly estimated PS. The heterogeneity of estimates across the subgroups was evaluated using the Wald test for homogeneity [39].

We undertook several steps to mitigate the potential for unmeasured confounding. First, to reduce unmeasured confounding by kidney function, we restricted the study cohort to patients with at least two dispensed prescriptions for metformin (recommended first line therapy for patients without severely compromised kidney function) during the 6 months prior to cohort entry [40], and without any insulin prescriptions during the baseline period. Second, we restricted the study population to patients with laboratory results data available [i.e., patients with non-missing hemoglobin A1c (HbA1c) and estimated glomerular filtration rate (eGFR)] in Clinformatics® or Marketscan, and matching was re-performed using claims-based variables, HbA1c, and eGFR. Third, we performed 1:1 high-dimensional PS matching, which enriched the original PS with 100 additional empirically identified covariates, based on thousands of candidate covariates in different care settings [41]. The algorithm automatically selects covariates based on their confounding potential and has been shown to improve adjustment for unmeasured confounding [41]. Fourth, we conducted bias analyses in which we re-estimated treatment effects after adjusting for HbA1c or eGFR to check if our estimates were robust even under assumptions of extreme imbalance in these unmeasured confounders between treatment groups [42].

To account for potential informative censoring, we conducted further analyses: (i) intent-to-treat (ITT) analyses which do not censor for treatment discontinuation or switching allowing maximum follow-up of up to two years, and (ii) censoring-weighted analyses which create pseudo-populations in which treatment discontinuation/switching was independent of baseline covariates [39]. Other sensitivity analyses included addressing potential exposure misclassification by varying the exposure assessment window from 60 to 30 days before censoring for treatment discontinuation, and restricting analyses to patients with at least 1 and 2 years of follow-up to account for longer follow-up time necessary for the development of cardiorenal outcomes. In these analyses, follow-up started at 1- and 2-years post-index until the end of available follow-up. We also compared empagliflozin with each individual GLP-1RA agent (liraglutide or dulaglutide) after re-matching them and re-estimating the PS for each pair.

All analyses were performed using the Aetion Evidence Platform® (2023), a software for real-world data analysis validated for a range of studies (Aetion, Inc.) [43], with R version 4.2 (R Foundation for Statistical Analysis) and SAS 9.4 Statistical Software (SAS Institute Inc., Cary, NC).

After matching, rates of the composite of MI or stroke were similar between empagliflozin and GLP-1RA initiators [13.1 and 13.4 events per 1,000 person-years (PY) with corresponding HR of 0.99 (0.92, 1.07) and RD of -0.23 (-1.25, 0.79) per 1,000 PY]. Empagliflozin was associated with lower rates of HHF relative to GLP-1RA with rates of 5.0 and 7.3 per 1,000 PY respectively, corresponding to HR of 0.69 (0.62, 0.77) and RD of -2.28 (-2.98, -1.59).

In analyses restricted to Medicare, empagliflozin was associated with a slightly lower risk of MACE compared with GLP-1RA initiators, based on rates of 22.6 and 25.1 per 1,000 PY respectively, HR of 0.90 (0.82, 0.99), and RD of -2.54 (-4.76, -0.32) per 1,000 PY. Rates of the composite of cardiovascular mortality or HHF were 14.2 and 18.3 events per 1,000 PY among empagliflozin and GLP-1RA initiators, respectively [HR: 0.77 (0.69, 0.86), RD: -4.11 (-5.95, -2.29) per 1,000 PY]. In a subgroup of patients with history of baseline CKD stages 3–4 (10,837 PS-matched pairs), empagliflozin was associated with a lower risk of ESKD compared with GLP-1RA initiators [HR: 0.75 (0.60, 0.94), RD: -6.77 (-11.97, -1.61) per 1,000 PY] (Table 2).

Estimates for secondary outcomes were overall consistent, with similar risks of MI, stroke, unstable angina, and coronary revascularization between the groups, while the risks of composite of primary outcomes, HHF (defined more broadly), all-cause, and cardiovascular mortality were lower in empagliflozin vs. GLP-1RA initiators (Table 2). Empagliflozin was also associated with lower risks of the composite of MI, stroke, or HHF in all patients, and the composite of MACE or HHF in older Medicare patients relative to GLP-1RA.

Database-specific estimates were overall consistent except for small differences in commercial claims databases due to the small numbers of events (Additional file: sTable). 1 6

Consistent with HR and RD estimates, CIF curves showed similar risks of the composite of MI or stroke, and lower risk of MACE among patients initiating empagliflozin relative to GLP-1RA. The risks of HHF and the composite outcome of cardiovascular mortality or HHF were also lower in empagliflozin relative to GLP-1RA initiators. (Fig. 1).

On the relative scales, estimates for the composite outcome of MI or stroke and MACE outcomes were similar in patients with and without baseline history of ASCVD or HF. Relative risk reductions in HHF and the composite of HHF and cardiovascular mortality were consistently observed independently of baseline ASCVD and HF. For all outcomes, absolute RDs were larger in patients with baseline ASCVD or HF than in those without these conditions (Fig. 2).

Stratified analyses by age showed that the relative risk of the composite outcome of MI or stroke was lower in older vs. younger patients, while the relative risk reductions for HHF were similar across age categories. Estimates for the remaining outcomes did not differ by age subgroups. Stratified analyses by sex produced similar relative hazards between male and female patients across all outcomes. Absolute RDs were larger in older vs. younger patients (Fig. 3). Subgroup analyses for secondary outcomes provided similar findings (Additional file 1: sTable 7).

Sensitivity analyses were overall consistent with primary analytical findings (Additional file: sTable). Rematching populations using laboratory results, high-dimensional PS matching (Additional file: sTables–), and bias analyses (Additional file: sFigures,) supported the robustness of the primary findings to unmeasured confounding. Analyses restricted to patients with ≥ 1–2 years of follow-up and comparing empagliflozin vs. each individual GLP-1RA agent (liraglutide or dulaglutide) revealed similar findings (Additional file: sTables,). 1 8 1 9 10 1 3 4 1 11 12

Several limitations should be considered. First, we cannot exclude unmeasured confounding. After PS matching using laboratory results in a subset of the cohort with laboratory results available, findings remained consistent with the primary analyses. Our 1:1 PS matched design incorporated a rich set of claims-based variables that have been shown to balance laboratory results and clinical parameters typically only available in electronic health records [46]. Second, our outcome definitions relied on claims-based algorithms previously validated to have high specificity and PPV but low sensitivity [39]. Third, the median follow-up was short due to the lower persistence on treatments of patients in routine clinical practice compared to RCTs. Fourth, the primary analysis may suffer from informative censoring, which we addressed using ITT and censoring-weighted analyses. Finally, beneficial effects of GLP-1RA may require longer follow-up time to become apparent especially the effects mediated by atherosclerosis progression and weight loss. Analyses restricted to patients with ≥ 1–2 years of available follow-up provided consistent findings.

The institutional review board at the Brigham and Women’s Hospital approved the protocol. Data use agreements with the data vendors were in place. Informed consent was waived since the databases used were secondary administrative healthcare databases.

All authors have provided the consent for publication.

PTH previously worked at Johnson & Johnson on an unrelated work. HT has nothing to disclose. SS reports investigator-initiated grants to the Brigham and Women’s Hospital from Boehringer Ingelheim and owns equity in a software manufacturer, Aetion, Inc. DJW reports serving on data monitoring committees for Novo Nordisk and consulting for Elsevier and UpToDate, and grants from PCORI. BME reports consulting for, Janssen, Eli Lilly and Company, Provention Bio, Ipsen Pharmaceuticals, Novo Nordisk, NIDDK, American Heart Association, and UptoDate and grants from PCORI and Novo Nordisk outside the submitted work. RJG reports grants from Amgen, AstraZeneca, Kowa, Novartis, Pfizer outside the submitted work. LK is an employee of Eli Lilly and Company and owns stock in Eli Lilly and Company. NS and ADL are employees of Boehringer Ingelheim International GmbH. None for JMP. Dr. Patorno reported receiving grants from Boehringer Ingelheim during the conduct of the study and grants from the Food and Drug Administration and the Patient-Centered Outcomes Research Institute outside the submitted work. The authors did not receive any payment related to the development of the manuscript. Boehringer Ingelheim was given the opportunity to review the manuscript for medical and scientific accuracy as well as intellectual property considerations.

Due to the data user agreements with the Centers for Medicare and Medicaid Services, Optum, and Marketscan, we cannot share the analytical datasets. Databases used in the study are available to purchase from the data vendors.