GLP-1RA Use and the Risk of Heart Failure in Patients with Diabetes: Association or Causation?

Glucagon-like peptide-1 receptor agonists (GLP-1RAs) have transformed the care of patients with cardiovascular-kidney-metabolic syndrome. Specifically, among adults with type 2 diabetes, the newer GLP-1RAs such as semaglutide and tirzepatide improve glycemic and blood pressure control, delay progression of chronic kidney disease, and reduce the risk of major adverse cardiovascular events such as myocardial infarction and stroke.¹ As a result, these agents are now recommended as first-line therapies in certain patients with diabetes who have or are at high risk of atherosclerotic cardiovascular disease.²

While the protective effect of GLP-1RA therapies on atherosclerotic cardiovascular disease is now well established, the effect on heart failure outcomes is uncertain. A recent systematic review of randomized controlled trials found a 13% reduction in the risk of hospitalization for heart failure in patients with diabetes treated with GLP-1RAs compared with placebo (hazard ratio, 0.87 [95% confidence interval, 0.81 to 0.95]),³ while an older systematic review that included smaller trials found no clear benefit (hazard ratio, 0.94 [95% confidence interval, 0.85 to 1.03]).⁴ Two recent randomized trials in patients with heart failure and preserved ejection fraction (HFpEF) concluded that treatment with GLP-1RAs improves symptoms, functional status, and clinical outcomes.^5,6 However, the effects of GLP-1RAs on heart failure outcomes in individuals without heart failure at baseline, as well as those with heart failure with reduced ejection fraction (HFrEF) are not fully understood. This is a critical knowledge gap given the substantial burden of heart failure among individuals with diabetes.

In this issue of Circulation,⁷ Xu and colleagues investigate the relationship between GLP-1RA therapy and heart failure outcomes among more than 30,000 adults with type 2 diabetes. To do so, they applied the target trial framework using data from the Stockholm CREAtinine Measurements (SCREAM) project, a population-based longitudinal cohort of more than 3 million residents of the Stockholm region, Sweden.⁸ Using unique personal identity numbers, SCREAM data were linked to regional and national administrative databases containing information on demographics, socioeconomic status, health care utilization, laboratory tests, dispensed medications, diagnoses, and vital status.⁸ They emulated two trials: one comparing GLP-1RAs with dipeptidyl peptidase-4 inhibitors (DPP4i), and another comparing GLP-1RAs with sodium–glucose cotransporter-2 inhibitor (SGLT2i). Median follow-up was 26 months in the GLP-1RA arms and 21–36months in the control arms. The first trial emulation found that initiation of GLP-1RA therapy was associated with a significantly lower risk of hospitalization for heart failure compared with initiation of a DPP4i (12.1 vs 15.8 events per 1000 person years; hazard ratio, 0.77 [95% confidence interval, 0.66 to 0.91]). The second trial emulation found that the risk of hospitalization for heart failure was similar with GLP-1RA therapy compared with SGLT2i therapy (12.8 vs 12.6 events per 1000 person-years; hazard ratio, 1.02 [95% confidence interval, 0.85 to 1.18]). In a “per protocol” analysis that required trial participants to persist with the assigned treatment rather than simply initiating it, the risk differences were consistent in direction but larger in magnitude in the trial comparing GLP-1RAs with DPP4is (hazard ratio, 0.62 [95% confidence interval, 0.50 to 0.75]).

The authors conclude that, to the extent that the observed associations could be interpreted causally, GLP-1RA therapy could be considered for the reduction of heart failure hospitalization in patients with type 2 diabetes. They note that GLP-1RA use was associated with a similar risk reduction as observed with SGLT2i use, and a vast body of randomized evidence supports the use of SGLT2is for reducing the risk of hospitalization for heart failure; by extension, GLP-1RAs may represent a potential therapeutic option for heart failure prevention in patients with type 2 diabetes. In a secondary analysis restricted to individuals without heart failure at baseline, the largest difference in heart failure risk was observed in individuals with the highest predicted risk of heart failure at baseline as estimated using the Thrombolysis in Myocardial Infarction Risk Score for Heart Failure in Diabetes. Future studies should seek to replicate these findings across other, more widely used heart failure risk scores.

A major strength of this study is its rigorous application of the target trial emulation framework, a state-of-the-science approach for causal inference using observational data. The three key steps in this approach are: 1) Defining the causal question of interest; 2) Specifying the protocol of a hypothetical randomized trial that would definitively answer the question of interest (i.e., the “target trial”), including eligibility criteria, treatment strategies, assignment procedure, follow-up period, outcome, causal contrast of interest, and analysis plan, and 3) Emulating the target trial using observational data.⁹ A critical element of this framework is the clear definition of time zero, which is the point in time when eligibility is assessed, treatment is initiated, and follow-up begins. Aligning these components mitigates key sources of bias commonly encountered in nonrandomized studies, including immortal time bias and selection bias. A reporting guideline for studies using the target trial emulation framework (The TARGET Statement) has been recently published.¹⁰

Although target trial emulation mitigates key sources of bias in observational studies, its findings can be interpreted as causal if key assumptions are met, such as the relevant patient characteristics and clinical factors are known and accounted for, the treatment strategy is well defined with a clearly defined start of therapy, and patients are followed forward in time to ascertain outcomes (i.e., absence of uncontrolled confounding [“conditional exchangeability”], a non-zero probability of receiving treatment [“positivity”], a well-defined intervention corresponding to the treatment strategy under study [“consistency”], and absence of informative loss to follow-up [“non-informative censoring”]). The use of an active comparator in the present study by Xu and colleagues could mitigate the potential bias related to treatment indications (i.e., confounding by indication) in some scenarios, and adjustment for a comprehensive set of baseline and time-varying covariates reduces confounding. Accounting for time-varying covariates allows adjustment for time-varying confounding, which is particularly important when estimating the per-protocol effect because patients who discontinue or switch the treatment often do so in response to changes in their health status. These clinical factors both influence subsequent treatment decisions and predict cardiovascular outcomes; so, failing to account for them may bias estimates of the true effect of sustained GLP-1RA use.

In addition, the validity of the estimated effects depends on the comprehensive and accurate measurement of covariates. In the present study by Xu and colleagues,⁷ unmeasured or incompletely captured determinants of prescribing GLP-1RAs, such as socioeconomic status, frailty, or subtle differences in cardiovascular risk not reflected in structured data, may influence both treatment selection and heart failure risk. Of note, the early divergence of the cumulative incidence curves for semaglutide versus DPP-4 inhibitors suggests that there may be some residual confounding, which may partially account for the observed differences in outcomes. Residual confounding can be assessed by analyzing negative control outcomes, i.e., outcomes that share similar sources of bias with the primary outcome yet are unrelated to GLP-1RA treatment.¹¹ However, negative control outcomes can be difficult to identify for a therapy like GLP-1RAs that affect a broad range of organ systems.¹² Measurement error in heart failure diagnoses and incomplete ascertainment of covariates can also introduce bias, while such errors are likely to be nondifferential between treatment strategies. These limitations do not diminish the study’s contributions but underscore the need for cautious interpretation of the causal relationship between GLP-1RAs and hospitalization for heart failure.

Looking ahead, several areas warrant further investigation. First, to address the anticipated need for lifelong therapy, we need to understand the long-term outcomes of GLP-1RAs therapy on heart failure beyond 3 years, and whether these vary by specific agent, formulation, or dose. Second, given the substantial phenotypic heterogeneity among individuals with diabetes, including variation in body mass index, blood pressure, kidney function, and metabolic profiles, it is important to examine whether the cardioprotective effects of GLP-1RAs vary by these or other characteristics. These assessments of heterogeneous treatment effects must go beyond univariable subgroup analyses (e.g., men vs. women; age <65 years vs. ≥65 years, etc.) to study multivariable scores of predicted risk or benefit.¹³ Such heterogeneity assessment may be particularly important for drugs like GLP-1RAs for which adherence in clinical practice is substantially lower than in clinical trials with notable differences among clinical and demographic subgroup. We have previously shown that showed that nearly 6 in 10 Medicare beneficiaries with diabetes discontinued injectable semagltuide within 12 months of initiation, and, paradoxically, rates of discontinuation were higher in groups at high risk of atherosclerotic cardiovascular disease compared with lower-risk populations.¹⁴ Whether the cardioprotective effect of GLP-1RAs differs between HFpEF and HFrEF is also an important question that could not be examined in the present study because data on ejection fraction were unavailable. Third, future studies should examine the effect of GLP-1RAs on risk of hospitalization for heart failure among patients without diabetes in the real-world clinical settings. This is important because, of the 137 million U.S. adults eligible for semaglutide therapy, the majority (98 million, or 71%) do not have diabetes.¹⁵ Finally, this study was conducted in Stockholm residents and within the universal health system of Sweden; therefore the findings may not be extendable to other healthcare systems, and confirmation in other populations is warranted.

In summary, Xu and colleagues should be commended for their rigorous application of the target trial emulation framework to advance understanding of the relationship between GLP-1RA therapy and risk of heart failure in patients with diabetes. As the global use of GLP-1RAs skyrockets, their study provides timely and clinically relevant insights that inform evidence-based practice and support more comprehensive cardiometabolic care.