Glucagon-like peptide-1 receptor agonists and heart failure in type 2 diabetes: systematic review and meta-analysis of randomized and observational studies

Background The effect of glucagon-like peptide-1(GLP-1) receptor agonists on heart failure remains uncertain. We therefore conducted a systematic review to assess the possible impact of GLP-1 agonists on heart failure or hospitalization for heart failure in patients with type 2 diabetes. Methods We searched MEDLINE, EMBASE, the Cochrane Central Register of Controlled Trials (CENTRAL) and ClinicalTrials.gov to identify randomized controlled trials (RCTs) and observational studies that addressed the effect of GLP-1 receptor agonists in adults with type 2 diabetes, and explicitly reported heart failure or hospitalization for heart failure. Two paired reviewers screened reports, collected data, and assessed the risk of bias. We pooled data from RCTs and observational studies separately, and used the GRADE approach to rate the quality of evidence. Results We identified 25 studies that were eligible for our review; 21 RCTs (n = 18,270) and 4 observational studies (n = 111,029). Low quality evidence from 20 RCTs suggested, if anything, a lower incidence of heart failure between GLP-1 agonists versus control (17/7,441 vs. 19/4,317; odds ratio (OR) 0.62, 95 % confidence interval (CI) 0.31 to 1.22; risk difference (RD) 19 fewer, 95 % CI 34 fewer to 11 more per 1000 over 5 years). Three cohort studies comparing GLP-1 agonists to alternative agents provided very low quality evidence that GLP-1 agonists do not increase the incidence of heart failure. One RCT provided moderate quality evidence that GLP-1 agonists were not associated with hospitalization for heart failure (lixisenatide vs placebo: 122/3,034 vs. 127/3,034; adjusted hazard ratio 0.96, 95 % CI 0.75 to 1.23; RD 4 fewer, 95 % CI 25 fewer to 23 more per 1000 over 5 years) and a case–control study provided very low quality evidence also suggesting no association (GLP-1 agonists vs. other anti-hyperglycemic drugs: 1118 cases and 17,626 controls, adjusted OR 0.67, 95 % CI 0.32 to 1.42). Conclusions The current evidence suggests that GLP-1 agonists do not increase the risk of heart failure or hospitalization for heart failure among patients with type 2 diabetes. Electronic supplementary material The online version of this article (doi:10.1186/s12872-016-0260-0) contains supplementary material, which is available to authorized users.


Background
Glucagon-like peptide-1 (GLP-1) receptor agonists are a relatively new class of incretin-based agents for the treatment of type 2 diabetes mellitus that lower blood glucose [1,2], reduce body weight [3], and possibly reduce cardiovascular risk compared to placebo [4,5]. The American Diabetes Association and the European Association for the Study of Diabetes recommend GLP-1 agonists as a second-line treatment option for type 2 diabetes [6].
In 2014, the US Food and Drug Administration raised concerns regarding heart failure risk with one dipeptidyl peptidase-4 (DPP-4) inhibitor, saxagliptin [7]. These concerns followed publication of studies that reported increased risk of hospitalization for heart failure in patients using DPP-4 inhibitors [8][9][10]. These observations raise the possibility that GLP-1 agonists, which share a similar pharmacological mechanism with DPP-4 inhibitors, might also cause heart failure.
Animal studies have shown that the GLP-1 agonist liraglutide can activate cytoprotective pathways in the heart, and improve outcomes after experimental myocardial infarction in mice [11]. Early clinical studies also suggested that GLP-1 agonists have positive effects on cardiovascular biomarkers, such as high-sensitivity Creactive protein and plasminogen activator inhibitor-1 [12,13], and improve regional and overall left ventricular function in patients with acute myocardial infarction and severe systolic dysfunction after successful primary angioplasty [14].
Clinical trial results often, however, prove inconsistent with laboratory and surrogate outcome studies, and emerging randomized trials and observational studies have, reported inconsistent results [15][16][17][18][19]. We therefore undertook a systematic review to address the effect of GLP-1 agonists on heart failure or hospitalization for heart failure in patients with type 2 diabetes.

Methods
We followed the PRISMA and MOOSE guidelines for conducting and reporting systematic reviews and meta-analyses of randomized controlled trials (RCTs) and observational studies [20,21].

Data sources and search strategy
We searched MEDLINE, EMBASE, and the Cochrane Central Register of Controlled Trials (CENTRAL) from inception to 25 June, 2015. We used both MeSH and free text terms to identify relevant articles. An information expert (DP) developed each database-specific search strategy (Additional file 1). We also searched Clinical-Trials.gov as well as conference abstracts published by the American Diabetes Association, European Association for the Study of Diabetes, and European Society of Cardiology for additional eligible studies and trial information.

Eligibility criteria
We included RCTs, cohort studies, or case-control studies that compared GLP-1 agonists against placebo, lifestyle modification, or active anti-hyperglycemic medication in adult type 2 diabetes patients, reported ≥ 12 weeks follow-up data (not applicable to case-control studies), and explicitly reported the outcome of heart failure or hospitalization for heart failure.

Study selection
Paired reviewers, trained in research methods, independently screened titles/abstracts and then full texts for eligibility, assessed risk of bias, and collected data from each included study, using pilot-tested standardized forms with corresponding detailed instructions. Any disagreement between the two reviewers was resolved through discussion or adjudication by a third reviewer (XS).

Risk of bias and quality of evidence assessment
We assessed the risk of bias of RCTs according to modified version of the Cochrane Collaboration's tool [22,23] in which the response options are "probably yes" and "probably no" instead of "unclear"; the approach has shown to be reliable and valid for blinding [24]. The items include randomization sequence generation; allocation concealment; blinding of participants, caregivers, outcome assessors (i.e., heart failure or hospitalization for heart failure), and outcome adjudicators; prognostic balance between treatment groups; and incomplete outcome data.
We used a modified version of the Newcastle -Ottawa Quality Assessment Scale [25][26][27] for assessing risk of bias of observational studies. Specifically, we removed two items "representativeness of the exposed cohort" and "was follow-up long enough for outcomes to occur" that we judge related to applicability, and added two items -ascertainment of type 2 diabetes and adjustment for potential confounding factors. We planned to assess for risk of publication bias, but were unable to do so due to low power of the relevant tests in the presence of low events rates.
We rated the quality of evidence for heart failure and hospitalization for heart failure as high, moderate, low, or very low using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) methodology [28][29][30][31][32][33][34].

Data extraction
We collected the following information from each eligible studies: study characteristics (e.g., author name, year of publication, study design, sample size, length of follow-up), patient characteristics (e.g., gender, age, diabetes duration, body mass index (BMI), baseline HbA1c level), interventions (e.g., details of GLP-1 agonists therapy and control group), and outcomes (number of events and patients included for analyses in each group, as well as adjusted data if available). For trials with multiple reports, we collated all data into a single study [35]; for trials with reports both from ClincialTrials.gov and journal publications, we carefully checked the data for consistency; for trials reporting outcome data of multiple follow up points, we used the data with longest follow up. For observational studies, we also collected information on data source, methods used to control confounding, and reported adjustment factors.

Statistical analysis
We analyzed RCTs and observational studies separately. We did not combine the outcomes of heart failure and hospitalization for heart failure, as hospitalization for heart failure is likely more serious and of greater importance to patients than heart failure not requiring hospitalization.
We assessed statistical heterogeneity with the Cochran chi-square test and I-squared statistic. We used Peto's method to pool data from RCTs [36,37] using random effects models and reported pooled Peto odds ratios (ORs) and associated 95 % confidence intervals (CIs). We conducted four a priori subgroup analyses to explore heterogeneity associated with our pooled estimates: (1) type of control (placebo vs. active treatment), (2) length of follow up (52 weeks or shorter vs. over 52 weeks), (3) mode of therapy (GLP-1 agonists monotherapy vs. addon/combination therapy), and (4) individual GLP-1 agonists agents (different GLP-1 agonists agents vs. control). We also carried out sensitivity analyses to explore the robustness of our findings using different effect measures, pooling methods, and statistical models.
We pooled adjusted estimates of heart failure from cohort studies using random effects model due to significant variations in the comparison and patient populations among eligible studies.

Ethics
Ethical approval was not necessary as this study is a Systematic Review and Meta-Analysis.

Study selection
Our literature search yielded 11,441 reports; 821 were potentially eligible after title and abstract screening, and 25 studies proved eligible after full text screening. These included 21 RCTs involving 18,270 patients from 30 reports [15,16, and four observational studies [17][18][19]66] involving 111,029 patients (three cohort studies and one nested case-control study) (Fig. 1).
Evidence from randomized controlled trials RCTs reporting heart failure Twenty trials reported on heart failure; 18 (80 %) were multi-center studies, and 18 (90 %) were clearly labeled as phase III trials. These trials enrolled 46 to 1,091 patients (total 12,199); the mean age of patients ranged 52.9 to 67.2 years old, mean BMI 25.6 to 33.3 kg/m 2 , mean baseline HbA1c 7.6 to 8.5 %, mean FPG 7.1 to 10.0 mmol/L, and mean or median duration of diabetes was 2.6 to 11.5 years (Table 1). Five used GLP-1 agonists as monotherapy, 15 as add-on or combination therapy ( Table 2). The length of follow-up ranged from 16 to 164 weeks (median 52; 10 trials followed patients for > 52 weeks).
All the trials reported industry funding; 18 were identified from ClinicalTrials.gov, of which 12 had no corresponding journal publications. Because of the limited information provided in the trial registry, we were unable to adequately assess the risk of bias for these 12 trials. Additional file 2 presents the details of the assessment for risk of bias. The baseline demographics and clinical characteristics of patients in each included trials were generally balanced between groups. The overall risk bias of eligible RCTs was moderate.
Twenty trials reported 36 heart failure events in 11,758 patients using at least one medication (raw event rate 0.3 %). The pooling of those trials showed no statistically significant difference in the risk of heart failure between GLP-1 agonists treatment and control (17/7,441 in GLP-1 agonists and 19/4,317 control; OR 0.62, 95 % CI 0.31 to 1.22, I-square = 0 %; risk difference (RD) 19 fewer, 95 % CI 34 fewer to 11 more per 1000 over 5 years) (Fig. 2). We rated the quality of evidence as low because of risk of bias and imprecision (Table 3).
Subgroup analysis by type of control (interaction p = 0.79), mode of therapy (interaction p = 0.84) and length of follow up (interaction p = 0.64) showed no differential treatment effects (Additional files 3, 4, 5 and 6). The subgroup analysis of heart failure risk by individual GLP-1 agonists agents suggested a possibility of differential treatment effect across individual agents (interaction p = 0.07), with liraglutide associated with a non-significant increased risk for heart failure (OR 4.85, 95 % CI 0.75 to 31.36); this finding was however based on a limited number of events (five in total) and characterized with very wide confidence interval.
Sensitivity analysis using alternative effect measures, statistical methods, and analysis models did not show important changes in pooled effects.

Trials reporting hospitalization for heart failure
The Evaluation of LIXisenatide in Acute Coronary Syndrome (ELIXA) trial, designed to assess the cardiovascular safety of lixisenatide, reported hospitalization for heart failure [15,16] (Table 1). The ELIXA trial randomized 6,068 patients with type 2 diabetes and a recent acute coronary syndrome to lixisenatide or placebo, with a median of follow up of 2.1 years. In this trial, 122 patients were hospitalized for heart failure among 3,034 patients taking lixisenatide (4.0 %) and 127 in 3034 patients taking placebo (4.2 %), and no statistically significant difference was present between the groups (hazard ratio (HR) 0.96, 95 % CI 0.75 to 1.23; RD 4 fewer, 95 % CI 25 fewer to 23 more per 1000 over 5 years). The trial authors' subgroup analysis by type of history of heart failure showed no differential treatment effects (lixisenatide vs. placebo: patients with history of heart failure: HR 0.93, 95%CI 0.66 to 1.30; patients with no history of heart failure: HR 0.97, 95 % CI 0.67 to 1.40). We rated the quality of evidence as moderate (Table 3).

Evidence from observational studies Studies reporting heart failure
Three cohort studies [17,18,66] reported heart failure. Of these, one prospectively designed study [66] examined exenatide versus basal insulin; the other two [17,18] retrospective in design -assessed GLP-1 agonists versus sulfonylureas, and exenatide or exenatide plus insulin versus insulin (Tables 4 and 5). The sample sizes ranged from 882 to 39,225, and length of follow up ranged from 1 to 4 years. The mean age ranged from 58.28 to 62.5 years, BMI 32.6 to 35.3 kg/m 2 , and mean baseline HbA1c 7.9 to 8.9 %.
The three studies used electronic heath records or claims data for their analyses. Type 2 diabetes patients were ascertained by specialists in outpatient setting in the prospective cohort study [66]; the other two retrospective cohort study [17,18] did not explicitly state the  [17] demonstrated that outcome of interest was not present at start of study, and mentioned the method used to assess the outcome of interest. Two studies [18,19] used advanced statistical model to control for the influence of confounding factors. Overall, the risk of bias associated with these studies was moderate to high (Additional file 7). All three studies reported raw data, for a total of 2,868 heart failures among 53,292 patients (raw event rate 5.4 %); two retrospective cohort studies [17,18] reported adjusted effect estimates (Tables 5 and 6). The prospective cohort study [66], enrolling 882 patients with one year follow-up, found that two patients (2/438) in the basal insulin had heart failure events and no patients (0/444) in exenatide group. One retrospective cohort study [17], including 13,185 patients and with a median follow-up of four years, reported that GLP-1 agonists were associated with a nonsignificant increase in heart failure versus sulfonylureas (adjusted HR 1.10, 95 % CI 0.99 to 1.22). The other retrospective cohort study [18], involving 39,225 patients and with a median follow-up of 3.5 years, found that both exenatide and exenatide plus insulin were associated with a lower risk of heart failure versus insulin alone (adjusted HR 0.34, 95 % CI 0.22 to 0.52; adjusted HR 0.40, 95 % CI 0.32 to 0.50, respectively, Fig. 3). Using GRADE, we rated the quality of evidence in the included studies as very low, due to risk of bias, indirectness and heterogeneity in addition to the inherent risk for confounding associated with observational studies.

Studies reporting hospitalization for heart failure
One nested case-control study [19] assessed with GLP-1 agonists versus other oral anti-hyperglycemic drugs Using GRADE, we rated the quality of evidence as very low, due to risk of bias and imprecision in addition to the inherent risk for confounding associated with observational studies.

Main findings
Our pooled analysis of 20 RCTs addressing use of GLP-1 agonists for type 2 diabetes found moderate quality evidence suggesting no increase in heart failure. The only RCT provided high quality evidence that lixisenatide did not increase the risk of hospitalization due to heart failure. Though the four observational studies provide only very low quality evidence, their results are consistent with those from the randomized trials.

Strengths and limitations
We are the first to systematically review the evidence regarding GLP-1 agonists for type 2 diabetes and risk of heart failure. Our study has several strengths. First, we used rigorous methods to systematically identify both randomized and observational studies that reported data to inform this issue, including a large number of trials that were not published in journals. Second, we carefully checked the data reported in ClinicalTrials.gov and journal publications for consistency to ensure accuracy of the data. Third, we analysed the data on heart failure and hospitalization for heart failure separately, because those outcomes are likely to be of different importance to patients. Fourth, we used the GRADE approach to assess the quality of evidence on an outcome-by-outcome basis.
Our study also has limitations. First, the available evidence is not strength to give definitive answer for this question, since the included RCTs reported few heart failure events and the follow-up was not enough for heart failure to occur, and much findings came from observational studies of very low quality evidence. Second, we have included some observational studies at moderate to high risk of bias. This has made the inference about the effects of GLP-1 agonists challenging. Third, the diversity of observational studies also made our analysis of the evidence difficult. One retrospective cohort study [18], assessing exenatide and/or insulin on heart failure outcome, included patients with heart failure at baseline, and the proportion of patients with history of heart failure was higher in the insulin group (3.2 %) than in the exenatide group (1.7 %) and exenatide + insulin group (2.4 %), which made the finding from this study biased.

Other researches
Ours is the first systematic review addressing the impact of GLP-1 agonists on heart failure. There is some evidence from human studies that GLP-1 agonists might provide protection against heart failure: preliminary study [67] showed that GLP-1 treatment might have a trend towards improvement of cardiac function in type 2 diabetes patients with stable heart failure; intrinsic GLP-1 expression has been shown to compensatorily upregulate in patients with left heart failure [68]; and GLP-1  Several trials probably had risk of bias on random sequence generation, allocation concealment and blinding (Additional file 2), and the follow up (median of 52 weeks) was not long enough for heart failure to occur in patients with low risk of cardiovascular disease b Baseline risk estimate for heart failure in a 5-year time frame comes from the control arm of the cohort study we identified to best represent our target population (Kannan 2015 [17]), with 528 events of heart failure in 13,185 participants (4.0 %) at four year follow up across control and intervention arm c Baseline risk estimate for hospitalization for heart failure in 5-year time frame comes from the control arm of the only included ELIXA trial [16] we identified to best represent our target population with 127 events in 3034 participants (42 per 1000) over a 2.1 year follow up period, in the absence of observational studies providing more credible baseline risk estimates  These two studies accessed incretin agents and the risk of heart failure, and the data of events/cases and total number of analyzed patients regarding glucagon-like peptide-1 receptor agonists and dipeptidyl peptidase-4 inhibitors were not reported separately, so the data above were the data of total study patients CI confidence interval, NR not reported, HR hazard ratio, OR odds ratio, CVD cardiovascular disease, BMI body mass index agonists are also shown to be associated with a modest increase of ejection fraction in diabetic patients [69]. A recent meta-analysis of RCTs [70] found that GLP-1 agonists were associated with a modest reduction in blood pressure and a slight increase in heart rate. These biological studies suggest that GLP-1 agonists might, if anything, reduce the incidence of heart failure. Though results of RCTs fail to show this decrease, confidence intervals do not exclude the possibility of a modest reduction.

Conclusions
The current evidence suggests that GLP-1 agonists do not increase the risk of heart failure or hospitalization for heart failure. The current body of evidence, however, is not definitive. More carefully designed, conducted, adequately powered trials and observational studies are warranted to confirm the effects of GLP-1 agonists on incidence of heart failure and hospitalization for heart failure. Future studies should also examine whether the