Glucagon-Like Peptide 1 Receptor Agonists and the Risk of Incident Diabetic Retinopathy

This study was conducted using the U.K. Clinical Practice Research Datalink (CPRD), which was linked to the Hospital Episode Statistics (HES) repository. The CPRD contains the medical records of more than 14 million people enrolled across 700 general practices (12). Medical diagnoses and procedures are recorded using the Read code classification, and drugs prescribed by general practitioners are coded using the U.K. Prescription Pricing Authority dictionary. The CPRD contains information on anthropometric variables (e.g., BMI) and lifestyle variables (e.g., smoking), and diagnoses recorded in this database have been previously validated and shown to be of high quality (13). The HES contains all inpatient admissions, including primary and secondary diagnoses (coded using the ICD-10) and hospital-related procedures. The linkage of the CPRD to the HES is possible from 1 April 1997 onward and is limited to general practices in England that have consented to the linkage scheme (currently representing 75% of all practices in England) (13). The study protocol was approved by the CPRD Independent Scientific Advisory Committee (protocol number 16_287R) and by the Research Ethics Board of the Jewish General Hospital, Montreal, Canada.

We first identified a base cohort of patients newly treated for type 2 diabetes. This included all patients, at least 18 years of age, initiating a new noninsulin antidiabetic drug (metformin, sulfonylureas, prandial glucose regulators, acarbose, thiazolidinediones, dipeptidyl peptidase 4 [DPP-4] inhibitors, GLP-1 RAs, and sodium–glucose cotransporter 2 inhibitors) between 1 April 1998 and 30 September 2015. All patients were required to have at least 1 year of medical history recorded in the CPRD before their first prescription. We excluded patients initially treated with insulin (because these patients are likely to have advanced disease) and women with a history of polycystic ovarian syndrome or a history of gestational diabetes mellitus in the last year (other known metformin indications).

Using this base cohort, we assembled a study cohort of all patients initiating a new antidiabetic drug class on or after 1 January 2007; the first GLP-1 RA, exenatide, was approved in the U.K. on 20 November 2006 (14). These patients included those newly treated with a noninsulin antidiabetic drug as well as those who switched to or added on an antidiabetic drug from a class not previously used in their treatment history. Cohort entry was defined by the date of this new prescription.

We excluded patients previously diagnosed with any retinal disease (codes available upon request) or with medical conditions associated with retinopathy (HIV infection, history of bariatric surgery), and those previously using drugs associated with retinopathy (imatinib, acitretin, nicotinic acid, rituximab, taxanes, interferon, zilovudine, rifabutine, and fingolimod) (10,15,16).

Patients were monitored until the earliest of the following events: incident diagnosis of diabetic retinopathy recorded in the outpatient or inpatient setting (Read and ICD-10 codes listed in Supplementary Table 1), end of registration with the general practice, death from any cause, or end of the study period (30 September 2015).

We used a time-varying exposure definition, in which each person-day of follow-up was classified hierarchically into one of the following four mutually exclusive categories: current use of GLP-1 RAs (exenatide, liraglutide, lixisenatide; alone or in combination with other antidiabetic drugs other than insulin), current use of DPP-4 inhibitors (alone or in combination with other antidiabetic drugs other than insulin), current use of two or more oral antidiabetic drugs, and others, which included current use of other antidiabetic drugs and treatment combinations as well as no current use of antidiabetic drugs. Current use refers to each patient’s exposure category on the day of follow-up included in the risk set. For all exposure categories, we defined exposed person-time by the prescription duration plus a 30-day grace period. Thus, continuous use was assumed if the prescription duration overlapped with the date of the next prescription, allowing for the 30-day grace period in the case of two nonoverlapping successive prescriptions. Because GLP-1 RAs are recommended as second- or third-line treatment in the management of type 2 diabetes (1), the reference category for our analyses consisted of current use of two or more oral antidiabetic drugs.

All models were adjusted for potential confounders assessed at study cohort entry. These included year of cohort entry, age, sex, quintiles of the Index of Multiple Deprivation (17), smoking status, BMI category (<25 kg/m², 25–29 kg/m², ≥30 kg/m², unknown), hemoglobin A_1c (HbA_1c) level (≤7% or ≤53 mmol/mol, 7.1–8.0% or 54–64 mmol/mol, >8% or >64 mmol/mol, unknown; last measurement before cohort entry), systolic and diastolic blood pressure (last measurement before cohort entry), and duration of treated diabetes. Age and duration of treated diabetes were modeled flexibly (i.e., restricted cubic splines with five interior knots) as continuous variables to account for potentially nonlinear relationships with the outcome (18). Due to expected missing data, BMI and HbA_1c were modeled as categorical variables. The models were also adjusted for the following variables recorded at any time before cohort entry: alcohol-related disorders, dyslipidemia, diabetic arterial complications (neuropathy, nephropathy, peripheral arteriopathy, myocardial infarction, ischemic stroke), history of cataract surgery, albuminuria or proteinuria, uveitis, and sickle cell disease. We adjusted for the use of statins, fibrates, antihypertension drugs, ophthalmic agents, antimalarial drugs, fluconazole, and tamoxifen in the year before cohort entry. Finally, we also adjusted for the overall number of nonantidiabetic drugs and the number of physician visits in the year before cohort entry.

We used descriptive statistics to summarize the characteristics of the cohort. We calculated crude incidence rates, with 95% CIs based on the Poisson distribution, of diabetic retinopathy overall and for each exposure category. Cox proportional hazards regression models using the time-varying exposure definition were used to estimate adjusted HRs with 95% CIs of incident diabetic retinopathy associated with current use of GLP-1 RAs, compared with current use of two or more oral antidiabetic drugs. All models were adjusted for the potential confounders listed above.

We conducted three prespecified secondary analyses. First, to explore whether the use of GLP-1 RAs is associated with a transient increased risk of diabetic retinopathy, current use was further categorized according to three predefined continuous durations of use (≤6 months, 6.1–12 months, >12 months). The P value for heterogeneity across the HRs of the different durations of use was calculated using a χ² test. Second, we assessed the association with individual GLP-1 RAs (liraglutide and exenatide; no lixisenatide-specific analyses were conducted due to the low number of exposed patients). Finally, we assessed whether duration of treated diabetes modified the association by including an interaction term between duration of treated diabetes (<5 vs. ≥5 years) and exposure to GLP-1 RAs.

We also conducted three post hoc secondary analyses. In the first two analyses, we assessed whether arterial hypertension (defined as systolic blood pressure ≥140 mmHg or diastolic blood pressure ≥90 mmHg or use of any antihypertension medication) or use of ACE inhibitors or angiotensin receptor blockers (ARBs) modified the association. Finally, we also assessed whether baseline HbA_1c level (≤7% or ≤53 mmol/mol, 7.1–8.0% or 54–64 mmol/mol, >8% or >64 mmol/mol) modified the association. For this analysis, exposure was defined based on the drugs received at cohort entry with a maximum follow-up of 1 year.

We conducted seven prespecified sensitivity analyses to assess the robustness of our findings:

1. To assess possible exposure misclassification, we repeated the primary analysis by lengthening the grace period to 90 days between prescriptions that did not overlap.

2. To adjust further for potential confounding by disease severity, we additionally adjusted for the use of antidiabetic drugs in the year before cohort entry.

3. Because an association between thiazolidinediones and macular edema has been suggested (16), we redefined our exposure variable by removing current use of thiazolidinediones from the GLP-1 RA and reference (two or more oral antidiabetic drugs) exposure groups and including such use in the “other treatment combinations” group.

4. We used multiple imputations for variables with missing information (BMI, HbA_1c level, smoking status, Index of Multiple Deprivation).

5. We used the high-dimensional disease risk score approach as an alternate means to control for confounding (19). This method uses an algorithm that empirically selects covariates along with prespecified covariates for the statistical analysis and is recommended in settings similar to our study where the exposure of interest is relatively rare but the outcome of interest is common (20).

6. We repeated the analyses to account for competing risk due to deaths from any cause (21).

7. To assess the possible inclusion of prevalent retinopathy events (i.e., patients who had the condition before study cohort entry but were not yet diagnosed), we stratified the cohort based on the presence of at least one retinopathy screening visit in the year before cohort entry (the recommended screening regimen according to U.K. guidelines [22]).

Finally, in a post hoc analysis, we assessed the potential effect of residual confounding on our point estimates using the approach proposed by Ding and VanderWeele (23), a model that does not impose any assumptions on the unmeasured confounder or confounders, such as having an unmeasured confounder that is binary, or having no interaction between the effects of the exposure and the confounder on the outcome, or having only one unmeasured confounder.

To further assess the validity of our findings, we conducted two additional analyses using negative and positive control exposures (24). For the negative control exposure, we repeated the analyses comparing current use of DPP-4 inhibitors with current use of two or more oral antidiabetic drugs; the former are incretin-based drugs, but have not been associated with diabetic retinopathy in previous large RCTs (25–27). For the positive control exposure, we repeated the analyses comparing current use of insulin with current use of two or more oral antidiabetic drugs; the former has been previously associated with a transient increased risk of diabetic retinopathy (9).

We conducted a post hoc ancillary analysis head-to-head comparison of new users of GLP-1 RAs with new users of insulin. For this analysis, we identified patients in our base cohort initiating a GLP-1 RA or insulin after January 2007. Cohort entry was defined as the date of the first prescription of GLP-1 RA or insulin. We then estimated propensity scores using multivariate logistic regression to estimate the predicted probability of receiving a GLP-1 RA versus insulin, conditional on the variables listed above. Patients with nonoverlapping propensity score distributions were trimmed from the analysis. The remaining patients were monitored from cohort entry until they switched from GLP-1 RAs to insulin or vice versa, discontinued treatment, or experienced the outcome, whichever occurred first. Finally, the HR of incident diabetic retinopathy was estimated using a Cox proportional hazards model adjusted for the propensity score, which was included in the model as an interaction term between propensity score deciles and the propensity score as a continuous variable. All analyses were conducted with SAS 9.4 software (SAS Institute, Cary, NC).

The cohort included 77,115 new users of antidiabetic drugs (Fig. 1). The median duration of follow-up was 2.8 (maximum 8.8) years, generating 245,825 person-years. During follow-up, 3,047 patients received GLP-1 RAs (97% in combination therapy), with the median duration of use being 0.8 (maximum 7.3) years (median duration of use for the comparator group, i.e., two or more oral antidiabetic drugs, was 0.6 [maximum 8.7] years). Overall, 10,763 patients were newly diagnosed with diabetic retinopathy during follow-up, corresponding to an overall incidence rate of 43.8 (95% CI 43.0–44.6) per 1,000 persons per year.

Table 1 presents the characteristics of patients who received GLP-1 RAs versus two or more oral antidiabetic drugs at cohort entry. Compared with users of two or more oral antidiabetic drugs, users of GLP-1 RAs were younger, more likely to be women, and more likely to be obese. They were also more likely to have elevated HbA_1c levels (>8% or >64 mmol/mol), have a history of neuropathy, nephropathy, or proteinuria, and to have used antihypertension drugs.

Table 2 presents the results related to the use of GLP-1 RAs. Compared with current use of two or more oral antidiabetic drugs, current use of GLP-1 RAs was not associated with an overall higher risk of diabetic retinopathy (crude incidence rates, 40.4 vs. 49.0 per 1,000 persons per year; adjusted HR 1.00, 95% CI 0.85–1.17). However, there was a suggestion of heterogeneity across the duration categories. A duration of use ranging between 6.1 and 12 months was associated with a 44% increased risk of diabetic retinopathy (crude incidence rates, 56.6 vs. 45.9 per 1,000 persons per year; adjusted HR 1.44, 95% CI 1.06–1.95). This association was not observed for shorter and longer durations of use of ≤6 months (crude incidence rates, 38.2 vs. 51.3 per 1,000 persons per year; adjusted HR 0.94, 95% CI 0.76–1.17) and >12 months (crude incidence rates, 33.2 vs. 47.4 per 1,000 persons per year; adjusted HR 0.83; 95% CI 0.60–1.15) (P for heterogeneity = 0.07). Drug-specific analyses revealed similar overall and duration patterns for liraglutide and exenatide, but these did not achieve statistical significance due to the smaller number of exposed events (Supplementary Table 2). Duration of treated diabetes and HbA_1c level did not modify the association between GLP-1 RA use and risk of diabetic retinopathy (Supplementary Tables 3 and 4). However, the risk of GLP-1 RA–associated diabetic retinopathy was increased among patients with arterial hypertension or use of ACE inhibitors or ARBs (Supplementary Tables 5 and 6).

Figure 2 summarizes the results of the sensitivity analyses (presented in detail in Supplementary Tables 7–14). The results remained consistent with those of the primary and secondary analyses in overall use and duration of use. With respect to the latter, GLP-1 RA durations ranging between 6.1 and 12 months were consistently associated with an increased risk of diabetic retinopathy across the different sensitivity analyses. Based on a post hoc analysis, these findings are unlikely to be the result of an unmeasured confounder under most plausible exposure-confounder and confounder-outcome associations (Supplementary Table 15).

With respect to our control exposures (patient characteristics in Supplementary Table 16; results summarized in Supplementary Fig. 1 and presented in Supplementary Tables 17 and 18), the use of DPP-4 inhibitors (negative control exposure) was not associated with an increased risk of diabetic retinopathy overall or by duration of use. In contrast, insulin (positive control exposure) was associated with an increased risk of diabetic retinopathy overall, and with evidence of a duration-response relation. Finally, Table 3 presents the comparison of new users of GLP-1 RAs with new users of insulin (cohort assembly shown in Supplementary Fig. 2, patient characteristics reported in Supplementary Table 19). Overall, the use of GLP-1 RAs was associated with a decreased risk of diabetic retinopathy (HR 0.67; 95% CI 0.51–0.90). The decreased risk was observed after a duration of at least 12 months of use (HR 0.48; 95% CI 0.31–0.76), whereas the HRs were close to the null for shorter durations of ≤6 months (HR 0.84; 95% CI 0.55–1.27) and 6.1–12 months (HR 1.05; 95% CI 0.64–1.72).

The results of this population-based study indicate that when compared with the use of two or more oral antidiabetic drugs, the use of GLP-1 RAs is not associated with an overall increased risk of diabetic retinopathy. In a secondary analysis, there was a suggestion of a transient 44% increased risk with GLP-1 RA durations ranging 6 and 12 months, an effect that appeared to be more pronounced in patients with arterial hypertension. Compared with insulin, GLP-1 RAs were associated with a 33% decreased risk of diabetic retinopathy.

Although the findings of our primary analysis suggest a null association between GLP-1 RAs and incident diabetic retinopathy overall, the results of our duration-response analyses suggest a potential transient increase risk in this outcome. A possible mechanism for this observation may involve large and rapid improvements in glycemic control, which have previously been linked to a transient worsening of diabetic retinopathy (9,10) via increased insulin-like growth factor levels and retinal ischemia (28,29). Interestingly, in the SUSTAIN-6 and LEADER trials, a divergence in the Kaplan-Meier curves was observed during the 1st year after randomization (4,30). Moreover, the time interval of the increased risk corresponds to when GLP-1 RA users achieve their greatest drops in HbA_1c levels (31). The fact that the association decreased with longer durations of use (>12 months) may relate to the depletion of susceptible phenomenon, where patients susceptible of developing retinopathy selected themselves out of the exposure group in the early phase of treatment (32).

Another possible mechanism might involve a direct drug effect on retinal cells, because GLP-1 receptors have been shown to be abundantly expressed in the retina (11). However, the fact that DPP-4 inhibitors, drugs that increase endogenous GLP-1 levels, were not associated with an increased risk of diabetic retinopathy argues against this hypothesis. Finally, it would be of interest to further investigate the transient increased risk observed in our study by reanalyzing data from the recent large RCTs (e.g., LEADER and SUSTAIN-6 trials) and providing information on temporal patterns of incident diabetic retinopathy (3,4).

In a post hoc ancillary analysis, GLP-1 RAs were associated with a decreased risk of incident diabetic retinopathy compared with insulin. However, this finding should be interpreted with caution. First, insulin users were older and more likely to have a history of diabetes-related complications than GLP-1 RA users. Thus, residual confounding is possible. Moreover, because insulin has been shown to cause a transient worsening of diabetic retinopathy (9), the results could also be a reflection of the increased risk of the comparator drug rather than a decreased risk associated with GLP-1 RAs.

Our study has several strengths. First, with a cohort of more than 77,000 newly treated patients with type 2 diabetes and close to 11,000 events, we had the statistical precision required to robustly assess this important safety question. Second, the use of a time-varying exposure definition eliminated the risk of immortal time bias (33) and was deemed to be an appropriate exposure definition given the dynamic nature of pharmacotherapy in type 2 diabetes. Third, the use of a base cohort eliminated left truncation, thereby allowing us to precisely assess important clinical characteristics, including duration of treated diabetes. Moreover, this method minimized the inclusion of prevalent users (34), because all patients entering the base and study cohorts were required to be new users of antidiabetic drugs.

This study also has some limitations. First, due to its observational nature, it is susceptible to potential residual confounding. However, with treatment guidelines recommending GLP-1 RAs as a second- or third-line therapy (1), the use of oral antidiabetic drug combinations as an active comparator likely reduced this bias while providing a clinically meaningful treatment comparison. Furthermore, to mitigate this potential bias, our models included several variables such as HbA_1c level and BMI. Inclusion of these variables along with more than 30 other potential confounders had a minimal effect on the point estimate (crude HR 0.92 vs. adjusted HR 1.00). Moreover, we conducted several sensitivity analyses, such as the use of the high-dimensional disease risk score, which included more than 500 variables and yielded consistent findings.

Second, prescriptions in the CPRD represent those issued by primary care physicians and have not been previously validated. Thus, misclassification of exposure is possible if patient adherence was low or if they were treated by specialists. However, such exposure misclassification is unlikely to have been differential between the exposure groups, because the prescribing of GLP-1 RAs is not restricted to specialists in the U.K. (35).

Third, diagnostic codes of diabetic retinopathy have not been validated in the CPRD and HES. However, since 2003, all U.K. citizens with diabetes are invited for annual retinopathy screens as part of the Diabetic Eye Screening Program (36). Moreover, from 2004 to 2014, the recording of diabetic retinopathy was remunerated as part of the Quality and Outcomes Framework of the National Health Service, a reward and incentive program detailing primary care practice (37). As such, we expect the recording of diabetic retinopathy in the CPRD to have been high during the study period. Furthermore, our overall rate and patterns of diabetic retinopathy according to duration of treated diabetes are comparable to those previously reported (Supplementary Fig. 3) (38,39).

Finally, assessing the association between GLP-1 RAs and various severities of diabetic retinopathy (e.g., nonproliferative vs. proliferative disease) was not possible because these are not well distinguished in the CPRD and HES. Therefore, assessing whether GLP-1 RAs are associated with a worsening of existing diabetic retinopathy will need to be investigated in other settings where this information is routinely recorded.

In summary, the results of this large population-based study indicate that the use of GLP-1 RAs is not associated with an increased risk of diabetic retinopathy overall. Although there was a suggestion of a transient increased risk with durations of use ranging between 6 and 12 months, this finding needs to be interpreted with caution. The apparent lower risk of diabetic retinopathy associated with GLP-1 RAs compared with insulin may be due to residual confounding.

Funding. A.D. received a Research Fellowship from the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG). K.B.F. holds a Canadian Institutes of Health Research New Investigator Award. This research was funded by a Canadian Institutes of Health Research Foundation Scheme grant and a New Investigator Award to K.B.F. J.A.U. holds a Heart and Stroke Foundation of Canada National New Investigator/Ontario Clinician Scientist (Phase I) Award. L.A. holds a Chercheur-Boursier Senior Award from the Fonds de recherche du Québec – Santé (FRQS) and is the recipient of a William Dawson Scholar award from McGill University. All authors have completed the International Committee of Medical Journal Editors uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare this study was funded by the Canadian Institutes of Health Research. The sponsors were not directly involved in the design and conduct of the study; the collection, management, analysis, and interpretation of the data; or the preparation, review, or approval of the manuscript.

Duality of Interest. M.E. is a consultant on the finasteride and sexual dysfunction litigation. J.A.U. has consultant or advisory relationships with Amgen, Boehringer Ingelheim, Janssen, Merck, Novartis, and Sanofi Pasteur; has received research support from Novartis; and has received honoraria for sponsored lectures from Janssen. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. A.D. drafted the manuscript. A.D., K.B.F., H.Y., O.H.Y., M.E., J.A.U., and L.A. contributed to study concept and design, to the analysis and interpretation of data, and to critical revision of the manuscript for important intellectual content. A.D., H.Y., and L.A. contributed to statistical analysis. L.A. acquired data, obtained funding, and supervised the study. L.A. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.