Glucagon Like Peptide 1 Receptor Agonists and Age-Related Macular Degeneration Incidence

Abstract

Purpose

To evaluate the association between glucagon-like peptide-1 receptor agonist (GLP-1RA) use and the incidence of age-related macular degeneration (AMD) and neovascular AMD (nAMD) among adults with diabetes.

Design

Retrospective, propensity score-matched cohort study.

Participants

Adults with diabetes initiating a GLP-1RA or another non-insulin antidiabetic agent. In Epic Cosmos, 2,797,686 GLP-1RA users were matched 1:1 to 2,797,686 active comparators.

Methods

We analyzed de-identified electronic health record data from Epic Cosmos and independently replicated the primary analyses in the University of California Health Data Warehouse. Individuals with prevalent AMD or nAMD and those with prior anti-vascular endothelial growth factor treatment were excluded. One-to-one nearest-neighbor propensity score matching balanced demographics, diabetes duration, comorbidities, smoking, diabetic retinopathy, obesity, and eye-care utilization. Cumulative incidence at 1, 5, and 10 years was estimated using Kaplan-Meier methods, and adjusted associations were evaluated using multivariable Cox proportional hazards models.

Main Outcome Measures

Incident any AMD and incident nAMD.

Results

In the matched Epic Cosmos cohort (total N=5,595,372), GLP-1RA use was associated with lower cumulative incidence of any AMD at 1 year (0.187% [95% CI, 0.182%–0.192%] vs 0.294% [95% CI, 0.287%–0.301%]), 5 years (0.924% [95% CI, 0.907%–0.941%] vs 1.278% [95% CI, 1.260%–1.296%]), and 10 years (5.561% [95% CI, 4.719%–6.549%] vs 6.540% [95% CI, 6.017%–7.108%]); all P<.001. Unadjusted cumulative incidence of nAMD was also lower among GLP-1RA users at 1 year (0.047% vs 0.065%), 5 years (0.280% vs 0.340%), and 10 years (1.874% vs 2.393%); all P<.001. In multivariable Cox models, GLP-1RA use was independently associated with lower hazard of any AMD (hazard ratio [HR], 0.84; 95% CI, 0.83–0.86; P<.001) and was not associated with increased hazard of nAMD (HR, 1.00; 95% CI, 0.97–1.04; P=.79). In the independent University of California Health Data Warehouse cohort, GLP-1RA use was likewise associated with lower AMD incidence at 5 and 10 years, whereas nAMD incidence did not differ significantly between groups.

Conclusions

Among adults with diabetes, GLP-1RA use was associated with lower incident AMD and was not associated with increased nAMD risk. These findings provide reassurance regarding ocular safety and support a potential protective association for earlier AMD phenotypes.

Precis

In this cohort of over 5.5 million adults with diabetes, glucagon like peptide-1 receptor agonist use was associated with a lower incidence of age-related macular degeneration, and there was no increased risk of neovascular age-related macular degeneration.

Introduction

Age-related macular degeneration (AMD) is a leading cause of irreversible vision loss among older adults.¹ As global populations age, the prevalence of AMD is expected to rise substantially, with estimates projecting more than 280 million affected individuals worldwide by 2040.^2,3 The pathophysiology of AMD involves a combination of oxidative stress, mitochondrial dysfunction, and chronic inflammation in the retina.^1,2,4

Glucagon-like peptide-1 receptor agonists (GLP-1RAs), widely prescribed in type 2 diabetes mellitus (DM) and weight loss, modulate several inflammatory pathways and may attenuate retinal inflammation and protect against neurodegeneration.^5,6 Therefore, there is growing interest in the anti-inflammatory effects of GLP-1RAs on the risk of ocular diseases. Studies have yielded mixed results, with some reporting increased risk rather than protective effects. A population-based cohort study from Canada reporting on more than 130,000 adults with DM suggested that GLP-1RA use was associated with a more than twofold increased risk of neovascular AMD (nAMD) over three years.⁷ They used data from a single province in Canada, limiting its sample size and generalizability, and it did not have an active comparator group. Despite its limitations, it generated substantial lay media coverage and growing concern among patients about the use of GLP-1RAs and potential ophthalmic implications.

Importantly, there is no established biological mechanism by which GLP1-RAs would increase AMD risk. The discrepancy between mechanistic expectations and reports of increased risk highlights the need for rigorous, comparator-based analyses in diverse populations. Given the benefits of GLP1-RAs and their growing use, clarifying ocular safety carries strong public health importance. Observational signals without mechanistic support have led to patient anxiety and inappropriate avoidance of effective therapies; therefore, a careful, bias-resistant evaluation is warranted.

While there are 6 additional publications investigating this research question^13–17,20 (Table 3), 5 of them used Trinet X and therefore reporting on the exact same population. The only other paper on this topic used a single center database with limited power in multivariable analysis. Our study aims to evaluate the association between GLP-1RA use and the incidence of any AMD and nAMD and it improves on prior work in several ways. First, we leverage two large databases, the multi-system global electronic health records (EHR) network Cosmos (Epic Systems, Verona, WI) and the University of California Health Data Warehouse (UCHDW)⁸, providing the largest reported sample size, broader geographic, racial/ethnic, and clinical diversity than single province-datasets. Second, we used an active comparator design with matching on comorbidities and healthcare utilization to minimize bias. Third, we collected retrospective data of patients for up to 10 years after the medication exposure, enabling estimates of absolute risks at clinically relevant horizons (1, 5, 10 years).

Table 3:

Review of published literature on the association of GLP-1 RAs and AMD.

Reference	Journal	GLP-1 RA Indication	Publication Date	Sample Size	Database	HR/RR
7	JAMA Ophthalmology	Diabetes	Jul 2025	139,002	Comprehensive administrative health and demographic data from patients in Ontario, Canada	nAMD= 2.21
13	Ophthalmology	Diabetes	Jul 2025	9,669	TrinetX	nAMD= 0.7
17	Retina	Diabetes	Jul 2025	600,816	TrinetX	AMD = 0.715, nAMD = 0.729
20	JAMA Ophthalmology	Weight Loss	Oct 2025	91,408	TrinetX	AMD= 0.09–0.016
16	Ophthalmology Retina	Diabetes or Weight Loss	Nov 2025	< 88,000	TrinetX	AMD= 0.75–0.84
15	Ophthalmology Retina	Diabetes or Weight Loss	Dec 2025	30,244	Cole Eye Institute Database	AMD (RR vs. insulin) = 0.45
14	Ophthalmology Retina	Weight Loss among Diabetic Patients	Dec 2025	20,959	TrinetX	nAMD = 0.47 AMD = 0.61

AMD = age-related macular degeneration, nAMD = neovascular age-related macular degeneration, GLP-1RA = glucagon like peptide 1 receptor agonist

Methods

We designed the analysis pipeline in UCHDW and implemented the same analysis in Epic Cosmos database to assess the robustness and generalizability of the findings.^8–10 At the time of this analysis, UCHDW consisted of the EHR and medical claims information from 9.8 million individuals.¹⁰ The Epic Cosmos data warehouse, a dataset created in collaboration with a community of health systems using Epic, represents more than 300 million patient records from over 1,762 hospitals and 40,700 clinics as of August 2025. The community represents patients from all 50 states, D.C., Canada, Lebanon, and Saudi Arabia. Results of the analysis on the UCHDW are reported in Supplemental File 1. Last follow-up was on March 04, 2026. Statistical analysis was conducted from July to August 2025. The University of California, San Diego Institutional Review Board considered this study not human subjects research and granted a waiver of informed consent. This study adhered to the tenets of the Declaration of Helsinki. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

Study Design and Cohort Selection

This study was a matched retrospective cohort study. The GLP-1RA group included individuals with a history of any DM who initiated any GLP-1RA and maintained therapy for at least six months. The “other antidiabetic” group included individuals with DM who initiated any other antidiabetic, other than insulin, with at least six months of continuous use, and no documented history of GLP-1RA intake during the entirety of follow-up observation. The first 6-month or more exposure period was used to define index medication use, but patients were subsequently followed longitudinally for up to 10 years to assess incident AMD and nAMD outcomes. Date of first exposure to GLP-1RAs was used as the index date in the analysis. Exclusion criteria were pre-existing diagnosis of AMD/nAMD prior to the index date, any history of anti-vascular endothelial growth factor treatment prior to the index date.

Cohort Identification and Sample Size Calculation

Our initial query of individuals with diabetes identified 4,116,189 eligible GLP-1RA users and 5,903,952 eligible individuals in the other antidiabetics group (Figure 1). Eyes in the GLP-1RA group were excluded from the other antidiabetics group. We performed propensity score matching, to ensure balanced comparison. Propensity scores were calculated using logistic regression incorporating the following baseline covariates: hypertension, coronary artery disease, cerebrovascular disease, chronic kidney disease, diabetic retinopathy, age, duration of diabetes mellitus, obesity, smoking, race, ethnicity, and sex. Obesity was defined as BMI≥30 or any diagnosis of obesity within a year from the index date. The number of ophthalmology or optometry visits before and after the index date was also included to account for differential likelihood of AMD or nAMD diagnosis based on healthcare utilization.¹¹ One-to-one nearest neighbor matching without replacement and with a caliper of 1×10⁻⁵ SD of the logit of the propensity score was applied, indicating minimal probability of matching issues. This resulted in 2,797,686 individuals in each group, totaling 5,595,372 individuals. The primary outcomes were the incidence of any AMD and nAMD. These outcomes were evaluated at 1, 5, and 10 years of follow-up.

Figure 1:

Study flowchart demonstrating the query and study cohorts; GLP-1RA: glucagon like peptide-1 receptor agonists.

Statistical Analysis

Demographic information on race and ethnicity was obtained through self-report according to local site procedures. When feasible, these data were subsequently harmonized to the standard categories defined by the US Centers for Disease Control and Prevention.¹² Summary statistics were reported as medians with interquartile ranges (IQR). The quality of matching between the GLP-1RA and other antidiabetics groups was assessed using standardized mean differences (SMDs). Incidence rates for AMD and nAMD were calculated and compared between groups, with 95% confidence intervals estimated using 1000 bootstrap resampling to preserve within-pair dependence and avoid conditioning on observed person-time inherent to standard Poisson CIs.Log-rank tests were used only for comparison of unadjusted event proportions among participants observed at each time point, and number of events and at-risk individuals were reported along with the Kaplan-Meier (KM) estimated incidence rate. Because these cross-sectional comparisons do not account for censoring, all formal inference on time-to-event outcomes was based on KM survival and multivariable Cox regression models, which appropriately handle dropout and variable follow-up. Because of this large proportion of missingness, HbA1c was compared descriptively among individuals with available results and was not used in propensity score matching or outcome models. To account for multiple comparisons across outcomes and time points (6 primary tests: 2 outcomes × 3 time points), we considered p-values < .0083 (Bonferroni-adjusted) to indicate statistical significance. Two-sided p-values <.05 were considered statistically significant for the survival analyses. As a sensitivity analysis to address potential nAMD outcome misclassification, we compared post-index anti-VEGF treatment rates between groups. Sub analyses of AMD and nAMD incidence were performed for each medication which were considered exploratory. Sequential Cox models were fit as sensitivity analyses to evaluate the impact of demographic factors, follow-up duration, and additional covariables on the estimated association between GLP-1RA exposure and nAMD incidence. The databases were queried using structured query language, and the analysis was performed using Python (version 3.12, Python Software Foundation, Wilmington, DE, USA). Cells with fewer than 11 patients or events were suppressed in accordance with data-use policy.

Results

Baseline Characteristics

Table 1 presents the baseline characteristics of participants in the GLP-1RA and other antidiabetics groups after matching. A total of 2,797,686 individuals were included in each group. The median age (GLP-1RA: 60 [IQR: 17] years vs. other antidiabetics: 59 [IQR: 20] years; SMD: 0.046) and duration of diabetes mellitus (GLP-1RA: 927 [IQR: 1824] days vs. other antidiabetics: 528 [IQR: 1743], SMD: 0.062) were similar between the groups. The distribution of sex was similar between the groups (GLP-1RA: 1,454,824 [52%] female vs. 1,477,233 [52.8%] female: other antidiabetics group). The study cohorts were well balanced (Table 1). Median HbA1c was similar between the GLP-1RA group (7.4 [IQR: 0.6]) and other antidiabetics group (7.4 [IQR: 0], p=0.43). HbA1c measurement within a year of the index date was available in 55.5% (n=1,552,848) of the individuals in the GLP-1RA group and 39.4% (n=1,103,484) of the individuals in the other antidiabetics group. The median follow-up period was shorter in the GLP-1RA group (2.95 [IQR: 3.27] years vs. 4.44 [IQR: 5.57] years).

Table 1:

Baseline characteristics of the participants in the GLP-1RA and the other antidiabetics groups.

Variable	GLP-1RA	Other antidiabetics	SMD
Total N	2,797,686	2,797,686
Age, median [IQR] years*	60 [17]	59 [20]	0.046
Diabetes Mellitus Duration, median [IQR] days	927 [1,743]	528 [1,834]	0.062
Body Mass Index, Mean (SD)	35.185 (7.244)	35.506 (7.549)	0.043
Hypertension, n (%)	2,154,883 (77)	2,161,069 (77.2)	0.005
Coronary Artery Disease, n (%)	526,305 (18.8)	522,395 (18.7)	0.004
Cerebrovascular Disease, n (%)	256,900 (9.2)	256,161 (9.2)	0.001
Chronic Kidney Disease, n (%)	452,278 (16.2)	457,742 (16.4)	0.005
Diabetic Retinopathy, n (%)	126,738 (4.5)	123,962 (4.4)	0.005
Obesity, n (%)	2,187,845 (78.2)	2,242,048 (80.1)	0.048
Current Smoker, n (%)	78,310 (2.8)	83,283 (3)	0.011
Former Smoker, n (%)	112,114 (4)	118,024 (4.2)	0.011
Race a : American Indian or Alaska Native, n (%)	36,244 (1.3)	36,856 (1.3)	0.002
Race: Asian, n (%)	92,846 (3.3)	90,128 (3.2)	0.005
Race: Black or African American, n (%)	484,718 (17.3)	482,923 (17.3)	0.002
Race: Native Hawaiian or Other Pacific Islander, n (%)	18,441 (0.7)	18,655 (0.7)	0.001
Race: Other b , n (%)	391,769 (14)	343,421 (12.3)	0.051
Race: Unknown, n (%)	339,415 (12.1)	332,734 (11.9)	0.007
Race: White, n (%)	1,789,817 (64)	1,790,624 (64)	0.001
Ethnicity: Hispanic or Latino, n (%)	305,412 (10.9)	303,385 (10.8)	0.002
Ethnicity: Unknown, n (%)	401,971 (14.4)	406,814 (14.5)	0.005
Sex: Female, n (%)	1,454,824 (52)	1,477,233 (52.8)	0.016
Sex: Male, n (%)	1,342,618 (48)	1,320,069 (47.2)	0.016
Sex: Other, n (%)	6 (0)	21 (0)	0.002

IQR = interquartile range; SMD = standardized mean difference.

^aRace is patient-reported according to procedures that vary by source organization (e.g., organizations may use closed or open categories), then standardized to US Centers for Disease Control and Prevention (CDC) categories within the Epic Cosmos data pipeline. Patients were able to report more than one race, and all reported races are represented (i.e., categories may overlap).

^bIncludes reported races that could not be mapped to CDC categories.

Incidence of AMD and nAMD and Time to Event Analysis

Table 2 summarizes the incidence rates of AMD and nAMD at 1, 5, and 10 years of follow-up. For any AMD, significant lower incidence was observed in the GLP-1RA group at all time points. At 1 year, KM estimated incidence of AMD was 0.187% (95% CI: 0.182–0.192) in the GLP-1RA group and 0.294% (95% CI: 0.287–0.301) in the other antidiabetics group (log-rank p = <.001). At 5 years, KM estimated incidence was 0.924% (95% CI: 0.907–0.941) in the GLP-1RA group and 1.278% (95% CI: 1.260–1.296) in the other antidiabetics group (log-rank p = <.001). At 10 years, KM estimated incidence was 5.561% (95% CI: 0.4.719–6.549) in the GLP-1RA group versus 6.540% (95% CI: 6.017–7.108) in the other antidiabetics group (log-rank p < .001). Similarly, there were significant differences in nAMD incidence between GLP-1RA and antidiabetics groups at any time point. At 1 year, KM estimated incidence of nAMD was 0.047% (95% CI: 0.044–0.050) in the GLP-1RA group and 0.065% (95% CI: 0.062–0.069) in the other antidiabetics group (log-rank p = <.001). At 5 years, KM estimated incidence was 0.280% (95% CI: 0.2710.290) in the GLP-1RA group versus 0.340% (95% CI: 0.331–0.350) in the other antidiabetics group (log-rank p = <.001). At 10 years, KM estimated incidence was 1.874% (95% CI: 1.4752.381) in the GLP-1RA group compared to 2.393% (95% CI: 2.057 −2.785) in the other antidiabetics group (log-rank p = <.001). In the independent UCHDW cohort, GLP-1RA use was also associated with lower crude AMD incidence at 5 and 10 years (5 years: 1.08% vs 1.48%, p = 0.001; 10 years: 1.23% vs 1.74%, p < .001), while crude nAMD incidence did not differ significantly between groups at 1, 5, or 10 years (Supplemental File 1).

Table 2.

Incidence of age-related macular degeneration (AMD) and neovascular age-related macular degeneration (nAMD) at 1, 5, and 10 years

Time point	AMD, GLP-1RA	AMD, Other antidiabetics	p value	nAMD, GLP-1RA	nAMD, Other antidiabetics	p value
1 year	0.187% (95% CI, 0.182–0.192) Events: 4,834; At risk: 2,383,362	0.294% (95% CI, 0.287–0.301) Events: 6,818; At risk: 2,170,244	<.001	0.047% (95% CI, 0.044–0.050) Events: 1,212; At risk: 2,385,375	0.065% (95% CI, 0.062–0.069) Events: 1,511; At risk: 2,169,464	<.001
5 years	0.924% (95% CI, 0.907–0.941) Events: 14,586; At risk: 584,800	1.278% (95% CI, 1.260–1.296) Events: 21,245; At risk: 865,779	<.001	0.280% (95% CI, 0.271–0.290) Events: 4,235; At risk: 587,277	0.340% (95% CI, 0.331–0.350) Events: 5,454; At risk: 869,581	<.001
10 years	5.561% (95% CI, 4.719–6.549) Events: 18,107; At risk: 325	6.540% (95% CI, 6.017–7.108) Events: 29,283; At risk: 834	<.001	1.874% (95% CI, 1.475–2.381) Events: 5,456; At risk: 325	2.393% (95% CI, 2.057–2.785) Events: 8,290; At risk: 841	<.001

AMD = age-related macular degeneration; nAMD = neovascular age-related macular degeneration; GLP-1RA = glucagon-like peptide-1 receptor agonist; CI = confidence interval.

Time to Event Analyses

Kaplan-Meier survival analysis demonstrated significantly longer time to AMD development (log-rank p < .001; Figure 2a) and nAMD (log-rank p < .001; Figure 2b) in the GLP-1RA group compared with other antidiabetics group. In multivariable Cox regression, GLP-1RA treatment was independently associated with reduced risk of AMD (HR, 0.84 [95% CI, 0.83–0.86]; p < .001). (Figure 3a). In the multivariable model, GLP-1RA exposure was not independently associated with the risk of nAMD (HR, 1.00 [95% CI, 0.97–1.04]; p = .79) (figure 3b). Figure 3 shows the full set of covariables included in the Cox models and their corresponding HRs (95% CIs).

Figure 2:

Kaplan Meier survival curves comparing time to developing a) age related macular degeneration and b) neovascular age related macular degeneration between the glucagon like peptide-1 receptor agonists and other antidiabetics groups; GLP-1RA: glucagon like peptide-1 receptor agonists.

Figure 3:

Forest plots demonstrating the results of the multivariable Cox proportional hazards analysis of a) age-related macular degeneration outcome and b) neovascular age related macular degeneration outcome; GLP-1RA: glucagon like peptide-1 receptor agonists.

In sequential analyses, attenuation of the nAMD association after demographic adjustment was driven primarily by age; however, after accounting for follow-up duration, the estimate shifted back toward a protective association (age+sex+follow-up model: GLP1-RA HR, 0.670; 95% CI, 0.560–0.803; p<0.001), and the fully adjusted follow-up-accounting model likewise showed lower nAMD hazard among GLP-1RA users (HR, 0.70; 95% CI, 0.67–0.72; p<0.001).

Dose Response relationship

Dose-stratified analyses of AMD-free survival suggested that the relationship between dose and AMD risk was not uniform across agents (figure 4). For dulaglutide, liraglutide, and tirzepatide, higher-dose strata demonstrated better AMD-free survival over follow-up (all log-rank p<0.001). In contrast, semaglutide showed inconsistent patterns across oral and injectable formulations, and exenatide did not demonstrate a statistically significant dose-stratified difference (log-rank p=0.073).

Figure 4:

Kaplan Meier survival curves demonstrating protective dose-response relationship across different dose groups of glucagon-like peptide-1 receptor agonists.

Sensitivity and Subgroup Analyses

Exploratory AMD and nAMD incidence results across dose categories are shown in supplemental file 2, Table S2_2. In additional agent-specific analyses comparing each GLP-1RA agent with its own matched control cohort, associations with AMD and nAMD were heterogeneous across agents and are summarized in Table S2_3. We observed nearly identical rates of post-index anti-VEGF treatment in the GLP-1RA and comparator cohorts (0.296% vs 0.298%; p=0.732). To further evaluate the validity of the nAMD phenotype definition, we examined post-index anti-VEGF treatment across phenotype-defined cohorts. Anti-VEGF treatment occurred substantially more often among patients classified as nAMD than among those with AMD overall (p<0.001), supporting that the nAMD definition captured clinically treated neovascular disease. Only 345 of 50,243 (0.7%) individuals in the AMD group received anti-VEGF agents without ever receiving an nAMD diagnosis code.

Discussion

In this study using data from Epic Cosmos, a large, global real-world database and independently validated in a second multi-center database, GLP-1RA use was associated with lower AMD incidence over 1 , 5 , and 10 years of follow-up compared with other antidiabetic medications. Survival analysis showed a potential independent association with reduced AMD risk, whereas when accounting for other covariates, GLP-1RA exposure was not associated with nAMD incidence. In the independent UCHDW cohort, a similar pattern was observed for AMD, while nAMD incidence did not differ significantly between groups. These findings suggest that long-term GLP-1RA use was not associated with increased incidence of AMD or nAMD and may be associated with reduced incidence in individuals with DM. To alleviate concerns about coding practices and inter-eye correlation, we conducted the analysis at the individual level and used first reported any AMD (both nonexudative and neovascular) and nAMD as the primary outcome measure. Previous studies have explored the relationship between GLP-1RA use and ocular outcomes, though results have been inconsistent.

A population-based cohort study from Canada reported that GLP-1RA use was associated with a more than twofold increased risk of incident nAMD (adjusted HR 2.21) among individuals aged 66 and older with type 2 diabetes.⁷ This study raised concerns about a potential pro-angiogenic association of GLP-1RAs in susceptible populations. In contrast, the current study did not observe a higher incidence of nAMD among GLP-1RA users. Differences in age, data sources, and population diversity between the two studies may explain the divergent findings. Shor et al. analyzed older adults using administrative claims data from a single Canadian province without an active comparator and with limited adjustment for surveillance intensity, factors that increase residual confounding and reduce generalizability.⁷ More recently, a U.S. Trinet X-based study associated GLP-1RA use with a lower incidence of nonexudative AMD compared to multiple comparator drugs, including metformin, insulin, and statins; in each comparison, GLP-1RA use was associated with a significantly lower incidence of nonexudative AMD.¹³ However, that analysis relied on claims-based definitions, required prolonged exposure windows, and used platform-level data with limited transparency in site-level heterogeneity and medication adherence.¹³ Consistent with the present findings, Myers et al., demonstrated that prescriptions for GLP-1 RAs were associated with a lower incidence of nonexudative AMD diagnosis among non-diabetic patients who were eligible for weight loss pharmacotheraphy.¹⁴ In comparison to Myers et al.’s study, which did not include any analyses of smoking status, our study accounted for former or current smoking status and had a much larger sample size. Interestingly, another single center study from Cleaveland clinic demonstrated an association between GLP1 RA use and lower crude relative risk of non-exudative AMD compared with using alternative antidiabetics in patients older than 50 years.¹⁵ Of note, their multivariable Cox proportional hazards analysis only showed significant protective association when compared GLP-1RA with insulin.

Consistently, Allan and colleagues found that prescription of GLP1 RAs may provide protective effects against the incidence of non-exudative AMD compared with other antihyperglycemic agents and antihyperlipidemic agents in patients above 60 years old using Trinet X platform.¹⁶ McLaughlin et al. used Trinet X database with smaller matched cohorts (9,669 per arm) and multiple ocular outcomes, which can amplify multiplicity and selection biases, yet also reported protective associations for AMD and nAMD.¹⁷ Our findings are consistent with these latter results, suggesting a potential protective association between GLP-1RA use and AMD incidence in real-world settings. However, in our study, the evidence was stronger and more consistent for AMD than for nAMD, as the adjusted nAMD association was null and the independent UCHDW cohort also showed no significant between-group difference in nAMD incidence. Compared with prior studies, our analysis focused specifically on AMD and nAMD, included longer follow-up, and evaluated 2,797,686 in each group, providing substantial power to estimate small between-group differences. While there are 6 additional publications other than the Canadian study investigating this research question, 5 of them used Trinet X and therefore reporting on the exact same population. The only other paper on this topic used a single center database with limited power in the multivariable analysis (Table 3).

The differing results for nAMD and any AMD may be due to both biological and data-related factors. While early AMD is primarily characterized by drusen accumulation, retinal pigment epithelium dysfunction, and chronic inflammation, nAMD involves additional angiogenic processes such as choroidal neovascularization driven by hypoxia and inflammatory cytokines.⁴ GLP-1RAs have demonstrated anti-inflammatory and neuroprotective properties in preclinical retinal models,^5,18,19 which may help reduce the incidence of non-neovascular AMD. Another possibility is outcome misclassification. Neovascular AMD diagnoses often rely on imaging findings and treatment history, but diagnostic and billing codes alone may not reliably capture these distinctions. Although there may have been diagnostic misclassification in some cases, it is expected to be similar in both groups. A treatment-based sensitivity analysis also showed no meaningful between-group difference in anti-VEGF use, which provides supportive evidence that differential ascertainment of clinically treated neovascular retinal disease is unlikely to explain the findings. However, because anti-VEGF therapy in patients with diabetes is not specific to nAMD, this analysis should be interpreted as supportive rather than definitive. In exploratory grouped-dose analyses, some agents, particularly dulaglutide and tirzepatide, showed numerically lower AMD incidence in higher dose categories, although these patterns were not uniformly monotonic and should be interpreted cautiously. Potential misclassification between nAMD and diabetic macular edema (DME) is another limitation of structured EHR-based phenotyping in patients with diabetes. As an internal check, we examined incident DME and identified 81,455 cases in the GLP-1RA group and 70,254 cases in the comparator group after index. While some diagnostic overlap cannot be excluded, these findings suggest DME was captured as a distinct retinal outcome and is unlikely to fully explain the between-group nAMD findings.

This study has several limitations. We used retrospective data from EHR to define outcomes, without confirmatory clinical validation of AMD diagnoses, since imaging data were not available. Moreover, the study lacked information on some key risk factors such as family history, and genetic predisposition. However, these elements are rarely available in large claims-or EHR-based cohorts, particularly genetic data, and prior population studies shared these limitations while often being smaller and less diverse compared to our study. The follow up period was expectedly shorter in the GLP-1RA group. To address this issue, we implemented time to event analyses using KM and Cox Proportional Hazards models which inherently account for loss to follow up. In addition, some of the centers across the University of California contribute data to Cosmos as well, so some overlap between the two databases should be considered when comparing between the two databases. AMD represents a heterogeneous set of conditions with overlapping but distinct biological pathways, making generalization across populations and drug classes inherently difficult.²

Another important consideration is differential time in clinical use across GLP-1–based therapies: exenatide has been available since 2005, liraglutide since 2010, dulaglutide since 2014, lixisenatide since 2016, semaglutide since 2017 (oral formulation since 2019), and tirzepatide since 2022. Accordingly, apparent differences in safety signals may partly reflect unequal cumulative post-marketing exposure, with newer agents having substantially shorter real-world follow-up. Despite the limitations, our study had multiple strengths. We used data from a global database of real-world routine care data, providing one of the largest sample sizes in the field, and to our knowledge based on a search of the existing literature, the largest sample size to date for any study examining GLP-1RA use and ophthalmic complications. Further, we validated our results in a large, ethnically diverse population within an integrated U.S. health system spanning numerous clinical sites, with longitudinal follow-up of up to 10 years. Medication use was defined using structured prescription data, and propensity score matching with a caliper of 1×10⁻⁵ SD (indicating powerful matching) and use of an active comparator cohort was applied to reduce confounding. The greater hazard with age, female sex, hypertension, history of diabetic retinopathy, and White race have all been known risk factors for AMD and nAMD. Lower hazard observed among Asian or Black/African American racial groups and Hispanic/Latino participants is also directionally consistent with epidemiologic data showing lower AMD risk in populations with greater ocular pigmentation relative to non-Hispanic White cohorts. The concordant directionality supports the construct validity of our models and does not alter the null association between GLP-1RA exposure and nAMD incidence. We replicated the primary analyses in two databases (i.e., Cosmos and the UCHDW) with broadly similar findings for AMD and no independent signal suggesting increased nAMD risk in the external cohort, supporting robustness to database-specific coding or case-mix differences. Our study also analyzes several million adults, whereas prior AMD GLP-1RA cohorts were typically <140,000 (Table 3), enabling precise absolute risk estimates with greater power. Moreover, two independent analysts implemented the analyses separately in each database, adapting the code to the local data model. Concordant results across these parallel implementations reduce pipeline-specific bias and further strengthen the reproducibility of our findings.

In conclusion, GLP-1RA use was associated with a lower risk of overall AMD and was not associated with increased risk of nAMD in adults with DM. These findings may offer reassurance regarding the ocular safety of GLP-1RAs and suggest a possible association with protection in the early retinal changes. Given the growing use of GLP-1RAs, further prospective studies that incorporate retinal imaging, and drug-specific analyses, and evaluating the association among non-diabetic individuals will be important to clarify their differential association with different stages of AMD. Potential ocular adverse effects of GLP-1 receptor agonists should be interpreted in the context of a generally favorable ocular safety profile. The most consistently reported signal is transient worsening of diabetic retinopathy, particularly in patients with pre-existing retinopathy and rapid glycemic improvement after treatment initiation, which likely reflects an “early worsening” phenomenon rather than direct retinal toxicity. Rare post-marketing and observational reports have also raised concern for nonarteritic anterior ischemic optic neuropathy with semaglutide, but the absolute risk appears very low and causality remains uncertain. Overall, current evidence does not support a broad class-related increase in ocular toxicity, though closer ophthalmic monitoring may be reasonable in high-risk patients early after treatment initiation.

Data Sharing Statement:

The databases used in this study are available only for investigators from institutions that contribute data. Otherwise, sharing the data openly is prohibited by the license agreement of each database.