GLP-1RA and the Risk of Non-Arteritic Anterior Ischemic Optic Neuropathy in Patients with Type 2 Diabetes: A Population-based Study

Abstract

Objective:

Evidence on the safety of semaglutide and other glucagon-like peptide-1 receptor agonists (GLP-1RA) concerning non-arteritic anterior ischemic optic neuropathy (NAION) is inconclusive, with several studies published to date presenting methodological challenges. We sought to compare the risk of presumed NAION in patients with type 2 diabetes (T2D), initiating a GLP-1RA versus a sodium-glucose cotransporter-2 inhibitor (SGLT2i), using a new-user active comparator cohort study design within a target trial emulation framework.

Research Design and Methods:

Using insurance claims data from three databases (01/2016–08/2024), we identified 482,912 propensity score matched (1:1) pairs of GLP-1RA and SGLT2i adult initiators with T2D and without prior ischemic optic neuropathy (ION) or other optic nerve disorders. NAION was defined as a diagnosis of ION without other causes of optic neuropathy, combined with an ophthalmologist or optometrist visit on the same day. We estimated pooled hazard ratios (HRs) and rate differences (RDs).

Results:

Over a median follow-up of 6.7 months on treatment, the risk of incident NAION was higher among initiators of GLP-1RA compared with SGLT2i [HR, 1.85; 95% CI (1.51–2.27); RD, 0.29 (0.19, 0.39) per 1,000 person-years]. Results were consistent across subgroups and sensitivity analyses. A meta-analysis of our semaglutide analysis with previously published results also showed an elevated risk, with a HR of 2.78 (1.39–5.56).

Conclusions:

In this large observational study, the initiation of GLP-1RA was associated with an 85% increase in the risk of presumed NAION, compared with the initiation of SGLT2is, though incidence rates and absolute increase in risk were small.

INTRODUCTION

Despite the widespread use of glucagon-like peptide 1 receptor agonist (GLP-1RA) for the treatment of type 2 diabetes (T2D) and chronic weight loss management, there have been relatively few permanent, medically significant complications with GLP-1RA. Recently, the safety of GLP-1RAs, primarily semaglutide, has been questioned with respect to the risk of non-arteritic anterior ischemic optic neuropathy (NAION), a form of optic neuropathy which can cause acute blindness in adults.^1,2 This concern was raised by a single referral-center study by Hathaway et al., which found that semaglutide, compared with non-GLP-1RA medications, was associated with a 4.3-fold increased risk of NAION in patients with T2D and a 7.6-fold increased risk of NAION in patients with overweight or obesity, with the majority of events occurring within 12 months of drug initiation.³

Subsequently, nine studies (a meta-analysis of clinical trials, a case series, and seven population-based cohort studies) have been published to date, reporting no or elevated risk of NAION associated with the use of GLP-1RA, primarily semaglutide.^4–12 A meta-analysis by Silverii et al. evaluated ischemic optic neuropathy (ION) outcomes using serious adverse event data of clinical trials and reported inconclusive findings.¹⁰ A case series by Katz et al. was unable to establish a causal link between exposure to those drugs and the observed ophthalmologic complications.⁹ Of the population-based cohort studies published to date, several presented methodological challenges, including potential prevalent user bias,^{4,6–8,12–14} confounding by indication bias,^15,16 or incomplete capture of longitudinal information for study participants (eTable 1).^4,6,8,12 Only a few studies appeared less prone to these challenges, although they only focused on semaglutide, without evaluating the entire GLP-1RA class or tirzepatide, a dual glucose-dependent insulinotropic polypeptide and GLP1 receptor agonist.^5,11 Refer to eTable 1 for a detailed evaluation of the publications to date.

Given the irreversible vision loss associated with NAION, the growing popularity of GLP-1RA for T2D management, weight loss, and among patients with cardiovascular and kidney disease, and the limitations of the published literature to date, it is critical to further evaluate the previously observed safety signal with respect to NAION using rigorous methodology. Therefore, our study, which includes two authors from the Hathaway study, sought to evaluate the association between GLP-1RAs and the risk of NAION in patients with T2D using a new-user study design within a target trial emulation framework, paired with comparable active comparator, i.e. sodium-glucose cotransporter-2 inhibitors (SGLT2i), with the same indication of use, similar line of treatment and expected access to health care as GLP-1RA.

METHODS

Data Sources

The study was conducted using claims data from two US employer-sponsored insurance databases (Optum’s de-identified Clinformatics^® Data Mart Database (CDM) and Merative Marketscan) and one federal insurance database (Medicare fee-for-service (FFS)). The three databases contain longitudinal, individual-level information on health plan enrollment, demographics, inpatient and outpatient diagnostic and procedural codes, and outpatient pharmacy dispensing records, including quantity filled, strength, and days’ supply. Laboratory test results were available for a subset of beneficiaries (approximately 45%) in CDM. The study protocol was approved by the Mass General Brigham’s Institutional Review Board, and data use agreements were in place for the three databases. Informed consent was waived because the study used de-identified secondary data. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines.

Study Design and Population

Within a target trial emulation framework (eTable 2), we conducted a new-user, active comparator population-based cohort study of patients newly receiving a GLP-1RA or SGLT2i between January 1, 2016 and May 31, 2024 (CDM), September 30, 2022 (Marketscan) and September 30, 2020 (Medicare FFS) (eFigure 1). Eligible patients were individuals with a diagnosis of T2D who were at least 18 years old in commercial databases and 65 years and older in Medicare. We excluded patients with a diagnosis of type 1, secondary or gestational diabetes, end-stage kidney disease, pancreatitis, tirzepatide use, multiple endocrine neoplasia type 2, solid organ transplant, prior history of ischemic optic neuropathy or optic atrophy, and anti-neutrophilic cytoplasmic autoantibody and other causes of vasculitis, including giant cell (or temporal) arteritis. Additionally, we excluded patients who had other conditions associated with optic neuropathy or systemic shock. eTable 3 provides a detailed and complete list of eligibility criteria.

Exposure and Study Follow-Up

Drug exposure was defined using claims for outpatient pharmacy-filled prescriptions of either a GLP-1RA (exenatide, liraglutide, dulaglutide, or semaglutide) or an SGLT2i (canagliflozin, dapagliflozin, empagliflozin, or ertugliflozin) (eTable 4). Albiglutide, lixisenatide, and combination products with insulin were excluded from the definition of exposure to a GLP-1RA because the former two are no longer marketed in the U.S. and the use of insulin in T2D is indicative of a more advanced or severely uncontrolled diabetes.^17,18 Tirzepatide was also excluded from the primary exposure definition because it is a dual GLP-1/glucose-dependent insulinotropic polypeptide receptor agonist. We selected SGLT2is as the comparator because, together with GLP-1RAs, they are guideline-recommended glucose-lowering therapies used at similar stages of diabetes severity and cardiovascular risk.¹⁹ Furthermore, to our knowledge, SGLT2i use has not been associated with an increased risk of NAION.

The cohort entry date was the day of the first filled prescription of a GLP-1RA or SGLT2i after a 365-day washout period with no use of either medication class. We allowed initiators of a GLP-1RA or SGLT2i to enter the study cohort up to 3 months before the end of data availability in each dataset to allow for a minimum follow-up of 90 days. Study follow-up started on the day after the cohort entry date and continued until the earliest occurrence of a NAION outcome, disenrollment from a health plan, death, switching to the other treatment arm, or treatment discontinuation after the allowed 60-day grace period (eFigure 1).

NAION Outcome

There is no specific diagnosis code for NAION. Therefore, our primary definition of presumed NAION was based on inpatient or outpatient International Classification of Diseases, Tenth Revision, Clinical Modification diagnosis codes of ischemic optic neuropathy (H47.01x) combined with an ophthalmologist or optometrist visit on the same day to improve the likelihood of an accurate diagnosis code after excluding other conditions associated with optic neuropathy, as described above and below (eTable 5). A previous validation study found that an algorithm combining the aforementioned ION code with a visit to an ophthalmologist produced a positive predictive value (PPV) of 75.8% (95% CI: 67.2% - 80.7%).²⁰ Our algorithm expanded the provider definition by allowing visits with an optometrist, acknowledging the insufficient availability of ophthalmologists, particularly in non-metro and rural regions of the U.S.²¹ To further improve the specificity of our algorithm, we required the visit to occur on the same day as the date of the diagnosis and also restricted our study cohort to those without any prior history of optic nerve disorders, inflammatory causes of ischemic optic neuropathies such as giant cell arteritis, and diseases that could potentially lead to misdiagnosis of NAION, including demyelinating diseases and uveitis (see eTable 3).

Baseline Characteristics

Separately in each database, we measured baseline characteristics during the 365 days prior to and including the cohort entry date. We selected covariates based on existing literature and clinical expert knowledge. All databases included information on demographics, claims-based measures of disease burden and frailty status, lifestyle factors, proxies of diabetes severity and complications, cardiovascular history, glucose-lowering medication use, proxies of socio-economic status, indicators of healthcare resource utilization such as preventative care, laboratory test orders, and provider visits (Table 1 and eTable 6). Information on hemoglobin A1c (HbA1c) and serum creatinine lab test results were available for a subset of commercially insured enrollees with available laboratory test results (~45%). Estimated glomerular filtration rate (eGFR) was calculated using the CKD Epidemiology Collaboration (CKD-EPI) equation for individuals with creatinine lab values available.²²

Table 1.

Pooled baseline characteristics of study participants before and after 1:1 propensity score matching

	Before PS Matching			After Matching
Characteristic	GLP-1RA (N=685,626)	SGLT2i (N=709,511)	Absolute standardized difference	GLP-1RA (N=482,912)	SGLT2i (N=482,912)	Absolute standardized difference
DEMOGRAPHICS
Age, years; mean (SD)	61.8 (10.4)	64.4 (9.9)	0.258	62.9 (10.3)	62.9 (9.9)	0.004
Non-differentiated sex or gender; n (%)
…Male	312,585 (45.6%)	406,165 (57.2%)	0.234	249,145 (51.6%)	247,484 (51.2%)	0.008
…Female	373,041 (54.4%)	303,346 (42.8%)	0.234	233,767 (48.4%)	235,428 (48.8%)	0.008
Race categories1; n (%)
…White	295,519 (63.5%)	302,006 (62.4%)	0.023	206,912 (64.0%)	206,914 (64.0%)	0.000
…Black	51,776 (11.1%)	53,234 (11.0%)	0.003	35,122 (10.9%)	35,234 (10.9%)	0.000
…Asian	9,934 (2.1%)	22,951 (4.7%)	0.144	8,889 (2.8%)	8,474 (2.6%)	0.012
…Other or unknown	103,813 (22.3%)	98,910 (20.5%)	0.044	68,771 (21.3%)	69,083 (21.4%)	0.002
Hispanic ethnicity2; n (%)	39,016 (13.1%)	45,104 (15.2%)	0.060	28,403 (14.2%)	28,393 (14.2%)	0.000
Geographic region; n (%)
…Northeast	83,810 (12.2%)	99,832 (14.1%)	0.056	62,766 (13.0%)	62,498 (12.9%)	0.003
…South	341,533 (49.8%)	344,433 (48.5%)	0.026	238,440 (49.4%)	238,984 (49.5%)	0.002
…Midwest	153,980 (22.5%)	144,727 (20.4%)	0.051	103,968 (21.5%)	104,070 (21.6%)	0.002
…West	106,303 (15.5%)	120,519 (17.0%)	0.041	77,738 (16.1%)	77,360 (16.0%)	0.003
BURDEN OF COMORBIDITIES
Combined comorbidity score; mean (SD)	1.6 (2.1)	1.6 (2.2)	0.023	1.5 (2.1)	1.5 (2.2)	0.000
Claims-based frailty index3; n (%)
…Robust	338,681 (49.4%)	364,741 (51.4%)	0.040	249,234 (51.6%)	249,616 (51.7%)	0.002
…Prefrail	304,499 (44.4%)	307,661 (43.4%)	0.020	208,050 (43.1%)	207,777 (43.0%)	0.002
…Frail	42,446 (6.2%)	37,109 (5.2%)	0.043	25,628 (5.3%)	25,519 (5.3%)	0.000
LIFESTYLE FACTORS; n (%)
Body mass index categories
…Overweight	57,249 (8.3%)	77,278 (10.9%)	0.088	45,708 (9.5%)	45,771 (9.5%)	0.000
…Class 1 or 2 obesity	159,047 (23.2%)	137,762 (19.4%)	0.093	103,800 (21.5%)	104,459 (21.6%)	0.002
…Class 3 obesity	181,249 (26.4%)	104,247 (14.7%)	0.293	92,257 (19.1%)	92,227 (19.1%)	0.000
…Unspecified obesity	182,616 (26.6%)	138,035 (19.5%)	0.169	109,943 (22.8%)	110,250 (22.8%)	0.000
Smoking	141,099 (20.6%)	140,047 (19.7%)	0.022	94,497 (19.6%)	94,422 (19.6%)	0.000
DIABETES-SEVERITY RELATED CONDITIONS; n (%)
Diabetic nephropathy	121,639 (17.7%)	126,524 (17.8%)	0.003	80,909 (16.8%)	80,313 (16.6%)	0.005
Diabetic neuropathy	157,699 (23.0%)	146,579 (20.7%)	0.056	104,677 (21.7%)	104,266 (21.6%)	0.002
Hyperglycemia	363,857 (53.1%)	346,635 (48.9%)	0.084	248,216 (51.4%)	248,328 (51.4%)	0.000
Diabetic ketoacidosis	1,900 (0.3%)	1,549 (0.2%)	0.020	1,140 (0.2%)	1,127 (0.2%)	0.000
EYE DISEASES AND RELATED CARE; n (%)
Diabetic retinopathy	68,772 (10.0%)	68,013 (9.6%)	0.014	46,761 (9.7%)	46,423 (9.6%)	0.003
Glaucoma (any)	37,832 (5.5%)	45,451 (6.4%)	0.038	28,021 (5.8%)	27,866 (5.8%)	0.000
Cataract	153,248 (22.4%)	168,918 (23.8%)	0.033	111,521 (23.1%)	111,014 (23.0%)	0.002
Pan retinal photocoagulation	2,077 (0.3%)	1,891 (0.3%)	0.000	1,324 (0.3%)	1,378 (0.3%)	0.000
Age-related macular degeneration	18,021 (2.6%)	22,616 (3.2%)	0.036	13,566 (2.8%)	13,537 (2.8%)	0.000
Hemorrhage in the eye	9,129 (1.3%)	9,416 (1.3%)	0.000	6,302 (1.3%)	6,254 (1.3%)	0.000
History of vitrectomy	1,450 (0.2%)	1,505 (0.2%)	0.000	975 (0.2%)	1,042 (0.2%)	0.000
Vision impairment or loss	4,734 (0.7%)	5,331 (0.8%)	0.012	3,355 (0.7%)	3,338 (0.7%)	0.000
Screening or vision or retinal eye examination	264,956 (38.6%)	284,708 (40.1%)	0.031	189,635 (39.3%)	189,257 (39.2%)	0.002
Ophthalmic diagnostic imaging	148,829 (21.7%)	161,498 (22.8%)	0.026	105,929 (21.9%)	106,459 (22.0%)	0.002
Ophthalmologist or optometrist visit	280,649 (40.9%)	294,627 (41.5%)	0.012	199,112 (41.2%)	198,566 (41.1%)	0.002
GLUCOSE LOWERING THERAPIES; n (%)
No. of GLMs at cohort entry; mean (SD)	2.2 (0.9)	2.3 (0.9)	0.099	2.3 (0.9)	2.3 (0.9)	0.000
Metformin	498,542 (72.7%)	556,603 (78.4%)	0.133	370,694 (76.8%)	371,196 (76.9%)	0.002
Dipeptidyl peptidase-4 inhibitors	138,321 (20.2%)	192,356 (27.1%)	0.163	113,229 (23.4%)	112,895 (23.4%)	0.000
Sulfonylureas - 2nd generation	227,165 (33.1%)	254,837 (35.9%)	0.059	171,474 (35.5%)	172,200 (35.7%)	0.004
Insulin	203,825 (29.7%)	127,862 (18.0%)	0.277	114,101 (23.6%)	112,586 (23.3%)	0.007
CARDIOVASCULAR COMORBIDITIES; n (%)
Hypertension	543,228 (79.2%)	574,191 (80.9%)	0.043	385,566 (79.8%)	385,016 (79.7%)	0.003
Hyperlipidemia	523,639 (76.4%)	561,952 (79.2%)	0.067	375,065 (77.7%)	374,523 (77.6%)	0.002
Coronary artery diseases4	139,850 (20.4%)	184,793 (26.0%)	0.133	106,267 (22.0%)	105,704 (21.9%)	0.002
Congestive heart failure	58,972 (8.6%)	85,937 (12.1%)	0.115	43,302 (9.0%)	42,962 (8.9%)	0.004
Hypotension (except intracranial hypotension)	12,773 (1.9%)	15,409 (2.2%)	0.021	8,783 (1.8%)	8,808 (1.8%)	0.000
Ischemic stroke	43,151 (6.3%)	53,919 (7.6%)	0.051	31,986 (6.6%)	31,917 (6.6%)	0.000
KIDNEY DISEASES; n (%)
Acute Kidney Injury	27,270 (4.0%)	28,732 (4.0%)	0.000	16,734 (3.5%)	16,736 (3.5%)	0.000
Chronic kidney disease (any)	116,122 (16.9%)	120,139 (16.9%)	0.000	75,592 (15.7%)	74,616 (15.5%)	0.006
Chronic kidney disease stages 3–4	88,044 (12.8%)	88,272 (12.4%)	0.012	54,868 (11.4%)	54,438 (11.3%)	0.003
OTHER COMORBIDITIES; n (%)
Venous thromboembolism	18,664 (2.7%)	17,347 (2.4%)	0.019	11,894 (2.5%)	11,808 (2.4%)	0.007
Hypercoagulable states	5,258 (0.8%)	6,193 (0.9%)	0.011	3,630 (0.8%)	3,427 (0.7%)	0.012
Hemorrhage or major bleeding	51,436 (7.5%)	53,391 (7.5%)	0.000	34,949 (7.2%)	35,734 (7.4%)	0.008
Anemia	114,938 (16.8%)	121,579 (17.1%)	0.008	77,458 (16.0%)	77,279 (16.0%)	0.000
Obstructive sleep apnea	161,481 (23.6%)	125,528 (17.7%)	0.146	97,590 (20.2%)	97,388 (20.2%)	0.000
Osteoarthritis of knee	75,148 (11.0%)	62,557 (8.8%)	0.074	46,652 (9.7%)	46,754 (9.7%)	0.000
MASH/MAFLD	54,319 (7.9%)	45,171 (6.4%)	0.058	34,248 (7.1%)	34,134 (7.1%)	0.000
Migraine	21,764 (3.2%)	14,100 (2.0%)	0.076	11,723 (2.4%)	11,606 (2.4%)	0.000
OTHER MEDICATIONS; n (%)
No. of antihypertensive drugs at cohort entry
…0	169,189 (24.7%)	156,206 (22.0%)	0.064	113,887 (23.6%)	114,328 (23.7%)	0.002
…1 – 2	355,318 (51.8%)	376,569 (53.1%)	0.026	254,549 (52.7%)	254,349 (52.7%)	0.000
… ≥ 3	161,119 (23.5%)	176,736 (24.9%)	0.033	114,476 (23.7%)	114,235 (23.7%)	0.000
ACEi or ARBs	482,686 (70.4%)	516,057 (72.7%)	0.051	348,093 (72.1%)	347,636 (72.0%)	0.002
β-blockers	242,833 (35.4%)	277,068 (39.1%)	0.077	174,631 (36.2%)	174,155 (36.1%)	0.002
Calcium channel blockers	193,981 (28.3%)	212,255 (29.9%)	0.035	139,237 (28.8%)	138,861 (28.8%)	0.000
Loop diuretics	97,974 (14.3%)	100,506 (14.2%)	0.003	63,200 (13.1%)	62,864 (13.0%)	0.003
Thiazide and thiazide-like diuretics	227,631 (33.2%)	218,947 (30.9%)	0.049	156,427 (32.4%)	157,031 (32.5%)	0.002
Medications to treat hypotension5	1,891 (0.3%)	1,580 (0.2%)	0.020	1,141 (0.2%)	1,008 (0.2%)	0.000
Amiodarone	5,479 (0.8%)	9,455 (1.3%)	0.049	4,251 (0.9%)	4,190 (0.9%)	0.000
Statins	490,636 (71.6%)	539,282 (76.0%)	0.100	356,993 (73.9%)	356,186 (73.8%)	0.002
Anticoagulants (oral or injectables)	52,039 (7.6%)	64,443 (9.1%)	0.054	37,321 (7.7%)	37,188 (7.7%)	0.000
Medications to reduce intraocular pressure6	30,921 (4.5%)	37,664 (5.3%)	0.037	23,027 (4.8%)	22,943 (4.8%)	0.000
Phosphodiesterase type 5 inhibitors	16,789 (2.4%)	19,844 (2.8%)	0.025	12,996 (2.7%)	12,948 (2.7%)	0.000
Opioids	206,905 (30.2%)	185,083 (26.1%)	0.091	135,122 (28.0%)	134,770 (27.9%)	0.002
Corticosteroids (oral)	141,623 (20.7%)	126,651 (17.9%)	0.071	90,930 (18.8%)	90,934 (18.8%)	0.000
Disease-modifying antirheumatic drugs	17,053 (2.5%)	14,104 (2.0%)	0.034	10,190 (2.1%)	10,728 (2.2%)	0.007
HEALTHCARE RESOURCE UTILIZATION; n (%)
No. of office visits; mean (SD)	10.01 (7.33)	9.24 (6.86)	0.109	9.43 (7.34)	9.41 (6.84)	0.003
No. of hospitalization events; mean (SD)	0.12 (0.43)	0.14 (0.47)	0.044	0.12 (0.44)	0.12 (0.47)	0.000
Internist	614,706 (89.7%)	639,560 (90.1%)	0.013	433,070 (89.7%)	432,839 (89.6%)	0.003
Endocrinologist	131,942 (19.2%)	98,509 (13.9%)	0.143	79,609 (16.5%)	79,063 (16.4%)	0.003
Cardiologist	223,224 (32.6%)	254,532 (35.9%)	0.070	159,168 (33.0%)	158,848 (32.9%)	0.002
LABORATORY RESULTS 7 ; mean (SD)
Hemoglobin A1c, %	8.05 (1.8)	8.01 (1.7)	0.022	8.10 (1.8)	8.14 (1.7)	0.023
eGFR, mL/min/1.73m2	81.3 (22.4)	77.1 (21.8)	0.189	79.7 (22.4)	80.1 (21.7)	0.018

PS: propensity score; SGLT2i: Sodium-glucose cotransporter 2 inhibitors; GLP-1RA: Glucagon-like peptide 1 receptor agonists; SD: standard deviation; GLM: Glucose lowering medications; MASH: metabolic dysfunction-associated steatohepatitis; MAFLD: metabolic dysfunction-associated fatty liver disease; ACEi: angiotensin-converting enzyme inhibitors; ARBs: angiotensin receptor blockers; eGFR: estimated glomerular filtration rate calculated using CKD-EPI 2021 formula²²

¹Only available in Medicare FFS and CDM, other/unknown racial category includes North American Native, unknown, or missing

²Only available in Medicare FFS and CDM

³Defined using Validation of a Claims-Based Frailty Index Against Physical Performance and Adverse Health Outcomes in the Health and Retirement Study

⁴Defined as the presence of acute myocardial infarction (MI), MI sequelae/old MI, stable angina, unstable angina, coronary atherosclerosis, coronary procedure, history of coronary procedure, congestive heart failure, ischemic stroke, cerebrovascular procedure, generalized and unspecified atherosclerosis, peripheral arterial disease or atherosclerotic cerebrovascular disease assessed during the baseline period

⁵Defined as prescription fill claims for fludrocortisone, droxidopa, atomoxetine, or midodrine during the baseline period

⁶Defined as prescription eye drop fills for β-adrenergic antagonists, carbonic anhydrase inhibitors, cholinergics, α-adrenergic agonists, prostaglandins, or prostamides assessed during the baseline period

⁷Hemoglobin A1c is available for 43% and eGFR is available for 47% of CDM’s cohort and were not included in the PS model

Statistical Analysis

To address temporal changes in cardiovascular and kidney disease treatment guidelines and potential channeling bias associated with GLP-1RA and SGLT2i prescribing over the study period,^23–25 we calculated propensity scores (PS) within three calendar time periods for CDM and Marketscan (2016–2018, 2019–2020, and after 2020) and two time periods for Medicare FFS (2016–2018 and 2019–2020), based on participants’ year of cohort entry. To estimate the probability of receiving the index exposure, we used multivariable logistic regression models conditioning on baseline confounders and confounder proxies (see Table 1 and eTables 6), except laboratory test results which were only available in a subset of the study population. For confounding adjustment, we used 1:1 PS matching (PSM) using the nearest neighbor methodology with a maximum caliper of 0.01 of the PS within each time block and database.²⁶ We assessed the balance of each covariate before and after PSM using absolute standardized differences, with differences less than 10% indicating negligible differences.

We calculated the incident rate (IR) of NAION for each exposure group and IR differences (RD) per 1,000 person-years (PY) with a 95% confidence interval (CI) comparing the groups. To estimate hazard ratios (HRs) with 95% CIs, we used Cox proportional hazards models and pooled database-specific HRs using fixed-effects meta-analysis. Kaplan-Meier curves and log-rank test statistics were generated to visualize and compare the cumulative hazard of each exposure group over time. All analyses were conducted with Aetion Evidence Platform^® version 5.17.1, Microsoft Excel (Microsoft Corporation), SAS version 9.4 (SAS Institute Inc., 2018), and Stata version 17.0 (StataCorp LLC 2022).

Secondary, Sensitivity, and Subgroup Analyses

To evaluate the robustness of our findings, we conducted multiple secondary, sensitivity, and bias analyses. First, to preserve the number of events, we explored a less specific outcome definition based on ION diagnosis codes only without requirement of an ophthalmologist or optometrist visit on the same day. Second, to improve the precision of our results and assess the average treatment effect in the treated population, we repeated the primary analysis using PS fine stratification methodology. Third, to account for the introduction in the market of tirzepatide, we expanded our exposure definition to include tirzepatide and reconducted the primary analysis in the CDM database, which has the most recent data available. Fourth, as the NAION safety signal emerged specifically for semaglutide,³ we assessed the risk of NAION separately for initiators of semaglutide compared with SGLT2is. Fifth, we also evaluated the risk of NAION among initiators of dulaglutide, another frequently prescribed GLP-1RA in the US.²⁷ Sixth, since the diagnosis code used in our outcome definition is not specific to NAION, we conducted a quantitative bias analysis to quantify the degree of potential outcome misclassification and its impact on observed effect estimates.²⁸ Additionally, to determine the presence of any treatment effect heterogeneity, we performed subgroup analyses by age (<65 years, ≥ 65 years), non-differentiated sex or gender (female, male), race (White, non-White), obesity severity (body mass index [BMI] of 30–39.9 kg/m², ≥ 40 kg/m²), smoking, cardiovascular disease, insulin use, obstructive sleep apnea, diabetic retinopathy, use of three or more antihypertensive medications at cohort entry, and HbA1c categories (< 8%, ≥ 8%). Lastly, we meta-analyzed findings from our semaglutide analysis with results from the two previously published population-based cohort studies, which appeared less prone to methodological challenges.^5,11

RESULTS

Cohort Characteristics

After pooling patients across the three datasets and before PSM, we identified 685,626 initiators of GLP-1RA and 709,511 initiators of SGLT2i (Figure 1). Dulaglutide and semaglutide were the most frequently prescribed GLP-1RA (38.9% and 35.0%, respectively), whereas empagliflozin and dapagliflozin were the most frequently initiated SGLT2i (59.9% and 24.1%, respectively (eTable 4). Compared to SGLT2i users, patients receiving GLP-1RAs were younger (mean age: 62 vs. 64 years), more likely to be female (54% vs 43%), more likely to have a recorded obesity or BMI-related diagnosis code (58% vs 47%) and higher BMI, and more likely to be on insulin therapy at baseline (30% vs 18%) (Table 1 and eTable 6).

Figure 1.

Consort diagram of study participants initiating GLP-1RA or SGLT2i

Following PSM, we identified 482,912 matched pairs. Patient characteristics were balanced between exposure groups with absolute standardized differences of less than 10%, including HbA1c and eGFR results, which were not included in the PS adjustment as these results were only available in a subset of the study population (Table 1 and eTable 6). The mean (SD) age was 63 (12) years, approximately 49% were female, 64% were of White race, 10% had diabetic retinopathy, and 24% were taking at least three antihypertensive medications at the time of cohort entry. The median [IQR] follow-up months were 6.5 [3.7, 13.9] and 6.8 [3.8, 13.9] for GLP-1RA and SGLT2i initiators, respectively. Most study participants were censored during follow-up due to index drug discontinuation (53% in GLP-1RA and 50% in SGLT2i initiators) (eTable 7).

Primary NAION Outcome

In the overall PSM cohort, there were 270 presumed NAION events among individuals initiating a GLP-1RA (IR, 0.61 events per 1,000 PY) and 144 among those initiating an SGLT2i (IR, 0.33 events per 1,000 PY), resulting in an increased risk of NAION in the GLP-1RA group compared with the SGLT2i group (HR, 1.85; 95% CI, 1.51 to 2.27; RD/1,000 PY, 0.29; 95% CI, 0.19 to 0.38) (Table 2). The cumulative incidence curves of NAION started diverging after the first three months of follow-up (Figure 2).

Table 2.

NAION outcomes in study participants initiating GLP-1RA or SGLT2i

NAION outcome	Database	GLP-1RA	SGLT2i	HR (95% CI)	RD/1,000 PY (95% CI)
		N events (IR/1,000 PY)	N events (IR/1,000 PY)
Primary NAION definition 1	Medicare FFS	106 (0.95)	56 (0.51)	1.86 (1.35–2.58)	0.44 (0.21–0.66)
	Marketscan	60 (0.39)	35 (0.23)	1.69 (1.11–2.56)	0.16 (0.03–0.28)
	CDM	104 (0.60)	53 (0.30)	1.95 (1.39–2.73)	0.30 (0.16–0.44)
	Pooled	270 (0.61)	144 (0.33)	1.85 (1.51–2.27)	0.29 (0.19–0.38)
Secondary NAION definition 2	Medicare FFS	109 (0.98)	60 (0.55)	1.79 (1.30–2.45)	0.43 (0.20–0.66)
	Marketscan	75 (0.49)	41 (0.27)	1.80 (1.23–2.63)	0.21 (0.08–0.35)
	CDM	120 (0.69)	70 (0.39)	1.71 (1.27–2.31)	0.29 (0.14–0.45)
	Pooled	304 (0.69)	171 (0.39)	1.76 (1.46–2.12)	0.30 (0.20–0.40)

¹Based on inpatient or outpatient diagnosis codes of ischemic optic neuropathy combined with an ophthalmologist or optometrist visit on the same day

²Based on inpatient or outpatient diagnosis codes of ischemic optic neuropathy only code

Figure 2.

Cumulative incidence curves for NAION outcomes among study participants initiating GLP-1RA or SGLT2i

Secondary, Sensitivity, and Subgroup Analysis

Results were consistent, with an attenuation toward the null, when we used diagnosis codes without eye care provider visit to define the NAION outcome (HR, 1.76; 95% CI, 1.46 to 2.12; RD/1,000 PY, 0.30; 95% CI, 0.20 to 0.40) (Table 2). The risk of NAION was higher in patients initiating a GLP-1RA compared with an SGLT2i (HR, 1.71; 95% CI, 1.41 to 2.09; RD/1,000 PY, 0.24; 95% CI, 0.16 to 0.31) in an analysis using PS fine stratification for confounding adjustment (eTable 8). In an exposure definition that included tirzepatide, the risk of NAION was higher among GLP-1RA or tirzepatide initiators (HR, 2.67; 95% CI, 1.47 to 4.83; RD/1,000 PY, 0.45; 95% CI, 0.19 to 0.70), compared with SGLT2i initiators. Similar results were obtained when we compared initiators of semaglutide to new-users of SGLT2i (HR, 1.80; 95% CI, 1.23 to 2.64; RD/1,000 PY, 0.20; 95% CI, 0.05 to 0.34), whereas we did not observe a meaningfully higher risk of NAION when we compared dulaglutide with SGLT2i initiators (HR, 1.17; 95% CI, 0.80 to 1.70; RD/1,000 PY, 0.09; 95% CI, −0.03 to 0.20) (eTable 9).

Findings were consistent across subgroups of age (≤65 and ≥ 65 years), non-differentiated sex or gender, race, obesity severity, smoking, cardiovascular disease, insulin use, obstructive sleep apnea, diabetic retinopathy, use of three or more antihypertensive medications at cohort entry, and HbA1c (eFigure 2). Although precision of the effect estimates was reduced in the smaller subgroups, analyses using more granular age categories (≤49 years, 50–64 years, 65–74 years, and ≥75 years) produced consistent results, with no evidence of effect modification by age (eFigure 3). A random-effects meta-analysis of results from published studies, incorporating findings from our semaglutide analysis, resulted in a HR of 2.78 (1.39 to 5.56) (eTable 10).

Finally, a quantitative bias analysis, which assessed the impact of outcome misclassification on observed findings, showed that our observed estimate is likely an underestimate of the true association between GLP-1RA and NAION, as described in eTable 11.

DISCUSSION

In this large nationwide observational study of patients with T2D, using a new user active comparator study design within a target trial framework, the initiation of GLP-1RA was associated with an 85% increased risk of presumed NAION, compared with SGLT2i. The incidence of NAION in our population was 50 per 100,000 person-years, which is consistent with the incidence rate observed in a population of similar age,²⁹ and the absolute increase in NAION risk associated with GLP-1RA vs. SGLT2i initiation was less than 30 additional cases per 100,000 person-years. Our study findings were consistent across multiple prespecified sensitivity and subgroup analyses and robust to different assumptions of outcome misclassification. Meta-analysis of semaglutide analyses resulted in a 176% increased risk of presumed NAION, compared with non-GLP-1RA antidiabetics.

A meta-analysis of clinical trial data that compared the incidence of ION events in patients randomized to GLP-1RA agents versus placebo did not find a significant difference between the groups (8/21,374 and 4/20,729, respectively, with odds ratio: 1.53; 95% CI, 0.53 to 4.44).¹⁰ However, the trials were not powered to assess NAION events. To date, eight observational studies and one case series assessed the risk of NAION among patients prescribed semaglutide with mixed findings.^3–12 The first study reported an increased risk of NAION among users of semaglutide compared with users of non-GLP-1RA with T2D (HR, 4.28; 95% CI, 1.62 to 11.29) and with overweight or obesity (HR, 7.64; 95% CI, 2.21 to 26.36).³ The authors acknowledge that the study utilized data from a tertiary care institution that specializes in ophthalmology and neuro-ophthalmology services, and thus could be subject to selection bias; however, the diagnosis of NAION was confirmed by expert examiners.³

Subsequent observational studies based on EHR data, claims, and Scandinavian registries also compared semaglutide against non-GLP-1RA drugs or other GLP-1RAs, but reported mixed findings on the association between semaglutide use and NAION in populations of patients with T2D, obesity, or T2D and obesity. Of the seven cohort studies, five incur a combination of significant study design flaws due to prevalent user design (risk of depletion of susceptible bias), confounding by indication bias due to comparator that are used at different stages of diabetes severity,¹⁹ unavailability of comprehensive capture of care data in patients’ longitudinal journey i.e. not fit-for-purpose data, or lack of information on loss to follow-up, such as death or plan disenrollment.^4,6–8,12 For example, Figure 2 in the study by Hsu et al. shows that more than 99% of the study participants remained at risk and under follow-up by the study’s fifth year, suggesting that information on loss to follow-up may not have been accounted for in the study analysis.⁸ Additionally, the EHR-based studies defined exposure to semaglutide using prescribing information rather than pharmacy fill information, which may pose challenges as prescribed drugs are not always dispensed due to factors such as shortages or cost barriers.³⁰ To reduce the chance of biases that could undermine validity in the current study and ensure longitudinal follow-up, we used a new-user, active-comparator study design within a target trial framework and accounted for participant enrollment information during baseline and follow-up periods in our analysis. We also used pharmacy dispensing claims to identify drug exposure.

One of the two studies less prone to the noted methodological challenges compared semaglutide vs. SGLT2i using Danish and Norwegian registry data and reported a 2.81-fold increased risk of NAION with 95% CI, 1.67 to 4.75 in an ITT analysis and 6.35-fold increase with 95% CI, 2.88 to 14.0 in an on-treatment analysis. The second study using OHDSI databases from patients with T2D compared semaglutide users with other individual GLP-1RAs or non-GLP-1RA antidiabetic agents and reported an increase in the risk of NAION among the semaglutide group (HR comparing empagliflozin, 2.27; 95% CI 1.59 to 4.46).⁵ However, both studies focused only on semaglutide and the former study relied on a low number of NAION events, while the latter did not report a granular adjustment for obesity and diseases associated with diabetes severity, and the study population was limited to metformin users, while simultaneously excluding those who used other antidiabetic medications at baseline.

In our analysis evaluating initiators of semaglutide vs. SGLT2i, we found a similar direction of the effect, albeit effect estimates were more precise and attenuated compared to findings from the Hathaway et al., Cai et al., and Simonsen et al. studies,^3,5,11 possibly driven by our larger study population (482,912 PS-matched pairs) and enhanced confounding control. When we compared initiators of dulaglutide vs. SGLT2i, we did not find an elevated risk of NAION among dulaglutide users consistent with findings from Cai et al., whereas the risk of NAION was 2.67-fold higher in initiators of GLP-1RA when including initiators of tirzepatide compared with SGLT2i. Although the cause of ischemia to the optic nerve head in NAION remains unknown,^1,2,31 we speculate that the higher risk of NAION observed for initiators of semaglutide and tirzepatide but not dulaglutide could be related to the increased potency in weight loss, coupled with reductions in blood pressure and HbA1c effects of these drugs compared to dulaglutide.^32–34 We hypothesize that GLP-1RAs may contribute to the development of NAION through a multifactorial pathway involving blood pressure-related vascular effects, and physiological and anatomical susceptibilities associated with weight loss. These factors may act synergistically to increase the risk of optic nerve damage. Hemodynamic changes secondary to blood pressure reduction may impair vascular autoregulation and promote endothelial dysfunction. Additionally, the rapid weight loss induced by potent GLP-1RAs could disrupt fluid balance and potentially compromise optic nerve perfusion, particularly in individuals with a small cup-to-disc ratio and with microvascular complications of diabetes. Our study has several strengths. We designed the study to specifically address the incidence of NAION among patients with T2D and without any prior history of ION, optic nerve head pathology such as papilledema or conditions like giant cell arteritis that can cause ION, optic atrophy of any cause, shock, uveitis, or any eye or systemic conditions that can cause other forms of optic neuropathy to reduce the chances of outcome misclassification. Our study design choices were guided by expert opinions from neuro-ophthalmologists on our team and using information obtained from the American Academy of Ophthalmology and previous publications on risk factors, patient presentations, clinical course, and diagnosis of NAION.^1,2,35–38 Second, despite reports of an elevated risk of developing NAION in patients with diabetes,²⁹ the true incidence in this group remains unknown. NAION estimates extrapolated from small local studies to the U.S. population suggest a mean annual incidence of 2 to 10 per 100,000 in adults over the age of 50 years and 0.54 per 100,000 for all ages,^39,40 while another study reported an incidence of 82 per 100,000 in adults over 67 years old.²⁹ Having a study population several orders of magnitude larger than previously published studies allowed us to quantify with more precision the association between GLP-1RAs and the risk of NAION in the overall population of patients with T2D and across multiple clinically relevant subgroups. Moreover, our study did not restrict to baseline users of metformin monotherapy, thereby reflecting more closely

evolving treatment practices over the study period and enhancing generalizability. Finally, our study could rely on the availability of information and adjustment for over 130 baseline variables and implemented multiple advanced analytic approaches to mitigate confounding bias.

Our study also has limitations. First, NAION events were primarily identified in claims data using non-specific ischemic optic neuropathy diagnosis codes. Hence, non-differential outcome misclassification is possible. We attempted to mitigate outcome misclassification by using stricter cohort entry requirements such as requiring a same-day ophthalmologist or optometrist visit and improving our outcome algorithm to increase specificity. We also conducted a quantitative bias analysis, which showed that in the setting of non-differential outcome misclassification, which would bias effect estimates towards the null, our observed estimates would be conservative. Another limitation of our study is the short median follow-up of 6.7 months, though this aligns with the 186-day median time to NAION onset reported in a disproportionality analysis of FDA adverse event data following semaglutide use.⁴¹ Moreover, due to the observational nature of our study, residual confounding by unmeasured or not completely measured baseline patient characteristics such as diabetes severity or BMI cannot be entirely ruled out, though it was mitigated by using a new-user, active comparator study design, adjusting for over 130 baseline variables, and implementing multiple advanced analytic approaches to address confounding.

In conclusion, in this large nationwide population-based cohort study of patients with T2D, using a new user active comparator study design within a target trial framework, the initiation of GLP-1RAs was associated with an increased risk of NAION, compared with the initiation of SGLT2is, though the incidence rate of NAION was low and the absolute increase in NAION risk associated with GLP-1RA vs. SGLT2i initiation was small in our population. These findings support and expand on the previously generated safety signal of an increased risk of NAION among users of GLP-1RAs, and through a more precise quantification of such risk may help clinicians balance the benefits and risks of treatment with GLP-1RAs in patients with T2D.