Reviews/GLP-1 Receptor Agonists/Increase non-arteritic anterior ischaemic optic neuropathy risk
Safety · GLP1-SAF-01 · v2
Borderline: within 4 points of the Likely/Uncertain boundary, so a re-grade could land either side.
Several large studies that followed people with type 2 diabetes found that those taking semaglutide developed NAION, a rare and usually permanent cause of vision loss, roughly twice as often as comparison patients ([1], [8], [24], [25], [4]). Reports collected from drug-safety databases point in the same direction and are very strong in size ([23], [27]). However, these are observational studies that cannot prove the drug itself is the cause, and the rate of the condition is low in everyone. Working against the claim, a pooled analysis of randomised trials found no increase ([13]), and two other real-world studies actually found a lower risk in the groups taking semaglutide ([10], [31]). The panel weighed the repeated two-fold signal against the null trial data and the contradictory studies and judged the association real but not proven to be causal.
Multiple independent cohorts converge on an approximately two-fold increased NAION hazard with semaglutide: a Danish national cohort [1] HR 2.19, a US Veterans active-comparator emulation [8] HR 2.33, a Danish-Norwegian cohort [24] HR 2.81, a TriNetX cohort at two years [25] HR 2.39, and a target-trial emulation [4] HR 1.76. A single-center neuro-ophthalmology registry [22] reports larger effects (HR 4.28 and 7.64), while a rigorous multi-database analysis [23] attenuates to IRR 1.32. Pharmacovigilance disproportionality is very large ([27] ROR 94.45; [11] ROR up to 74.89). Against this, pooled randomised placebo-controlled trials [13] found 3 confirmed cases on GLP-1RA versus 5 on placebo across 96,829 participants, and two cohorts report reduced risk ([10] OR 0.36; [31] HR 0.47). Absolute incidence is consistently low (e.g. [8] 123 vs 67 per 100,000 person-years). The contested point: The evidence divides on whether the replicated observational hazard reflects a real drug effect or confounding and bias: the randomised pool [13] is null and two primary cohorts ([10], [31]) reverse the sign, yet both sit in distinct populations (a military health system and an IIH cohort) and the trial pool rests on only 8 total events. The panel split on how much weight the null and sign-reversing evidence can bear against the directionally consistent cohort signal.
All 31 papers behind this claim with their DOIs and what each does to it, the panel’s arbitration, the evidence gap map and the open research questions. Free, and sent to your inbox.
We’ll also send you other claims from GLP-1 Receptor Agonists, one at a time, and tell you when a release or a pivotal trial changes a grade. Unsubscribe by replying.
One reviewer's only job is to attack the finding. The challenges below were weighed before the grade was set.
The highest-tier evidence contradicts the claim: pooled placebo-controlled RCTs [13] found 3 confirmed NAION cases on GLP-1RA vs 5 on placebo across 96,829 participants, i.e. no increase. The observational signal is not corroborated by randomised data.
Direct sign reversal in primary data: [10] reports OR 0.36 (reduced risk) in a 1.2M T2D cohort and [31] reports HR 0.47 in IIH patients. The corpus is internally contradictory on direction, not merely on magnitude.
Where it remains uncertain. Pooled randomised trials [13] found no increase (3 vs 5 cases in 96,829 participants) and two cohorts report reduced risk ([10] OR 0.36; [31] HR 0.47).
Aleatory
NAION is rare and event counts are small, so estimates carry wide random error, e.g. the pooled trial [13] rests on 8 total events and [22] reports CIs as wide as 1.62-11.29.
Epistemic
Cohorts cannot rule out that who receives semaglutide differs in NAION risk; [29] hypothesises rapid correction of hyperglycemia rather than drug toxicity drives events.
The load-bearing papers from a corpus of 31, each classified by what it does to the claim. Every DOI is taken from the corpus the claim was graded on.
Large national prospective cohort establishing ~2-fold increased hazard.
“Semaglutide exposure was associated with a higher incidence rate (0·228 vs. 0·093 per 1000 person-years, p < 0·001) and independently predicted a higher risk of upcoming NAION (HR 2·19, 95% confidence interval 1·54 - 3·12)”
Panel auditInternal validity: medium · scope of inference: Persons with type 2 diabetes in Denmark 2018-2024 on once-weekly semaglutide.Key limitations: observational · T2D only · residual confounding
Active-comparator target-trial emulation replicating the 2-fold signal against SGLT2i.
“semaglutide initiators had a 2.33-fold higher risk than SGLT2i initiators (hazard ratio, 2.33; 95% CI, 1.54-3.54; P < .001)”
Panel auditInternal validity: medium · scope of inference: US veterans with T2D on metformin initiating second-line therapy; absolute risk low.Key limitations: observational · 85.5% male veterans · active comparator is SGLT2i not placebo
Counter-evidence: highest-tier design finds no increase (but only 8 events).
“Data from randomised placebo-controlled trials with semaglutide and liraglutide do not show an increased incidence of NAION in participants receiving GLP-1RA treatment versus those receiving placebo”
Panel auditInternal validity: high · scope of inference: Randomised-trial populations with T2D and/or obesity on liraglutide or semaglutide.Key limitations: very few events · combines liraglutide and semaglutide · no formal test statistic reported
Counter-evidence: large cohort reports reduced risk.
“patients with type 2 diabetes prescribed semaglutide had a significantly lower risk of NAION (OR = 0.36, 95% CI = 0.25-0.51) compared to those prescribed non-GLP-1 RA medications”
Panel auditInternal validity: medium · scope of inference: US Military Health System beneficiaries with T2D or obesity.Key limitations: observational · obesity subgroup wide CI · possible confounding by indication
4 of 31 papers shown. The full review lists all of them, with the panel’s arbitration →
Claims are graded at each quarterly release, and re-graded between releases only when a pivotal trial lands. Every change is a new version of the claim.
The replicated, directionally consistent cohort signal supports a likely increased risk, but unproven causality, a null (if underpowered) RCT pool, and genuine contradictory cohorts preclude a higher tier.
Release 2026-q4 re-graded every claim from its corpus under protocol v0.19. An earlier release's grade is shown for the record; it came from an earlier protocol and is not a movement.
Assumptions that carry weight but are not themselves established.
The panel records the gap that would most move the claim, and the study that would close it.
Adequately powered randomised evidence. The only trial-level data [13] has 8 events and is consistent with both no effect and a doubling, leaving the strongest design mute on the claim. The panel rates this the highest-importance open gap for the claim, and names the study design that would close it: A prospective trial or pooled individual-patient-data analysis prespecifying adjudicated NAION as an outcome with enough person-years to detect a 2-fold hazard.
Open question: Does semaglutide cause NAION, or is the association driven by confounding by indication? Why it is still open: All supporting estimates ([1], [4], [8], [22], [23], [24], [25]) are observational and [29] offers rapid hyperglycemia correction as an alternative.
Priors is a sceptical first reader, not peer review, and not medical advice.
Release 2026-q4 · graded 21 Oct 2026 · protocol v0.19