A systematic review and meta-analysis published online May 21 in PLOS Medicine ↗ pooled five observational studies of type 2 diabetes patients and found that semaglutide ↗ use was associated with a roughly two-fold increase in the hazard of nonarteritic anterior ischemic optic neuropathy (NAION), a sudden vision-threatening optic-nerve event. The pooled fixed-effects hazard ratio versus all other glucose-lowering regimens was 2.17 (95% CI 1.73 to 2.74, p<0.001). The absolute risk stayed low: about 0.014% per year, or roughly 1 additional NAION case per 7,000 semaglutide-treated patients each year.
The signal did not depend on the comparator. When the analysis excluded other GLP-1 receptor agonists from the non-semaglutide pool, the hazard ratio was 2.13 (1.60 to 2.83). Versus SGLT2 inhibitors alone the hazard ratio was 1.96 (1.28 to 2.99). Seven additional studies were added in a sensitivity round; leave-one-out analysis did not surface a single dominant trial driving the signal, and the authors found no small-study or publication-bias pattern.
NAION is the most common acute optic neuropathy in adults over fifty. The clinical picture is sudden, painless, monocular vision loss from infarction of the head of the optic nerve, usually noticed on waking. Most cases occur in people with a crowded optic-disc anatomy ("disc at risk"), and the population that develops NAION is already enriched for hypertension, sleep apnea, and diabetes. The same vascular-risk profile drives both the underlying disease state for which semaglutide is prescribed and the baseline NAION rate, which is what makes the registry-based signal hard to isolate. There is no specific treatment; the resulting visual deficit is usually permanent and partial.
The concern emerged from case reports in 2024 and prompted precautionary labeling changes. The PLOS Medicine paper is the first attempt to put a pooled number on the size of the signal across multiple cohorts, and the authors note their result "corroborated decisions presented in regulatory communications" already on file.
All five core studies were retrospective, registry-based, and exposed to two well-known confounders: outcome misclassification (NAION is rare and gets coded inconsistently in administrative data) and indication-by-population confounding (patients prescribed semaglutide differ from non-users on age, glycemic control, vascular disease, and other variables that themselves change NAION risk). The certainty of evidence under the GRADE framework was rated low. In plain terms, the direction of the signal is consistent across every cut the authors tried, but the precise magnitude could move once prospective data arrive.
At 1 additional case per 7,000 patient-years, the absolute risk does not change the prescribing calculus for an unselected T2DM or obesity patient. It does change the math for two narrow subgroups. People who have already had NAION in one eye face elevated second-eye risk regardless of treatment, and the second eye is the only one left. People with documented "disc at risk" anatomy on a recent eye exam carry a higher baseline rate. For both groups the meta-analysis tightens the case for an ophthalmology conversation before starting semaglutide, and the existing label recommendation to discontinue if NAION occurs is unchanged.
The class question is half-answered. Hazard ratios versus pools that excluded GLP-1 receptor agonists (2.13) and versus SGLT2 inhibitors alone (1.96) sit close to the primary 2.17, which means most of the signal is not coming from a comparator-arm artifact. But excluding other GLP-1 RAs from the comparator did not collapse the estimate either, hinting that the effect is semaglutide-weighted within the class rather than uniformly distributed across it. Mechanism remains open. Candidate hypotheses include rapid HbA1c lowering driving a hemodynamic insult to a borderline-perfused optic-nerve head, GLP-1 receptor signaling on retinal vasculature, or residual selection patterns that registry analyses cannot fully wash out.
The next data point will be the EU-mandated post-authorization safety study, the first cohort designed from scratch to ask the question rather than reconstructed from claims. Until then, the meta-analysis is the largest pooled estimate on file, and "real but small, low certainty" is the fair summary for prescribers.