Adults with type 2 diabetes who started semaglutide were about half as likely to develop a new seizure disorder as adults who started other diabetes drugs. The reduction held up against a second modern comparator, and it did not appear to run through weight loss or better blood sugar.
That is the finding from a target trial emulation published in Neurology ↗, the journal of the American Academy of Neurology, posted online June 17. A target trial emulation is a way of using real-world health records to imitate a randomized trial: researchers define who would have been eligible, when their "treatment" started, and what counts as an outcome, then statistically balance the groups so the comparison is as close to a coin-flip assignment as observational data allows.
What they did
The team drew on the All of Us Research Program ↗, the National Institutes of Health cohort that has enrolled hundreds of thousands of Americans and links their medical records. They built two new-user, active-comparator cohorts of adults with type 2 diabetes who started a glucose-lowering drug between January 2018 and October 2023, then watched for a first diagnosis of epilepsy or seizure.
The first cohort set 2,586 semaglutide starters against 7,627 people who started other glucose-lowering drugs (10,213 in total). The second set 2,814 semaglutide starters against 5,791 who started SGLT2 inhibitors (8,605 in total). SGLT2 inhibitors are a newer class of diabetes and heart-kidney drugs, the fairest available stand-in for a comparably modern, comparably sick patient. The groups were balanced with inverse probability of treatment weighting, a statistical reweighting that makes the treated and untreated look alike on age, sex, other conditions, and medications. The analysis was then repeated with a second method, targeted maximum likelihood estimation, to check that the answer did not depend on the technique.
What they found
Semaglutide starters had a lower risk of adult-onset seizure than people on other glucose-lowering drugs (hazard ratio 0.44, 95% confidence interval 0.25 to 0.79). A hazard ratio of 0.44 means roughly 56 percent lower risk over the study window. Against SGLT2 inhibitors the signal was nearly identical (hazard ratio 0.48, 95 percent confidence interval 0.27 to 0.85), about half the risk.
Put in absolute terms, the four-year risk of a new seizure fell by 1.78 percentage points versus other diabetes drugs and 1.46 points versus SGLT2 inhibitors. That works out to a number needed to treat of 70 and 131: treat 70 patients with semaglutide instead of another glucose-lowering drug for about four years, and one fewer person develops a seizure disorder. Seizures are uncommon, so the relative halving rides on top of a small baseline, which is why the absolute numbers are modest.
The part that makes this more than another GLP-1 association is the mediation analysis. The authors asked how much of the effect ran through the two things semaglutide is famous for, lowering blood sugar and lowering weight. The answer was almost none: HbA1c explained 2.4 percent of the effect against other drugs and 6.5 percent against SGLT2 inhibitors, and body mass index explained 0 percent and 0.7 percent. If the seizure signal is real, it is not the weight loss and it is not the glucose control doing the work.
Why it is not proof
The authors graded their own result as Class II evidence, the tier below a randomized trial, and named the limits plainly. Seizure events were few, follow-up was short, and observational reweighting cannot erase every difference between people who get prescribed one drug versus another. Residual confounding, the possibility that semaglutide users differ from comparators in some unmeasured way that also affects seizure risk, is the standard caveat for this kind of study and applies here.
Semaglutide's effects on the brain have been the most contested chapter of the GLP-1 story. The drug's Alzheimer's disease trials read out as failures, and a lone re-analysis argued the data deserved a second look ↗. A seizure-prevention signal that bypasses weight and glucose lands in the same uncertain territory: a hint of a direct central-nervous-system effect, on a drug whose central-nervous-system effects keep failing to behave the way the early enthusiasm predicted.
Semaglutide ↗ is a peptide that mimics GLP-1 and acts on the GLP-1 receptor ↗, which sits not only on pancreatic and gut cells but throughout the brain. That receptor distribution is the mechanistic reason a metabolic drug keeps turning up in neurology datasets at all. Whether it belongs there is what a real randomized seizure trial would have to settle, and this study is an argument for running one, not a substitute for it.