Do GLP-1 Drugs Lower Seizure Risk? An Early Signal

GLP-1 receptor agonists may do more than lower glucose and body weight: a 2024 analysis found fewer reported late-onset seizures and epilepsy events among people taking these drugs. The finding is intriguing because two other glucose-lowering drug classes did not show the same pattern. It is also far too early to treat a GLP-1 drug as seizure prevention.

The best reading of the evidence is narrow. In older adults with type 2 diabetes and high cardiovascular or kidney risk, GLP-1 receptor agonists were associated with fewer passively reported seizure-related events in trials designed for cardiovascular and renal outcomes. That could reflect a direct effect in the brain, fewer strokes, chance, or some mixture of all three.

A signal that glucose control alone does not explain

A 2024 meta-analysis by Sindhu et al. pooled 27 placebo-controlled cardiovascular and renal outcome trials with 197,910 participants. The average participant was about 65 years old, 65.6% were male, and mean follow-up was 29.2 months. Across all newer glucose-lowering drugs studied, the analysis reported a statistically significant 24% lower risk of the combined endpoint of late-onset seizure and epilepsy versus placebo.

The subgroup results made the finding more interesting and less tidy. Eight GLP-1 receptor agonist trials, totaling roughly 60,000 participants, reported 30 seizure events in drug groups and 48 in placebo groups. The five DPP-4 inhibitor trials, with about 47,500 participants, reported 20 versus 22 events. Fourteen SGLT2 inhibitor trials, with about 90,000 participants, produced roughly 45 versus 48 events, with mixed results across individual trials. Neither non-GLP-1 subgroup had a statistically significant reduction.

That pattern argues against a simple explanation in which lower blood glucose automatically lowers seizure risk. But it does not prove that GLP-1 drugs act as anticonvulsants. Rare events create unstable estimates, and the patient populations, follow-up periods, and individual drugs differed across the trials.

An independent 2025 TriNetX study by Cheng et al. found lower epilepsy risk among people with type 2 diabetes using GLP-1 receptor agonists than among active comparators using DPP-4 inhibitors. This real-world result points in the same direction as the randomized-trial meta-analysis. It remains an observational association: matching can reduce measured differences between groups, but it cannot remove healthier-user effects, differences in clinical follow-up, or unequal recognition of seizures.

Why a direct brain effect is plausible

GLP-1 receptors are present in the hippocampus and cortex, brain regions involved in seizure generation. In animal and cell models, GLP-1 analogues have reduced inflammation, oxidative stress, and cell death; supported hippocampal neurogenesis; and altered GABA-related signaling. Those pathways could, in principle, affect seizure susceptibility.

A 2022 review by Manavi et al. summarizes these neuroprotective and neurotransmitter-related findings in preclinical epilepsy models. A 2024 systematic review of semaglutide in animal and cell-line studies likewise describes GLP-1 receptor expression in hippocampal and cortical circuits, along with effects on neuroinflammation, apoptosis, oxidative stress, and neurogenesis.

Biological plausibility is useful, but it is not human proof. Rodent seizure models do not reliably predict anticonvulsant effects in people, and no human pharmacodynamic study has shown that clinical GLP-1 receptor agonist doses raise seizure threshold or reduce electrographic seizure activity. A drug can improve several laboratory markers tied to brain injury and still fail to prevent seizures in patients.

The stroke explanation may be enough

The major alternative explanation is vascular rather than directly neurological. Cerebrovascular disease is a leading cause of late-onset epilepsy, especially in the older, high-risk population enrolled in these cardiovascular outcome trials. If a GLP-1 drug prevents strokes, it could also prevent some post-stroke seizures without changing neuronal excitability directly.

A 2026 propensity-matched multi-institutional cohort study by Costa et al. reported better outcomes with GLP-1 receptor agonist use for ischemic stroke, aneurysmal subarachnoid hemorrhage, and intracerebral hemorrhage. That evidence supports the vascular pathway; it does not establish a direct antiseizure effect.

Sindhu and colleagues extracted stroke counts, but they did not conduct a stroke-adjusted seizure analysis. As a result, the current data cannot estimate how much of the seizure signal, if any, comes from fewer cerebrovascular events. This distinction changes the clinical meaning. Preventing a stroke is enormously valuable, but it is different from demonstrating that a drug prevents epilepsy independent of stroke prevention.

The pooled result has real cracks

The meta-analysis was large, but its seizure data were thin. None of the 27 trials prospectively defined seizure or epilepsy as an outcome. Researchers drew events from adverse-event reports submitted to ClinicalTrials.gov, rather than from scheduled neurological assessment or adjudication. The authors rated selective-reporting risk as high across included studies because this was a post-hoc analysis of adverse-event categories.

Definitions also varied. The trials did not provide a standardized, verifiable distinction between seizure and epilepsy, and the investigators excluded alcohol-withdrawal seizures after the fact. A first unprovoked post-stroke seizure may have been coded differently across trials. Combining inconsistent labels can add enough diagnostic noise to distort a rare-event result.

Event rates stayed below 1% in every subgroup. The authors used Mantel-Haenszel methods for the primary analysis and Peto odds ratios as a sensitivity analysis. Sparse events strain both approaches, particularly when trials include zero-event groups. Statistical significance in a pooled rare-event analysis does not settle the size—or even the stability—of a clinical effect.

Individual GLP-1 trials were inconsistent. In REWIND, dulaglutide 1.5 mg produced 10 seizure events versus 9 with placebo among 9,901 participants. In AMPLITUDE-O, efpeglenatide had 4 events versus 2 with placebo among 4,076 participants. ELIXA, which studied lixisenatide in 6,068 participants, recorded 0 versus 1 event. These counts are too small to settle whether any individual agent helps, harms, or does neither.

The epilepsy-only analysis offers another warning. Across the GLP-1 trials, it counted 15 epilepsy events in treatment groups and 17 in placebo groups; the difference was not statistically significant. The combined endpoint may therefore reflect fewer acute seizures rather than less epileptogenesis, or it may simply lack enough events to answer the epilepsy question.

GLP-1 receptor agonists have a seizure-risk signal worth studying. They do not yet have evidence strong enough to justify prescribing them for seizure prevention.

What this means in practice

No GLP-1 receptor agonist is indicated for epilepsy or seizure prophylaxis. People with type 2 diabetes may have sound metabolic, cardiovascular, or weight-management reasons to use these medicines with a clinician. The possible seizure benefit should remain a secondary, low-confidence consideration—not the reason to start one.

The findings also do not automatically apply to younger people, people without diabetes, or people using semaglutide or related drugs solely for obesity. The randomized trials enrolled people with type 2 diabetes and substantial cardiovascular or renal risk, a population with a heavier vascular burden and a different mix of epilepsy causes.

The next useful study would prospectively define and adjudicate seizures, compare GLP-1 drugs with active treatments that achieve similar glycemic control, and separate stroke-mediated events from events that occur without a preceding vascular injury. Until then, the honest conclusion is modest: the signal is biologically coherent and independently echoed in real-world data, but it remains hypothesis-generating.