Patients on a GLP-1 drug for diabetes or weight, who then have a stroke severe enough to need a clot pulled out of a brain artery, were dying at about a third the rate of similar patients who had not been on a GLP-1. That is the headline a new propensity-matched cohort puts on the table, and it runs through the math hard enough that the headline is worth taking seriously.
The paper, published online May 30 in the Journal of Clinical Neuroscience ↗, pulls from the TriNetX research network. The researchers identified 286 adults who had been on a GLP-1 receptor agonist within three months before an acute ischemic stroke that required endovascular thrombectomy (the procedure where a catheter is threaded into the brain and the clot is mechanically pulled out), and 19,263 thrombectomy patients who had not been on a GLP-1 in the same window. After 1-to-1 propensity matching on age, sex, comorbidities, prior medications, body mass index, hemoglobin A1c, and the individual components of the National Institutes of Health Stroke Scale (the standardized scoring system clinicians use to rate how bad a stroke looks at presentation), 261 matched pairs remained. Both cohorts were then followed for three years.
All-cause mortality was 11.5 percent in the GLP-1 group and 29.9 percent in the unexposed group, a hazard ratio of 0.334 (95 percent confidence interval 0.219 to 0.508), p less than 0.001. Plain reading: the GLP-1 patients died at roughly one third the rate of their matched controls over three years. Inpatient hospitalizations over the same window ran 39.8 percent in the exposed group versus 54.8 percent in the unexposed group, a hazard ratio of 0.514 (95 percent CI 0.398 to 0.663), again p less than 0.001.
What the matching is supposed to do
Real-world cohort studies live or die on whether the matching makes the two groups exchangeable on every variable except the exposure. The pre-index window (the GLP-1 had to be on board within three months before the stroke, not after) was specified to handle a known statistical trap called immortal time bias, where patients who survive long enough to start a drug post-event look like the drug is helping them but really they are just the patients who survived long enough to start it. The matching variables include the right metabolic ones (HbA1c, BMI) plus the stroke-severity composite (NIHSS items, each scored separately rather than collapsed into a single sum). What that means in practice is that if a patient came in with an arm-motor score of 3 on the NIHSS in the GLP-1 cohort, their match in the unexposed cohort was scored 3 on the same item, not 3 on the total.
Two things the matching cannot do. It cannot adjust for variables that are not in the database, which is the standard caveat on any TriNetX analysis. It cannot rule out reverse causation through detection bias, where patients who get to a clinic to receive a GLP-1 prescription are systematically healthier in other ways than patients who cannot, in a way that overlaps with stroke outcomes but is not captured by the listed covariates. Those caveats apply to every real-world cohort study. They are not a reason to ignore the result. They are a reason not to call it causal.
Why the size of this is plausible
GLP-1 receptor agonists have been credited with cardiovascular and cerebrovascular benefit signals for years now, anchored on the LEADER, REWIND, and SUSTAIN-6 cardiovascular outcomes trials and then layered on by smaller real-world stroke cohorts. The biology is half-explained. Chronic GLP-1 receptor activation reduces vascular inflammation, improves endothelial function, slows atherosclerotic plaque progression, and lowers post-ischemic neuronal apoptosis in animal stroke models. A three-month pre-stroke exposure window is enough for the chronic mechanisms to be in play, even if the acute stroke itself is too fast for the drug to be doing anything new on the day of the event.
The size of the mortality effect, an HR of 0.334, is at the aggressive end of what GLP-1 cerebrovascular cohorts have produced. Previous real-world stroke analyses have landed roughly in the 0.5 to 0.7 range for similar outcomes. A 0.334 in post-thrombectomy patients specifically suggests the cohort is enriched for patients who would otherwise have done very badly (the thrombectomy population is, by selection, the severe-stroke population), and that the GLP-1 effect, whatever it is, may be largest in exactly that subgroup. The upper bound of the 95 percent CI still sits at HR 0.508, well past the null, which is consistent with that read.
What this is not
The paper does not name which specific GLP-1 drugs the exposed cohort was on. The TriNetX exposure flag is class-level, so the result is a GLP-1-class signal, not a drug-specific one. That matters because the field has been pushing toward arguing about which drug is better at which extra-glycemic outcome, and a class-level cohort study cannot answer that. It can only say the class, on average, is associated with better post-thrombectomy outcomes.
The paper also does not claim causality. Its conclusion is that GLP-1 exposure was associated with better outcomes. That is the right framing. What it sets up is the next, harder study: a prospective trial of pre-stroke GLP-1 maintenance in high-cardiovascular-risk patients, with stroke severity and mortality as prespecified outcomes. Until that runs, the HR of 0.334 is a strong real-world signal, not a clinical instruction.