In a database of nearly 878,000 people with obstructive sleep apnea, the ones who started a GLP-1 drug around the time of their diagnosis had about a quarter fewer ischemic strokes over the next year than the ones who did not. The same protective signal showed up for bleeding in the brain, emergency visits, hospital stays, and death. It also shrank every year the patients were followed, and the death number is large enough to be suspicious.
The analysis, published online June 26 in Respiratory Medicine ↗, drew on the TriNetX US Collaborative Network, a federated pool of electronic health records. The authors found adults diagnosed with obstructive sleep apnea between 2016 and 2025, defined exposure as starting a GLP-1 receptor agonist ↗ within six months before to one month after that diagnosis, then matched 438,844 treated patients to an equal number of untreated patients with similar baseline health. Obstructive sleep apnea is the condition where the airway keeps collapsing during sleep. It travels with obesity and raises the risk of stroke and other vascular damage, which is the whole reason a drug class built for weight and blood sugar is being looked at in these patients at all.
What the numbers said
At one year, clot-driven ischemic stroke was 25 percent lower in the GLP-1 group (hazard ratio 0.75). Bleeding inside the skull, called intracranial hemorrhage, was 56 percent lower (HR 0.44). Emergency department visits were 23 percent lower, hospital admissions 41 percent lower, and death from any cause was 62 percent lower (HR 0.38). Every one of those comparisons cleared the usual statistical bar.
The pattern survived two stress tests. When the authors looked only at patients on CPAP, the airway-pressure machine that is the standard sleep-apnea treatment, the direction of benefit held. When they looked only at tirzepatide ↗, the dual GIP and GLP-1 drug sold as Mounjaro and Zepbound, it held there too.
Why the death number is the tell
A 62 percent drop in death from starting a single drug is the kind of result that should make a careful reader lean back, not forward. Studies built from health records capture people who chose to start a medication and could afford to stay on it, and that group tends to be healthier, more engaged with their care, and more likely to survive for reasons that have nothing to do with the drug. The technical name is healthy-adherer bias, and a too-large mortality effect is its signature.
The shape of the results points the same way. Every outcome was strongest at one year and weaker at five. Ischemic stroke went from 25 percent lower at one year to 13 percent lower at five (HR 0.87). The mortality gap narrowed from 62 percent to 46 percent. A real drug effect on a chronic disease does not usually fade as exposure lengthens. A bias concentrated at the start does exactly that, because the healthiest early adopters are also the ones who drift out of a matched comparison over time. The authors say as much, calling the findings hypothesis-generating and in need of prospective trials.
This is also the second large TriNetX stroke signal for GLP-1 drugs in a month. A separate cohort tied the class to a two-thirds cut in three-year mortality after thrombectomy stroke ↗, a number with the same too-good shape and the same caveat. When a method keeps returning effect sizes that large, the method is part of the story.
What is probably real underneath
Strip out the implausible magnitude and something defensible is left. GLP-1 drugs have shown cardiovascular benefit in actual randomized trials, including the heart-event reductions ↗ measured for semaglutide ↗, and the biology that protects the heart's vessels plausibly protects the brain's. Sleep-apnea patients are a high-risk vascular group that has mostly been studied through the airway, not the bloodstream. A signal that GLP-1 drugs lower their stroke and brain-bleed risk is worth a real trial, even if the headline numbers from this one are inflated.
What the study cannot do is tell an individual patient with sleep apnea that a GLP-1 drug will cut their stroke risk by a quarter. What it can do is tell the people who design trials that the question is now specific enough to test.