In 171,917 Swedish adults with type 2 diabetes, the ones taking a GLP-1 drug were about half as likely to have the common kind of stroke. The ones taking an SGLT2 drug were not, even though a shelf of observational studies says they should have been.
That contrast is the whole point of the paper. It was published online September 23 in EClinicalMedicine ↗ by a team at the University of Gothenburg, led by Alexandros Mavridis with Katharina Stibrant Sunnerhagen as senior author, working from the Swedish National Diabetes Register. They tracked everyone newly entered in the register between January 2013 and December 2019, dropped anyone with a prior stroke, and watched for who had one next.
What the register showed
Three drug classes were in the comparison. GLP-1 receptor agonists, the injectable gut-hormone drugs that in Sweden over those years meant mostly liraglutide ↗ and dulaglutide ↗, with semaglutide ↗ arriving late in the window. SGLT2 inhibitors, the pills that make the kidneys dump sugar into urine. And DPP-4 inhibitors, an older pill class that nudges the body's own gut hormones up a little.
Over an average of three years, 2,734 people (about 1.6 percent) had an ischemic stroke, the type caused by a clot cutting off blood to part of the brain. Another 428 had a hemorrhagic stroke, the bleeding type.
People on a GLP-1 drug had a markedly lower rate of ischemic stroke (hazard ratio 0.46, 95% CI 0.29 to 0.72), which works out to roughly half the risk of comparable patients not on one. Their total stroke rate fell about as much (HR 0.51, 95% CI 0.33 to 0.79). The bleeding strokes showed no clear pattern either way. SGLT2 inhibitors and DPP-4 inhibitors showed no clear link to any kind of stroke.
Why the method is the story
Register studies like this one usually flatter drugs. The healthier, better-resourced patients tend to end up on the newer medicine, so the newer medicine looks like it prevented the bad outcome when really it just marked the people who were going to do fine anyway. That bias, confounding by indication, is why observational drug comparisons routinely report benefits that randomized trials later shrink or erase.
Mavridis and colleagues tried to strip that bias out. They used marginal structural models with inverse probability of treatment weighting, a method that reweights the population so that drug choice is no longer tangled up with how sick each patient was, and they updated each person's treatments, lab values, and conditions over time rather than freezing them at the start. It is the closest an observational dataset gets to imitating a randomized trial.
The result is the tell. After that correction, the GLP-1 benefit held, and the SGLT2 benefit did not appear. An uncorrected version of this data, the kind most observational papers report, would very likely have handed SGLT2 inhibitors a stroke benefit they did not earn here.
The 91-study backdrop
The authors then did something most single-cohort papers skip. They searched the literature to June 2026 and pulled 91 meta-analyses to see where their numbers landed. The split is clean. Meta-analyses of randomized trials credit GLP-1 drugs with cutting stroke and call SGLT2 and DPP-4 neutral. Meta-analyses of observational studies credit both GLP-1 and SGLT2 drugs. The Swedish cohort, an observational dataset run through trial-grade statistics, sided with the randomized trials: GLP-1 yes, SGLT2 no.
That is the sentence worth keeping. When an observational study bothers to correct for confounding properly, it can stop agreeing with other observational studies and start agreeing with the trials.
None of this makes GLP-1 drugs a stroke prevention therapy on its own. This is still one country, one register, and a secondary question riding on data collected for other reasons. The direction, though, is now consistent across two very different kinds of evidence, which is more than the SGLT2 stroke claim can say.
On peptidemodel, the GLP-1 agonists in this study sit under the GLP-1R target ↗. The stroke thread runs through the section already, from the other side: in July a separate 209,180-record analysis found that GLP-1 users who had a stroke anyway arrived with milder strokes and lower 30-day death ↗. This Swedish cohort speaks to the first question, whether the stroke happens at all. The two lines of evidence now point the same way.