Semaglutide, the drug sold as Ozempic and Wegovy, slowed a blood-based score that predicts a person's odds of getting dementia, according to a new look at the largest trial ever run on the drug. The catch is that the trial measured the score, not the disease.
That analysis published August 8 in Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring ↗. It reused blood samples from SELECT, the heart-outcomes trial that put semaglutide in front of thousands of people with excess weight and heart disease but no diabetes. The researchers pulled 2,970 participants aged 65 and older, all randomized to semaglutide 2.4 milligrams or a placebo, and ran their stored blood through a validated test called the dSST. That test reads 25 proteins and turns them into a number: a person's predicted risk of developing dementia over the next 5 and 20 years.
Over two years, the score climbed less in the people on semaglutide. For 5-year risk, the drug group's predicted event rate rose 26 percent less than placebo (odds ratio 0.74, meaning roughly a quarter lower odds of landing in a worse category). For 20-year risk the gap was smaller, about 9 percent (odds ratio 0.91). People on the drug were 36 percent less likely to shift into a higher dementia-risk band.
A prediction is not a diagnosis
Here is the load-bearing caveat, and the authors are clear about it. The dSST is a forecast, not a finding. Nobody in this analysis was newly diagnosed with dementia. The proteins in the panel track biological changes that tend to precede the disease, and the test has been validated against real outcomes, but moving the forecast is not the same as preventing the illness. It is the difference between lowering someone's cholesterol and preventing their heart attack. The first makes the second more likely to follow. It does not guarantee it, and in a two-year window you cannot watch it happen.
That is why the second study that landed the same day matters, and why it is not the clean confirmation it looks like at first.
The cohort found real cases, in the vaguest bucket
A separate team, publishing in The Journal of Prevention of Alzheimer's Disease ↗, went after actual diagnoses instead of a score. They used Taiwan's national insurance records from 2011 to 2021 and built 10,783 matched pairs of adults over 50 with type 2 diabetes, one of each pair started on a GLP-1 drug, the other on long-acting insulin. Then they counted who developed dementia. There were 375 cases. The rate was 4.86 per 1,000 person-years among GLP-1 users versus 7.56 among insulin users, which works out to about a third lower overall risk (hazard ratio 0.64).
Read the subtypes, though, and the picture gets uncomfortable. The entire benefit sat in "unspecified dementia," the catch-all code a clinician uses when the type is not pinned down (hazard ratio 0.41, a roughly 60 percent reduction). For the two specific, better-adjudicated diagnoses, the drug did nothing statistically: Alzheimer's disease came back at a hazard ratio of 1.48 and vascular dementia at 1.66, both non-significant and both pointing the wrong way. Liraglutide (Victoza and Saxenda) and dulaglutide (Trulicity) carried the unspecified-dementia signal.
A benefit that shows up only in the fuzziest diagnostic category, and vanishes in the sharp ones, is exactly the fingerprint you would expect from coding noise and confounding rather than a drug protecting neurons. The comparison group is another tell. Insulin tends to go to people whose diabetes is further along and harder to control, so GLP-1 users start out healthier in ways the matching cannot fully erase.
Two shapes of evidence, one gap
Put the two together and you get the honest state of the question. The randomized trial has the strong design but measured a proxy. The real-world cohort measured the disease but in a soft category, with a comparator that stacks the deck. Both lean the same direction, which is genuinely more than nothing, and it lines up with earlier signals, including our piece on how GLP-1 users with a high-pressure brain condition had half the cognitive decline ↗, where every outcome favoring the drug also carried the healthy-user fingerprint.
What is missing is the study that closes it: a trial that starts people on a GLP-1 drug and counts dementia diagnoses years later. Those trials exist and are running. Until one reads out, the drug has moved a forecast and nudged a fuzzy tally, and neither is proof.
On peptidemodel, the drugs named here have live cards, semaglutide ↗, liraglutide ↗, and dulaglutide ↗, all acting on the GLP-1 receptor ↗, the switch these results keep circling back to.