Semaglutide "cuts dementia risk", say the headlines. The number is modelled, and the real trial already failed

Semaglutide — the drug sold as Ozempic and Wegovy, and increasingly familiar to British readers through weight-loss coverage — is back in the health headlines with a claim that it lowers the risk of dementia. A new analysis published in Alzheimer’s & Dementia: Diagnosis, Assessment & Disease Monitoring is the source. The finding is real. The word doing the heavy lifting in the coverage is “predicted”, and it changes what the study actually shows.

What was measured, precisely

The work is a post hoc analysis of SELECT, a large cardiovascular-outcomes trial of semaglutide. Researchers looked at 2,970 older participants — aged 65 and over, overweight or obese, with established cardiovascular disease but without diabetes — and tracked a validated 25-protein blood signature that is associated with future dementia risk. Over 104 weeks, that signature worsened more slowly in the people on semaglutide than in those on placebo.

Translated into risk, the modelling suggests the five-year predicted dementia risk rose about 2.5 times less on the drug — a 26% lower modelled rate — and the twenty-year predicted risk rose roughly 1.67 times less, an 8.8% lower modelled rate.

Notice what every one of those figures describes: a change in a blood biomarker, run through a predictive model. Not a single one is a count of people who did or did not go on to develop dementia. Nobody in this analysis was diagnosed with fewer cases of the disease. A protein pattern that tends to precede dementia moved in a favourable direction, and a model turned that movement into a risk percentage.

Why a surrogate is not an outcome

The authors themselves flag the limits, and they are not small. The analysis was post hoc, with no pre-specified correction for testing multiple endpoints. Dementia was not systematically adjudicated in SELECT — the trial was designed to measure heart attacks and strokes, not cognition — so actual dementia events were few and under-counted. The result rests on a biomarker-based predicted score standing in for the real clinical endpoint.

In drug research this is the well-worn gap between a surrogate marker and a hard outcome. Biomarkers move for all sorts of reasons; the question is always whether that movement translates into people being better off. Often it does not.

The part the good-news write-ups skip

Here the caution is not hypothetical, because semaglutide has already been put to the direct test — and it did not pass. In November 2025, Novo Nordisk reported the results of evoke and evoke+, two dedicated Phase 3 trials of oral semaglutide in 3,808 people with early Alzheimer’s disease. Those trials did not show a statistically significant slowing of disease progression on the standard clinical measure. Tellingly, semaglutide improved Alzheimer’s-related biomarkers in both trials — and it still made no measurable difference to how patients actually fared. Novo Nordisk discontinued the extension periods.

That is the exact pattern the new SELECT analysis reproduces: the markers look better, the disease does not. When the drug was tested against real cognitive decline rather than a modelled proxy, the benefit did not appear.

None of this means semaglutide is worthless for the brain, or that the biology is a dead end — GLP-1 drugs remain a legitimate area of dementia research. It means a modelled reduction in a blood signature is a hypothesis, not a result, and the one time that hypothesis was tested head-on, it failed. “Ozempic prevents dementia” is not what this study found, and it is worth knowing why before the headline sticks.