Nobody has measured that, and what has been measured is a predictor of it — a different thing, and worth keeping separate from the question people are actually asking. In a post hoc analysis of SELECT covering 2,970 adults aged 65 and over, a validated 25-protein blood score that forecasts dementia rose in both arms. It rose less on semaglutide, at an odds ratio of 0.74 on the five-year signature and 0.91 on the twenty-year one.
A predictor moving is not the same as a disease not happening, and the gap between those two sentences is the whole of this page. The one study that looked at cognitive impairment itself found 33 cases on oral semaglutide against 73 on a comparator in 246 elderly outpatients, a reduction its regression put at 77% with no confidence interval published[2]. A larger study of the same drug against another DPP-4 inhibitor found nothing, which is why those two results have to be read together.
Nobody in this analysis developed dementia, and nobody avoided it. What was measured is a blood score built to predict it, and the distance between moving a predictor and moving an outcome is the same distance a biological clock sits at from a lifespan.
What was measured
Serum samples taken at baseline and at week 104 from 2,970 SELECT participants aged 65 and over with overweight or obesity and cardiovascular disease but no diabetes. [1] Those samples were run through a validated test that reads 25 proteins and returns a predicted risk of all-cause dementia over five and twenty years.
This is a post hoc analysis. SELECT was designed to test whether semaglutide prevents cardiovascular events — a composite endpoint with its own reading problems — not to answer anything about cognition, and the proteomic question was asked of its samples afterwards.
Rate, not level
Predicted dementia risk went up in both arms over the two years, which is what happens to a risk score as a cohort of 65-year-olds gets older. It went up less on semaglutide: 2.5-fold less on the five-year signature and 1.67-fold less on the twenty-year one.
That is the paper’s own framing — reduced increases — and it is worth keeping, because “reduced dementia risk” describes something that did not happen here in two separate ways.
What a validated score validates
That people with higher scores develop dementia more often than people with lower ones. That is what validation of a predictive test establishes, and it is a fact about the population the test was built on.
It does not establish that lowering somebody’s score lowers their risk. A marker can track a disease faithfully and still be a passenger rather than a driver. In which case moving it changes the reading and not the future — the reason the markers easiest to move are usually the ones proving least.
What would settle it
A trial with dementia as its endpoint, in people followed long enough to get it. That is expensive, slow, and the reason proteomic signatures are attractive in the first place.
Until one exists, this is a promising result about a score. The trial it came from was asking about hearts, reported honestly by authors who say their drug slowed the progression of a signature rather than of a disease — which is more care than most conclusions take.