EVOKE and EVOKE+: Why Semaglutide Did Not Slow Early Alzheimer's, and What the Biomarkers Showed

Semaglutide did not slow early Alzheimer's disease. In two identical phase 3 trials, EVOKE and EVOKE+, which together randomised 3,808 people with mild cognitive impairment or mild dementia due to Alzheimer's disease, once-daily oral semaglutide produced no difference from placebo on the primary endpoint: change in the Clinical Dementia Rating Sum of Boxes (CDR-SB) at 104 weeks 1. Novo Nordisk terminated both trials before their planned one-year extensions 1. The full results were published in The Lancet in March 2026 3.

Some Alzheimer's-related biomarkers did move. That is the part of the readout most likely to be misread, and the reason this article exists. It does not re-explain how semaglutide acts at the GLP-1 receptor, which is covered separately on this site. It covers the hypothesis, the design, the null clinical result, what the biomarkers showed and did not show, the adverse events, and the argument now being made that the drug class belongs in prevention rather than treatment — an argument these trials did not test.

The signal the trials were built to test

The rationale came largely from data not collected to answer the question. A 2022 analysis pooled three double-blind cardiovascular outcome trials of GLP-1 receptor agonists in type 2 diabetes, covering 15,820 patients, and reported fewer dementia diagnoses in those randomised to active drug than to placebo, with a hazard ratio of 0.47 5. The same paper examined a Danish nationwide registry cohort of 120,054 patients and found a hazard ratio of 0.89 for each additional year of exposure 5.

Both findings carry recognisable weaknesses. In the pooled trials, dementia was not a prespecified endpoint, diagnoses were few, and the confidence interval was wide. In the registry, who receives a newer diabetes drug is not random, and the factors that shape prescribing also shape dementia risk. The authors presented the results as grounds for a randomised trial, which is what they were 5. A mechanistic case sat alongside: GLP-1 receptors are expressed in the brain, and receptor agonism has anti-inflammatory and metabolic effects in preclinical models. EVOKE and EVOKE+ were the test.

Design: two trials, one question

The two trials began in 2021 and were identical except for one criterion: EVOKE excluded participants with significant subcortical vascular disease, and EVOKE+ admitted them 4. Participants had mild cognitive impairment or mild dementia, with Alzheimer's pathology confirmed by amyloid PET or cerebrospinal fluid testing. Mean age was about 72; around half had overweight or obesity, and about 14% had type 2 diabetes 4. Oral semaglutide or placebo was taken once daily on top of standard care, escalated to a target exposure. The primary endpoint was change in CDR-SB from baseline to week 104 1.

The primary result: no separation from placebo

Both trials missed. On CDR-SB at week 104, the semaglutide and placebo groups did not differ in either trial 13. Investigators presenting the results described the two groups as declining in parallel throughout follow-up — not diverging late, and not converging after an early gap 4. The secondary measures, including an activities-of-daily-living scale for mild cognitive impairment, the Montreal Cognitive Assessment, ADAS-Cog13, the Mini-Mental State Examination and a composite score, showed indistinguishable trajectories. Time to progression from mild cognitive impairment to dementia did not differ 4.

A null result replicated across two trials of this size is informative in its own right. The combined sample is larger than those of most Alzheimer's programmes that did produce approved drugs. The absence of any secondary signal removes the usual escape route, in which a missed primary endpoint is followed by a favourable secondary. And parallel decline leaves no subgroup-shaped hint in the curves themselves. The sponsor drew the same conclusion and stopped the programme 1.

What the biomarkers showed

The biomarker results are where the readout becomes easy to misstate. In a cerebrospinal fluid substudy of roughly a hundred participants per group, seven markers showed nominally significant reductions of 10% or less on semaglutide compared with placebo, among them phosphorylated tau species, total tau, neurogranin and YKL-40, a marker associated with glial inflammation 4. Novo Nordisk's own summary stated that Alzheimer's-related biomarkers improved but that the improvements did not translate into delayed decline 1.

Four features limit what these shifts can mean. They are small. They were measured in a substudy of a few hundred people rather than the full population, and they reached nominal significance rather than serving as prespecified confirmatory endpoints. They did not appear in plasma, where the same markers were measured in far more participants. And investigators noted that shifts of this size fall in the range seen in trials of other agents that showed no clinical efficacy 4.

A systemic anti-inflammatory effect is plausible, and the fall in plasma hs-CRP is consistent with what the drug class does in other settings. But systemic inflammation is not Alzheimer's pathology, and the two plasma markers of neural injury moved slightly in the unfavourable direction. A biomarker that moves while the clinical endpoint does not is, by definition, not functioning as a surrogate in this setting. It is evidence that the drug did something biological, not evidence that it did something useful.

Adverse events and discontinuation

The safety profile matched what the drug produces in its approved indications, and no new adverse effects were identified 3. Gastrointestinal events — nausea, reduced appetite, diarrhoea and vomiting — were more frequent on semaglutide, and more participants on semaglutide than on placebo discontinued because of them 4. Mean body-weight change was 5.8% below baseline on semaglutide against a gain of 0.6% on placebo 4.

Body-weight change matters here for a reason specific to this population. In older adults with cognitive impairment, unintended reduction in body weight is generally treated as an adverse sign, associated with frailty rather than with benefit. A treatment that suppresses appetite therefore carries a cost in this group that it does not carry in an obesity trial. EVOKE was not designed to weigh that cost against a benefit, because no benefit appeared. Any future trial of the class in older people will have to.

Prevention, not treatment: the reading that followed

After the readout, much of the discussion moved to a narrower claim: that GLP-1 receptor agonists might reduce the incidence of dementia when taken for years before symptoms appear, even if they cannot alter the course of established disease 4. An accompanying commentary in The Lancet took up the question of where the programme leaves the hypothesis 2. The distinction is legitimate. Treating symptomatic Alzheimer's disease means acting on a brain in which amyloid, tau and synaptic loss are already advanced. Preventing dementia means acting earlier, possibly on the vascular and metabolic contributors that the original signal in people with diabetes may have been picking up.

It is also, at present, a hypothesis rather than a finding, and it should be labelled as one. The evidence for prevention is the same pooled-trial and registry evidence that motivated EVOKE 5, and EVOKE has now shown that this evidence did not predict a treatment effect. A prevention trial would need to enrol people without cognitive impairment, follow them for years, and count dementia diagnoses as a prespecified endpoint. None of that has been done. Until it is, the accurate summary is that semaglutide has been tested in early Alzheimer's disease and did not work, and that whether the class prevents dementia is untested.

Limits of the trials themselves

• The formulation was oral. Absorption of oral semaglutide is low and variable, so central exposure may have been lower than an injectable form would achieve. Whether that matters is unknown, because no injectable arm was tested.
• Two years is short for a disease that develops over decades, although approved anti-amyloid agents separated from placebo on CDR-SB over comparable or shorter periods.
• The cerebrospinal fluid substudy was small and its findings nominal.
• Participants with type 2 diabetes were a minority, about one in seven, so the population differed from the one in which the original signal arose.
• The planned extensions were terminated, so no data exist on longer exposure in this population.

What EVOKE settled

EVOKE and EVOKE+ settled the treatment question for this drug, in this formulation, in this population: two concordant trials, 3,808 participants, no clinical effect at two years 13. They also produced a small, compartment-specific biomarker signal that is compatible with a biological effect of the drug and incompatible with a claim that the effect matters clinically. The association that started the programme remains unexplained 5. That is what a well-run negative trial is for. It removes one hypothesis cleanly and leaves a sharper one — prevention — that now needs a trial of its own.