Semaglutide and Cognition: 13,007 Records, One Null Result
A retrospective cohort study published in BMJ Mental Health on 20 August 2026 asked whether starting semaglutide is associated with fewer recorded cognitive signs and symptoms in people already carrying a psychiatric diagnosis. The cohort was 13,007 adults. The answer, against three of four comparators, was yes.
The fourth comparator is the interesting one, and almost no coverage will mention it.
A framing note first. Semaglutide is an approved medicine, this is a study of patients, and semaglutide is not in the Peptra Labs catalogue. Nothing here is guidance about human use.
How the semaglutide cohort was assembled
The work comes from the University of Oxford together with Holmusk, using NeuroBlu Data, a set of de-identified multicentre US electronic health records spanning 1999 to 2024.
Eligibility was tighter than most database studies. Adults needed any ICD-9 or ICD-10 psychiatric diagnosis, plus at least one semistructured, clinician-rated cognitive assessment in the 12 months before starting an antidiabetic drug, and at least one in the 12 months after. Requiring a structured assessment on both sides of the index date is what makes the outcome measurable rather than inferred from billing codes.
The primary outcome was a composite score from 0 to 100, where 0 means no cognitive signs or symptoms recorded and 100 means signs present across all domains, drawn from clinician ratings of memory, attention, orientation and other cognitive functions. Higher is worse. The analysis used a parametric g-formula and reported mean ratios, adjusting for baseline score and prespecified covariates.
Four comparator strategies were prespecified: sitagliptin, empagliflozin, glipizide, and no antidiabetic treatment at all. Four is unusual and it is a strength. A single comparator tells you about one contrast. Four tells you whether the contrast survives changing the reference.
What the analysis found
Of 13,007 individuals, mean age 62.0 years and 48.4 percent female, 1,261 initiated semaglutide.
| Comparator | Mean cognitive score | Mean ratio vs semaglutide (11.14) | 95% CI | p |
|---|---|---|---|---|
| No antidiabetic treatment | 14.91 | 0.75 | 0.65 to 0.85 | 0.00002 |
| Empagliflozin | 13.69 | 0.81 | 0.69 to 0.94 | 0.016 |
| Glipizide | 13.46 | 0.83 | 0.70 to 0.93 | 0.041 |
| Sitagliptin | 12.52 | 0.89 | 0.77 to 1.02 | 0.50 |
The p values are Bonferroni corrected, which matters when four comparisons are run against the same exposure.
Associations were stronger for memory and for broader cognitive functions than for the narrower domains. Effect sizes looked similar across major depression, psychosis and bipolar disorder, but reached consistent significance only in major depression, which is the largest of the three groups and therefore the best powered.
The comparator semaglutide did not beat
Sitagliptin is a DPP-4 inhibitor. It works by slowing the enzymatic breakdown of the body’s own GLP-1, so that endogenous incretin signalling lasts longer.
It is the one comparator in this study that also acts on the GLP-1 axis, and it is the one comparator against which semaglutide showed no significant difference. The confidence interval, 0.77 to 1.02, crosses 1 by a small margin.
The authors do not build an argument on this and neither should a reader. Sitagliptin patients also had the lowest comparator score at baseline of assessment, the comparison is not randomised, and a p of 0.50 after correction is a long way from a signal. But the pattern is at least consistent with an incretin-axis explanation rather than a semaglutide-specific one, and it is exactly the kind of detail that disappears when a study is summarised as showing a drug improves cognition.
What the design cannot deliver
The authors are explicit in their own clinical implications section: randomised clinical trials are needed to confirm causality and clinical utility.
Three limits sit underneath that sentence.
The outcome is what a clinician wrote down. A semistructured clinician-rated assessment is better than a diagnosis code, and it is still a rating made by someone who can see whether the patient is on semaglutide. Cognitive signs recorded in routine care are not neuropsychological test scores.
Who gets started on semaglutide is not random. In a US electronic health record spanning 1999 to 2024, access to a newer and more expensive agent tracks insurance, clinic type and comorbidity burden. The g-formula adjusts for prespecified covariates, which means the covariates the researchers could see and chose.
The data provider is a co-author. Two authors report employment with and equity ownership in the corporate group behind the record set, and a third reports consultancy fees from the same company. That is disclosed properly in the paper and it is not a reason to discount the result. It is a reason to read the choices about cohort construction as choices made by people with an interest in the platform performing well.
Where this sits in the GLP-1 literature
This is the third distinct non-metabolic endpoint semaglutide and its class have produced in as many weeks.
We reported a prespecified analysis of inflammation in the SELECT trial, where an inflammatory marker fell before weight did, and a 2026 review of GLP-1 and mitochondrial biology that proposed an organelle-level mechanism while conceding that no trial has ever included a prespecified mitochondrial endpoint. Cognition now joins inflammation and mitochondrial function on the list of things the class appears to touch, and for which the mechanism is inferred rather than measured.
The published cognitive literature thins quickly as you move along the class. Tirzepatide, the dual GIP and GLP-1 agonist, has a 2026 scoping review pulling together preclinical experiments and early clinical work. Retatrutide, the triple agonist, is earlier in development and has no comparable body of cognitive data.
It is worth remembering that peptides have been taken to cognitive endpoints before, with sobering results. Cerebrolysin, a peptide preparation studied for decades, was assessed in a 2019 Cochrane review for vascular dementia which found the evidence insufficient to support routine use. A large cohort and a plausible mechanism have not historically been enough.
For laboratories, all of the above is supplied on a research use only basis, and our ranking of the most-studied research peptides tracks how much published work each compound actually has behind it.
References
- De Giorgi R, Lipunova N, Mathias E, Shang F, Taquet M. Cognitive signs and symptoms in people with a psychiatric diagnosis on semaglutide: a retrospective cohort study of 13 007 patients in the USA. BMJ Mental Health, 20 August 2026. doi 10.1136/bmjment-2026-302828
- Zhou T, Tang H, Zhang B, et al. Association between glucagon-like peptide-1 receptor agonists and risk of dementia in older adults with type 2 diabetes: A target trial emulation. Diabetes, Obesity and Metabolism, 22 December 2025
- Hawas Y, Abouzid M, Gadelmawla AF, Ewis DK. The Neuroprotective Potentials of Dual GIP/GLP1-RA (Tirzepatide): From Preclinical Experiments to Clinical Trials: A Scoping Review. Brain and Behavior, August 2026
- Cui S, Chen N, Yang M, Guo J. Cerebrolysin for vascular dementia. Cochrane Database of Systematic Reviews, 11 November 2019
The products referenced on this site are supplied for laboratory research use only. They are not medicines and are not intended for human or veterinary use. This article summarises published research for informational purposes and is not medical advice. Statements about third-party studies belong to their authors.
