Do GLP-1 Drugs Reduce Disease—or Slow Ageing?

The strongest conclusion is a confirmed cardiovascular benefit for people with established cardiovascular disease, overweight or obesity, and no diabetes. Lower all-cause mortality is a supporting signal, but it did not pass the complete confirmatory testing sequence. A 2026 proteomic analysis adds a biomarker lead; two Alzheimer’s disease trials did not detect benefit on their primary cognition-and-function outcome. These findings require population- and endpoint-specific judgments and do not establish a general slowing of ageing.

Correction, 20 September 2026: After obtaining the SELECT main report and mortality paper, we corrected a mislabeled heart-failure endpoint and withdrew the fixed-duration interpretation of “67 people treated for 3.4 years.” We also corrected overly certain infection and subgroup claims, added patient-reported health and serious adverse-event results, and incorporated 2026 cognitive evidence. The original three figures and scenario calculations have been replaced.

01What SELECT confirmed

SELECT randomized 17,604 participants to weekly subcutaneous semaglutide 2.4 mg or placebo, alongside standard care. Participants were at least 45 years old, had BMI at least 27 and established cardiovascular disease, and did not have diabetes. Mean age was 61.6; 72.3% were men and approximately 84% were White. Mean follow-up was 39.8 months and median follow-up 41.8 months. This was secondary prevention in a high-risk population, not a direct test in normal-weight people or people without cardiovascular disease.s1s4s6

The table consistently uses the main report’s Table 2 definitions. Counts represent participants experiencing an endpoint during observation; hazard ratios come from time-to-first-event models. These are different statistics.

CV death had p=0.07, exceeding its prespecified nominal two-sided threshold of 0.023, so subsequent confirmatory testing stopped. The all-cause mortality interval below 1 is a meaningful supporting signal, but cannot bypass this multiplicity constraint to become confirmed lifespan extension. Lack of confirmation does not establish lack of effect.s1s3

The original article mislabeled the HF composite as hospitalization/urgent visits alone. The latter has different counts and an interval crossing 1. FDA Table 5 also reports fatal plus nonfatal MI and stroke, with counts of 243 versus 334 and 160 versus 178. These differ from the nonfatal endpoints above; their counts and hazard ratios must not be mixed.s2

02Absolute benefit requires a defined time horizon

The crude event fractions, 569/8803 and 701/8801, differ by 1.501 percentage points. The reciprocal is 66.6. That arithmetic is correct, but observation lengths vary. Calling this “67 people treated for 3.4 years to prevent one event” assigns a fixed-time meaning that was not estimated. A three- or four-year NNT requires cumulative-incidence differences at that time, with appropriate censoring and competing-event handling. We have neither participant data nor digitized curves, and withdraw the fixed-duration NNT.

Using an HR as a risk ratio in transport calculations is also only an approximation. With a fixed horizon, proportional hazards and no competing event, the appropriate relationship is:

Treated risk = 1 − (1 − control risk)^HR

Assuming three-year control risk of 8% and HR 0.80 gives model NNT 64.6; with control risk 4%, it gives 127.1. These are deliberately specified scenarios, not observed SELECT three-year risks or estimates for healthy people. The plotted range varies only HR and does not include uncertainty about the target population, long-term adherence or baseline risk.

Adding constant competing hazards shows why the same event HR can imply different absolute benefits: more early deaths from another cause leave fewer people alive to experience the event of interest. The right panel fixes control event risk at 8% to illustrate this relationship. In particular, more competing deaths can make target-event reduction look larger without improving overall health. This is not a net-benefit model or an estimate of SELECT outcomes in a world without COVID-19.

03What the mortality breakdown tells us

There were 375 versus 458 all-cause deaths, a gap of 83: 39 CV and 44 non-CV deaths. Infection deaths numbered 62 versus 87, including 43 versus 65 COVID-19 deaths. Four mutually exclusive categories therefore contribute count gaps of 39 CV, 22 COVID-19, 3 other infection and 19 other non-CV deaths. Deaths of undetermined cause were classified as CV deaths.s3

The ratios 25/44≈57% and 22/25=88% describe the composition of observed count differences. They are neither fractions mediated through an anti-inflammatory pathway nor effects in a hypothetical pandemic-free world. The investigators proposed competing mortality as a possible explanation for convergence of CV curves during the pandemic; this does not identify a counterfactual net CV benefit.

Reported COVID-19 cases numbered 2108 versus 2150, without a detected difference. Related serious adverse events numbered 232 versus 277, and COVID-19 deaths 43 versus 65, favoring semaglutide. These analyses lack full multiplicity protection. Failure to detect an infection-rate difference does not establish identical infection processes. Restricting a comparison to people infected after randomization also forfeits the original guarantee of group comparability. Deaths analyzed by randomized assignment and conditional comparisons among infected participants must remain distinct; “proven to improve survival after infection only” was too strong.

04Benefit and tolerability both matter

The main report found fewer serious adverse events overall and more discontinuation because of adverse events. Both observations belong in the assessment.s1

The main report gives premature permanent discontinuation as 26.7% versus 23.6%; the 2024 FDA label separately gives 31% versus 27% discontinuing before trial end. We have not obtained enough documentation to reconcile the definitions or observation windows fully. The table prioritizes the explicitly defined adverse-event discontinuation counts.

The 2024 label also reports hip/pelvis fractures in women, 24/2448 versus 5/2424, and in participants aged at least 75, 17/703 versus 4/663; urolithiasis occurred in 1.2% versus 0.8%. These subgroup or small-event signals deserve reporting but cannot provide precise individual risk predictions. The fracture subgroups may overlap. SELECT used targeted safety collection, so its table is not a complete inventory of minor symptoms. This article evaluates evidence and does not advise starting, stopping or adjusting treatment.s2

05Weight loss and inflammation: no identified causal shares

At 104 weeks, the main report gives weight changes of −9.39% versus −0.88% and a placebo-corrected relative hsCRP change of −37.8%. The hsCRP analysis also reports a within-semaglutide ratio to baseline of approximately 0.62 at week 52, maintained at week 104. A within-arm ratio and a placebo-corrected ratio require different labels.s1s5

Adding baseline LDL cholesterol, statin use, baseline hsCRP and time-varying hsCRP shifted the MACE HR from 0.80 to 0.86 and the all-cause mortality HR from 0.81 to 0.91. This is a change in a conditional model’s treatment coefficient, not a measure of the fraction of benefit mediated by inflammation. Post-treatment common causes, concurrent weight/metabolic changes, measurement error and HR noncollapsibility can matter; the model also adds other baseline covariates. The old calculation mislabeled the ratio of adjusted to unadjusted log coefficients as attenuation. It is the remaining coefficient ratio; neither quantity identifies causal mediation here.

Higher baseline hsCRP predicts poorer prognosis. Nonsignificant treatment interactions mean that clear heterogeneity was not detected, not that subgroup effects are equivalent or that hsCRP can never predict treatment benefit. Multiple analyses of SELECT remain publications from one trial, not independent replications.

06Cognitive evidence in 2026: scores versus clinical outcomes

A post hoc SELECT analysis examined 2970 participants aged at least 65 with paired baseline and week-104 serum samples. A 25-protein Dementia SomaSignal Test predicted five- and twenty-year dementia risk, yielding reported “ORs” of 0.74 (0.65–0.85) and 0.91 (0.88–0.94). The analysis models change in the logit of predicted probability and exponentiates the treatment coefficient. These estimates concern predicted-score trajectories, not observed dementia incidence, and cannot be translated into “26% fewer dementia cases.”s13

Specific limits matter: complete paired samples only, with missingness assumed completely at random; post hoc analysis without multiplicity control; and serum measurements applied to a score validated in EDTA plasma, with calibration transport still needing assessment. Some extremely low five-year predictions were clipped. SELECT did not conduct systematic cognitive assessment or dementia adjudication for this analysis. Changing peripheral metabolic or inflammatory proteins could change the score without proportionately changing neurodegeneration. Attenuation after BMI adjustment likewise does not prove that the remaining coefficient represents direct neuroprotection.

The 2026 EVOKE and EVOKE+ trials address a more direct clinical question. Together they randomized 3808 adults aged 55–85 with amyloid-confirmed early symptomatic Alzheimer’s disease to oral semaglutide or placebo. Differences in 104-week CDR-SB cognition-and-function change were −0.08 (−0.35 to 0.20; p=0.57) and 0.10 (−0.17 to 0.38; p=0.46), respectively. Neither trial detected primary-endpoint benefit. These values come from the published abstract, cross-checked against the authors’ congress poster and its stated hybrid HTC estimand. We have not obtained the complete supplementary methods or independently refit the models.s14s15

The poster also reports changes in some cerebrospinal-fluid markers in a small substudy, without multiplicity-adjusted comparisons. Biomarker improvements can coexist with a negative primary clinical result. EVOKE concerns established early AD, an oral formulation and a particular follow-up period; it does not rule out dementia prevention in every other population. It does prevent treating GLP-1-related biomarker changes as general proof of cognitive protection.

07Projected life-years and the broader ageing claim

SELECT did measure patient-reported health status. At 104 weeks, the between-group change difference was 0.01 (0.01–0.02) for EQ-5D-5L and 1.60 points (1.16–2.04) for its visual analogue scale. These were supportive outcomes without multiplicity-adjusted intervals. Statistical significance alone does not establish clinical importance, and neither measure directly estimates a whole-body ageing rate.s1

The often-quoted projection of 1.9 additional life-years applies SELECT effects to life tables from a UK electronic-health-record cohort. It assumes persistent long-term effects and does not fully incorporate discontinuation beyond trial follow-up. Some endpoint point estimates are set to 1 according to statistical significance. The paper acknowledges these limitations. Its uncertainty simulation primarily propagates selected HR uncertainties, not all transport assumptions, and the interval is not an observed range of lifelong benefit.s10

The strongest current use of the evidence is to assess specified disease benefits and burdens in specified populations. A general slowing-of-ageing claim needs testable evidence connecting long-term function, several disease domains, safety and survival, plus validation that treatment-induced biomarker changes predict real outcomes. SELECT’s primary cardiovascular conclusion stands; universal anti-ageing effects and a definite number of extra years do not follow directly.

Evidence checked through 20 September 2026. This revision examined the main and mortality reports’ load-bearing sections, recomputed public aggregates and hypothetical survival models, and replaced the figures. There are no participant data, Cox/imputation/cognitive-model refits, or complete EVOKE supplementary methods. The SELECT discontinuation discrepancy remains unresolved. Codex performed revision, self-review and editing; this is not human clinical professional review.

The reproduction package contains source-located aggregate inputs, code, independent numerical checks and five figures, without third-party article originals or personal data.