https://www.bmj.com/content/394/bmj-2026-100011
chatGPT(5.5paid):
Summary
This BMJ paper asks a clinically important question that the SURPASS-CVOT trial did not directly answer: what cardiovascular benefit would be expected from adding tirzepatide to usual treatment, compared with something approximating no additional cardiovascular-active therapy?
SURPASS-CVOT compared tirzepatide with dulaglutide and established that tirzepatide was non-inferior, but because dulaglutide itself reduces cardiovascular events, that comparison does not directly estimate the absolute benefit of starting tirzepatide versus background care. The authors therefore used sitagliptin as a cardiovascular-neutral “placebo proxy” in a real-world target-trial emulation.
They analysed two US insurance claims databases from 2022-25, including 52,971 people aged ≥40 with type 2 diabetes and established atherosclerotic cardiovascular disease: 35,353 starting tirzepatide and 17,618 starting sitagliptin. The analysis used propensity-score overlap weighting; the effective weighted comparison was about 7,442 patients in each group.
The principal result was:
- MACE at one year: 2.9% with tirzepatide versus 4.4% with sitagliptin
- absolute risk difference: −1.4 percentage points
- HR 0.68 (95% CI 0.58–0.80)
- estimated NNT 70 for one year.
The components are particularly informative. Myocardial infarction was reduced, HR 0.67, while stroke was essentially unchanged, HR 0.91. All-cause mortality was substantially lower, HR 0.55.
An unexpected finding was the apparent reduction in infection-related outcomes. Hospital admission for infection was reduced by about 36% (HR 0.64), infection-related mortality by about 60% (HR 0.40), and infections overall were also less frequent. The authors therefore suggest that some of tirzepatide’s mortality benefit could involve non-atherosclerotic mechanisms, rather than simply fewer cardiovascular events.
Importantly, the cardiovascular signal was not uniform. The absence of an apparent stroke effect, despite a clear MI effect, makes the result more interesting mechanistically and argues against simply interpreting the headline MACE HR of 0.68 as a uniform 32% reduction in all cardiovascular pathology.
What is novel
I think there are three distinct elements of novelty, with the methodological one perhaps more important than the pharmacological one.
1. Estimating tirzepatide versus something approximating placebo
This is the paper’s central novelty.
The randomised evidence already showed that tirzepatide performed at least as well as dulaglutide. What remained uncertain was:
How much better is adding tirzepatide than continuing contemporary background treatment?
The authors attempt to answer this by substituting sitagliptin, regarded as cardiovascular neutral, for placebo. They therefore obtain an estimate much closer to the incremental treatment effect clinicians might want when deciding whether to start tirzepatide.
That yields clinically interpretable absolute effects such as 1.4 fewer MACE events per 100 treated for one year, rather than simply a head-to-head relative comparison with another protective incretin drug.
2. The “trial-anchored real-world evidence” methodology
This may be the most important scientific novelty.
The investigators did not simply run another propensity-matched database analysis. They used a framework that they had previously tested by trying to reproduce the results of actual randomised trials from claims data. They had already emulated SURPASS-CVOT and SUSTAIN-6, allowing them to assess where their observational methodology did and did not reproduce known randomised results.
They also preregistered the protocol before the inferential analysis and explicitly structured it as a target-trial emulation.
So the broader methodological claim is:
Use RCTs to calibrate an observational analysis, and then extend that validated analysis to answer questions that the RCT did not answer.
That is potentially important for drug evaluation as placebo-controlled cardiovascular outcome trials become less practical or ethical.
3. The infection signal
The infection result is probably the most biologically novel finding.
The reduction was actually larger for serious infection than for the cardiovascular outcomes:
- infection hospitalisation HR ≈ 0.64
- infection-related mortality HR ≈ 0.40
- all-cause mortality HR ≈ 0.55.
The authors therefore raise the hypothesis that some apparent survival benefit of tirzepatide might come through infection susceptibility or host response, rather than exclusively through atherosclerosis.
That deserves further investigation, but I would regard it at present primarily as hypothesis-generating.
Critique
1. The fundamental problem is still confounding by indication
The biggest weakness is that tirzepatide and sitagliptin are prescribed to rather different patients.
Before weighting, tirzepatide recipients were substantially younger and generally healthier. Sitagliptin is particularly attractive in older, frailer patients because it is oral, familiar, relatively simple to use and does not induce substantial weight loss.
That creates exactly the kind of bias that could make tirzepatide appear to reduce:
- mortality
- severe infection
- hospitalisation
- perhaps even cardiovascular events.
Propensity weighting can balance measured variables, but it cannot balance characteristics that are absent or poorly captured in claims data.
The authors themselves acknowledge this, noting that earlier benchmarking suggested sitagliptin might preferentially be prescribed when clinicians anticipated shortened life expectancy.
This is particularly important because the extraordinary mortality and infection effects are precisely the outcomes most susceptible to such frailty confounding.
So I would have considerably more confidence in the MI finding than in the 55% reduction in mortality.
2. The infection result could itself be evidence of residual healthy-user bias
There are two interpretations of the infection result.
The authors’ interpretation is:
tirzepatide → altered biology/metabolism/inflammation → fewer serious infections → lower mortality.
But another interpretation is:
healthier, less frail patients → more likely to receive tirzepatide → less likely to develop serious infections or die.
That second pathway is entirely plausible.
The negative controls—lumbar radiculopathy and abdominal hernia—are useful, but they do not completely resolve this problem. An unmeasured frailty variable could strongly predict pneumonia, sepsis and mortality while having essentially no relationship with lumbar radiculopathy or hernia.
Thus a null negative-control analysis does not demonstrate absence of residual confounding.
3. Calling sitagliptin a “placebo proxy” risks overstating what the comparison means
Sitagliptin is cardiovascular-neutral in randomised trials, which makes the strategy reasonable. But it is still not placebo.
The treatment decision is very different:
- tirzepatide is injectable;
- sitagliptin is oral;
- tirzepatide produces major weight loss;
- sitagliptin generally does not;
- cost and insurance authorisation differ;
- patient motivation probably differs;
- clinicians may avoid major weight-loss drugs in frail or sarcopenic patients.
The authors explicitly acknowledge the oral-versus-subcutaneous difference as a limitation.
I would therefore describe the study as tirzepatide versus a cardiovascular-neutral active comparator, rather than psychologically treating the estimate as equivalent to tirzepatide versus randomised placebo.
4. Follow-up is surprisingly short
Although outcomes are expressed as one-year risks, actual median on-treatment follow-up was under half a year. The paper acknowledges this explicitly.
That matters for two reasons.
First, classical anti-atherosclerotic effects tend to accumulate over time. A cardiovascular benefit that emerges within the first few months is therefore less likely to represent regression or prevention of atherosclerotic lesions alone.
Second, extrapolating Kaplan-Meier estimates to one year when progressively fewer people remain under observation makes the NNT less robust than the headline presentation might suggest.
Thus NNT=70 “at one year” should not be interpreted with the same confidence as a one-year RCT in which almost everyone is actually observed for a year.
5. The very early divergence is both interesting and suspicious
The event curves reportedly begin to separate within approximately three months.
The authors suggest that this might mean mechanisms independent of weight loss—for example vascular or systemic inflammatory effects.
That is possible.
But in an observational study, an immediate treatment benefit is also a classic warning sign for:
- healthy-user bias,
- differences in frailty,
- treatment selection,
- differential healthcare behaviour.
The authors appropriately acknowledge exactly this ambiguity: early separation may represent biological effects or channeling of treatment to healthier individuals.
I would therefore resist using the rapid onset as strong mechanistic evidence.
6. The pattern across endpoints deserves more emphasis
If tirzepatide were producing a powerful general anti-atherosclerotic effect, one might expect MI and ischaemic stroke to move in broadly the same direction.
Instead:
MI: HR 0.67
Stroke: HR 0.91
The stroke confidence interval is wide, so this does not demonstrate biological heterogeneity, but it weakens a simplistic “32% less atherosclerotic disease” interpretation.
Moreover, when the endpoint is expanded to MI, stroke, revascularisation and unstable angina, the effect falls to only HR 0.92.
That contrast between:
- MACE HR 0.68,
- MI/stroke HR 0.74,
- expanded cardiovascular composite HR 0.92,
suggests that a surprisingly large part of the impressive headline result is associated with mortality and selected hard endpoints rather than a dramatic reduction across the cardiovascular spectrum.
7. All-cause mortality inside “MACE” inflates the cardiovascular interpretation
Their principal MACE definition is:
MI + stroke + all-cause mortality, rather than cardiovascular death.
That was chosen to mirror their target-trial emulation, but it creates an interpretational problem.
If tirzepatide reduces deaths due to infection—or if residual frailty confounding produces fewer infection deaths—those deaths improve the MACE endpoint despite having nothing to do with atherosclerosis.
This is important because the paper is titled “risk of atherosclerotic cardiovascular events.”
The cleaner cardiovascular result may therefore be the analysis excluding mortality:
MI or stroke HR 0.74, rather than MACE HR 0.68.
That distinction deserves more prominence.
8. Administrative claims cannot capture some of the most important confounders
The analysis adjusts for a large number of coded covariates, but claims data generally capture poorly or not at all:
- actual HbA1c
- blood pressure control
- measured physical frailty
- muscle mass
- diet
- exercise
- socioeconomic factors in detail
- adherence
- degree of obesity beyond coding
- unintentional versus intentional weight loss
- clinician judgement of prognosis.
The paper acknowledges residual confounding and some socioeconomic limitations.
In this particular drug comparison, those missing variables are not peripheral—they are precisely the factors likely to influence the choice between tirzepatide and sitagliptin.
Overall assessment
I would interpret the paper as good-quality observational evidence supporting a genuine cardiovascular benefit of tirzepatide, rather than as proof that tirzepatide reduces MACE by exactly 32%.
My confidence differs substantially by finding:
| Finding | My confidence |
|---|---|
| Tirzepatide provides cardiovascular benefit vs a neutral comparator | High-moderate |
| Reduction in MI | Moderate-high |
| Little/no effect on stroke in first year | Moderate, but underpowered |
| MACE HR exactly ~0.68 | Moderate |
| All-cause mortality reduced ~45% | Low-moderate |
| Infection mortality reduced ~60% because of tirzepatide biology | Low / hypothesis-generating |
| NNT ≈70 for one year | Useful estimate, but not equivalent to RCT-derived NNT |
The most important contribution of the paper may therefore not be the exact HR of 0.68. It is the demonstration of a potentially powerful strategy: benchmark an observational target-trial emulation against an actual RCT, then use the calibrated observational system to answer a clinically important counterfactual that the RCT cannot directly provide.
The most intriguing biological result is the infection signal, but paradoxically that is also the result I would treat with the greatest caution because infection and mortality are especially sensitive to unmeasured frailty and treatment-selection effects.
A particularly useful next analysis would be to ask whether the size and timing of the MI benefit can plausibly be explained by weight loss, glucose reduction, blood-pressure reduction, inflammatory effects, or a direct GLP-1/GIP vascular mechanism. The early separation of the curves makes that mechanistic question quite interesting.