Does SOUL show oral semaglutide is cardiovascularly equivalent to the injectable?

SOUL tested oral semaglutide 14 mg daily in people with type 2 diabetes and either atherosclerotic cardiovascular disease or chronic kidney disease, and reported a reduction in major adverse cardiovascular events. I have seen this used to argue that the oral formulation is interchangeable with injectable semaglutide for cardiovascular protection, and separately to argue the opposite - that its hazard ratio of about 0.86 is visibly weaker than SELECT's 0.80 and SUSTAIN-6's 0.74, so the oral route delivers less.

Both arguments look like the same statistical mistake to me, made in opposite directions: comparing point estimates from trials with different populations, different doses and different comparators as though the difference between them were itself measured. But I cannot articulate why cleanly, and I would like to be able to.

Also, if a class effect is established for injectable semaglutide, what work is a trial like SOUL actually doing? Is it confirming route-independence, or is it a regulatory requirement for a formulation, or something else?

Both arguments are the same error, and the clean way to say why is: no trial in this set measured a difference between the trials. Comparing 0.86 against 0.80 treats the gap as an estimate when nothing estimated it, and the uncertainty on that gap - if you constructed it - would be far wider than the gap itself.

What SOUL established

Roughly 9650 adults with type 2 diabetes and established atherosclerotic cardiovascular disease and/or chronic kidney disease, randomised to oral semaglutide 14 mg daily or placebo on top of standard care, median follow-up around four years. Three-point MACE: about 12.0% versus 13.8%, hazard ratio approximately 0.86, 95% CI 0.77 to 0.96 [1].

Absolute arithmetic: ARR = 13.8 - 12.0 = 1.8 percentage points, so NNT = 1 / 0.018 = 55.6, call it 56 treated over roughly four years to prevent one event.

Why you cannot rank the three hazard ratios

Four differences, each large enough to move a hazard ratio on its own:

• Population. SELECT excluded diabetes by design. SUSTAIN-6 and SOUL both enrolled type 2 diabetes, but SOUL additionally admitted chronic kidney disease without atherosclerotic disease, which changes the event mix.
• Background therapy era. SUSTAIN-6 randomised in 2013 to 2015; SOUL a decade later. Background statin intensity, antiplatelet strategy and SGLT2 inhibitor use are all substantially different, and a better-treated placebo arm compresses the achievable relative benefit. This alone could account for the whole 0.74-to-0.86 gap.
• Duration and event accrual. SUSTAIN-6 ran 104 weeks as a non-inferiority safety trial. SOUL ran roughly twice as long with a much larger event count and therefore a much tighter interval.
• Exposure. Oral semaglutide 14 mg daily does not produce the same systemic exposure as 2.4 mg weekly injected, and the two are not interconvertible by any simple ratio.

Look at the intervals rather than the points: roughly 0.72 to 0.90, 0.58 to 0.95, and 0.77 to 0.96. They overlap almost entirely. Three studies whose compatible-value ranges overlap that heavily are not distinguishable on effect size, and the correct summary is one estimate, not a ranking: across populations, doses and routes, semaglutide reduces MACE by something in the region of 15 to 25% relative.

What work SOUL is doing

Three things, none of which is "confirming route-independence" in the strict sense:

• It removes a real possibility rather than confirming an assumption. Oral semaglutide has around 1% bioavailability with wide between- and within-person variability driven by food, water volume and gastric conditions. It was not obvious a priori that daily oral dosing produces a steady enough exposure to deliver an outcome benefit. A null SOUL would have been interpretable and would have mattered.
• It supports a formulation-specific label claim. Regulators generally do not extrapolate cardiovascular outcome claims across formulations with materially different pharmacokinetics.
• It extends the population. SOUL's inclusion of chronic kidney disease alongside atherosclerotic disease widens the evidence base beyond either predecessor.

What it does not establish

It does not establish that oral 14 mg and injectable 2.4 mg are interchangeable, because nothing randomised anyone between them for cardiovascular outcomes. It does not establish equivalence of any kind - a trial against placebo cannot support an equivalence claim about a different arm that did not exist. And "similar hazard ratio therefore similar drug effect" is the same inferential error as "different hazard ratio therefore different drug effect", just more comfortable.

Adding the formal reason the point-estimate comparison fails, in case it helps to see the arithmetic.

If you genuinely wanted to compare two hazard ratios from separate trials, you would do it on the log scale. The variance of the difference of two independent log hazard ratios is the sum of their variances, so the standard error of the difference is the square root of that sum. Because you are adding variances, the interval on the difference is always wider than either original interval - by a factor approaching the square root of two when the two trials are similarly precise.

Concretely, with SELECT at log(0.80) and SOUL at log(0.86), the difference in log hazard ratios is small - about 0.07 - while the combined standard error is comfortably large enough that zero difference sits well inside any sensible interval. So the data are entirely compatible with the two hazard ratios being identical. You do not need the exact numbers to see this; you only need to notice that the difference between the point estimates is much smaller than the width of either interval.

The general rule worth internalising: if two confidence intervals overlap substantially, no formal test is going to find a difference. The converse is not true - non-overlapping intervals do not guarantee a significant difference either - but the overlapping case is a reliable stop signal.

This is also why network meta-analysis exists, and why its output is always accompanied by a warning about the transitivity assumption. An indirect comparison anchored on placebo arms assumes those placebo arms are exchangeable. Given the decade between SUSTAIN-6 and SOUL, and the change in background cardiovascular therapy over that decade, that assumption is doing a lot of unearned work here.

Practical footnote on what "oral semaglutide 14 mg" means, because it affects how you should read SOUL's exposure.

The 14 mg tablet is not a 14 mg systemic dose. Bioavailability is on the order of 1%, so the delivered quantity is closer to a tenth of a milligram and depends heavily on whether the dosing conditions were met: fasting, a small volume of plain water in the region of 120 mL, and a wait of at least half an hour before food, drink or other oral medication [1].

In a clinical trial with regular contact and reinforcement, adherence to those conditions is far better than it will be over years of ordinary life. So SOUL's result reflects trial-grade dosing technique as well as trial-grade adherence to the tablet itself. Whatever the effect size is in routine use, the mechanism by which it would be smaller is not pharmacological - it is that a meaningful fraction of doses deliver less drug than intended.

This is one of the few places where a formulation difference has a plausible route to a real difference in outcomes, and it is a reason to be careful with any claim of interchangeability rather than a reason to disbelieve SOUL.