Ten meta-analyses agreeing with each other looks like a settled question. This umbrella review went and checked what they were made of, and the answer is that they are largely made of the same trials — which is the difference between counting studies and counting independent observations.
The pair that settles it
Two of the three reviews of post-ablation atrial fibrillation recurrence included exactly the same six primary studies — a pairwise corrected covered area of 100%. [1] One reported a pooled hazard ratio of 0.58, 95% CI 0.42 to 0.79. The other reported 0.78, 95% CI 0.61 to 0.99.
Identical evidence, different published answers. Nothing separates them except what their authors chose: which comparator, which statistical model, which population. A reader who found one of those papers would come away with a meaningfully different impression from a reader who found the other — which is what a conclusion detached from its data looks like when it happens twice from one dataset.
How much of the literature is one literature
Corrected covered area measures how much the primary studies inside a set of reviews overlap. It came to 43.8% across the three recurrence reviews, 34.1% across the five semaglutide-only reviews of incident AF, and 22.5% across all six incident-AF reviews.
So when these papers agree, that agreement is substantially a shared base rather than independent replication, which is the authors’ own conclusion. Ten citations pointing the same way is not ten pieces of evidence — the same illusion as a pooled figure standing in for consensus.
The quality appraisal
Every review was independently appraised twice on a standard instrument. Five came out at low confidence and four at critically low. None reached moderate or high.
That is a striking result for a question with ten published syntheses. Quantity of reviews has no relationship to the certainty they can deliver, and the effect estimates still spread from an odds ratio of 0.54, 95% CI 0.39 to 0.77, up to 0.83, 95% CI 0.70 to 0.98 — intervals that overlap but describe quite different sized benefits, as the underlying trial-level evidence also shows.
What this does not say
That the effect is not real. The authors’ conclusion is that GLP-1 drugs are plausibly associated with lower atrial fibrillation burden, with certainty low for new cases and very low for recurrence after ablation.
What they ask for is adequately powered randomized trials with AF endpoints that are adjudicated and systematically monitored — because a rhythm problem that is often silent is only counted when somebody looks for it, and how hard a study looks is another choice that moves the number.