The evidence points both ways. Pooled across ten randomized tirzepatide trials, atrial fibrillation came out at an odds ratio of 2.20 [1]. Its credible interval runs from 0.81 to 6.75, which measures nothing. Ten meta-analyses of this drug class point the other way, toward less atrial fibrillation [2]. None of them reached moderate confidence. A network comparison of 2,171,267 patients found no difference against a near-neutral drug [3].
This is the safety question, not the treatment one. Whether a GLP-1 keeps atrial fibrillation from returning after ablation is a separate matter. The answer there is also unsettled, and it is set out in the atrial fibrillation evidence.
What was pooled
Ten randomized controlled trials of tirzepatide in adults with overweight or obesity [1]. In total 6,515 people, of whom 4,491 — 68.9% — were randomized to the drug. The analysis was Bayesian, so its ranges are credible intervals rather than confidence intervals. It looked at three nested categories of rhythm problem.
For atrial fibrillation the pooled odds ratio was 2.20, credible interval 0.81 to 6.75. For atrial arrhythmia more broadly, 2.16, credible interval 0.90 to 5.87. For any arrhythmia at all, 1.84, credible interval 1.04 to 3.90.
How to read intervals that wide
A range from 0.81 to 6.75 is not a measurement. It is consistent with a slight reduction and with a sevenfold rise. It cannot tell those apart. Intervals get that wide when there are very few events. That is what arrhythmia looks like in trials designed to measure weight.
So the two narrower categories found nothing. The broadest one did, barely, at a lower bound of 1.04. One more trial in either direction could undo it. What an interval includes decides more here than the point estimate does, the same reading applied in the vessel-function measurements.
The literature that points the other way
Ten meta-analyses have asked whether this drug class lowers atrial fibrillation. An umbrella review appraised them together [2]. Two shared an identical six-study base and reported hazard ratios of 0.58 and 0.78. Identical evidence, different published answers.
Overlap between the reviews was substantial: 43.8% for recurrence and 34.1% among the semaglutide-only incident reviews. So when these papers agree, the agreement is largely a shared base rather than independent replication. Five were rated low confidence and four critically low. None reached moderate or high. The full appraisal is in the umbrella review.
The largest single comparison sits between the two positions. A network and pairwise meta-analysis covered 2,171,267 patients with type 2 diabetes [3]. It found GLP-1 drugs no different from DPP-4 inhibitors on new-onset atrial fibrillation, at RR 0.93. SGLT2 inhibitors beat both. That is not a signal of harm and it is not a benefit either.
What would make this more than a signal
A trial with arrhythmia as a prespecified endpoint. Monitoring designed to catch it. Enough events to produce an interval narrower than a factor of eight. None of that exists. What exists is adverse-event reporting pooled after the fact.
There is also a confounder running in an awkward direction. Rapid weight loss, electrolyte shifts and a rising heart rate all accompany these drugs. All three have some relationship to rhythm. A pooled adverse-event count cannot say whether any of it is doing the work. Nor can a trial population answer questions about a clinic population, the gap measured in the ordinary-care results.
What to do with it
Not stop a prescribed drug on the strength of it. Three wide intervals from pooled adverse events are not a reason to change treatment. That decision belongs to a prescriber who knows your heart history.
What it is worth is a question. A history of atrial fibrillation, palpitations or any rhythm diagnosis belongs on an intake form before starting. It is exactly the kind of history a telehealth questionnaire may not ask for. The census of what sellers leave unasked counts that from the other side.