Two safety signals, one way to read them
Two headlines, two weeks apart, about the same class of medicines. One said GLP-1 drugs had been linked to a potentially deadly brain disorder. The other said tirzepatide was not linked to atrial fibrillation. Both are technically true. Both mislead, in opposite directions, and for the same reason: they report a conclusion and drop the number that would tell you how much to believe it. This post is about that number. It is not a verdict on either drug. It is a way of reading the next headline of this kind, because there will be one, and the class is now prescribed to tens of millions of people.
Start with the brain disorder. The study behind it, published in Clinical Nutrition in February and presented at a conference this year, searched the FDA's adverse-event reporting database and found fifteen cases of Wernicke encephalopathy, a thiamine-deficiency condition, among 195,979 people whose reports mentioned a GLP-1 receptor agonist. The reporting odds ratio was 2.35, with an interval from 1.38 to 4.01. Here is what that database is: doctors, patients and companies sending in reports of things that happened to people on a drug. Nobody knows how many people took the drug and had no problem, so no rate can be computed. What the number says is that Wernicke appears in these reports more often than you would expect by chance. What it cannot say is whether the drug caused it. Most of the fifteen had severe vomiting or malnutrition first, which is how thiamine runs out regardless of the drug. The honest reading is a reason to look, in patients with severe gastrointestinal symptoms. That is exactly how the authors put it.
Now the reassurance. The study behind it, in the Journal of the American Heart Association in September, pooled ten randomised trials, 6,515 adults with overweight or obesity, 4,491 of them on tirzepatide, and asked whether atrial fibrillation was more common on the drug. The pooled odds ratio was 2.20, with an interval from 0.81 to 6.75. Read that interval before the headline. It runs from a 19 percent lower risk to a nearly seven times higher risk, and it includes one, which is why the result is called not associated. But an interval that wide does not say there is no risk. It says these trials, built to measure weight loss, saw too few of these events to tell. The same paper found higher odds of any arrhythmia, 1.84 with an interval from 1.04 to 3.90, which clears one by a hair on a handful of events. The authors say so, plainly, and ask for caution. Both findings are real. Neither is the sentence the headline made of it.
The two studies fail in mirror image, and that is worth seeing clearly. A trial is designed to find a benefit in thousands of people over a year or two; it is the wrong instrument for a harm that happens to one person in ten thousand, and it will usually report no association for the same reason a bathroom scale reports no weight for a letter. A reporting database is designed to catch exactly those rare harms, and it will find them, but it cannot tell you how often they happen or whether the drug is the cause. So the alarming headline comes from the tool that cannot measure frequency, and the reassuring headline comes from the tool that cannot see rare events. Neither is wrong. Both are incomplete in a predictable direction, and once you know the direction, you can correct for it.
For an investor the practical question is what actually changes a label, because that is what changes prescribing and revenue. It is not a meta-analysis and it is not a case series in a journal. It is a regulator, the FDA or the EMA, reviewing adverse-event data and deciding whether the label needs a new warning, a new contraindication or a new monitoring requirement. That process takes years, and the great majority of signals never reach it. The ones that do tend to have three properties: a plausible mechanism, a consistent pattern across different data sources, and a rate high enough to matter at the scale of use. Wernicke has the first and part of the second. The arrhythmia finding has none of the three yet. What to watch is not the next journal press release but the regulators' own safety communications, which are dated, sourced and public, and which we will record here when they appear.
We track this because the questions being asked about these medicines are changing. Through 2024 the question was whether they worked, and the answer was yes, at scale. The question now is what they do over years in tens of millions of people, and the honest answer is that nobody knows yet, and each of these studies is a small, inconclusive contribution to finding out. Reporting them with their intervals and their limits is slower than the headline and less comforting than the reassurance, but it is the only version that will still be true when the next one arrives. Disclosure: the author holds shares in Eli Lilly, which markets tirzepatide, and in Novo Nordisk, which markets semaglutide; the full statement is on the About page. This is general information, not medical advice.