Seizures When You Already Have Epilepsy

If you already have epilepsy and are offered a GLP-1 for your diabetes, the question is whether it affects your seizures. A study of 8,688 matched pairs reports fewer — a hazard ratio of 0.82.^[1] It also reports 60% lower mortality from the same comparison, and that second number is the reason to be careful with the first.

What was done

Using electronic health records from a large multinational network, researchers identified adults with at least three epilepsy or recurrent-seizure diagnoses who also had type 2 diabetes. Those starting a GLP-1 were matched one-to-one against those starting a different glucose-lowering drug — an SGLT2 inhibitor, a DPP-4 inhibitor, a sulfonylurea or insulin — using propensity scores built from 82 covariates, producing 8,688 matched pairs.

That is a serious piece of methodology. Eighty-two covariates is thorough, the sample is large, and the follow-up runs well over a year in both arms.

Why the mortality figure is the tell

The mechanism is well understood and has a name, and we set it out at length in when a hazard ratio is too good, where the same shape appears in a study of psoriasis patients. In short: GLP-1 medicines go to people who are healthier, better insured and more engaged with their care, while the sickest and frailest disproportionately end up on insulin — which sits in the comparison group.

Propensity matching corrects for what is written in the record: diagnoses, medications, laboratory values, demographics. It cannot correct for what is not recorded, and general frailty — the thing that best predicts who dies within two years — is largely not recorded. Eighty-two covariates matched on paper can still leave two groups of visibly different people.

When the effect on death exceeds what randomized trials found, the extra is measuring who got the prescription.

What that does to the finding you came for

The study’s actual subject is seizures, and its seizure result is modest and biologically plausible: a hazard ratio of 0.82, an absolute risk difference of 2.1 percentage points. Taken alone, that is exactly the sort of finding worth taking seriously.

It cannot be taken alone. It comes from the same comparison between the same two groups whose mortality difference demonstrates those groups were not alike. If unmeasured health explains most of a 60% mortality gap, it is also available to explain a 2.1-point seizure gap — and healthier people with better-controlled diabetes have fewer seizures for reasons that have nothing to do with which drug they were given.

That is the part worth carrying: an implausible result does not sit quietly in its own row. It is evidence about the whole comparison, and it discredits the plausible findings beside it.

A separate question — whether these drugs affect the chance of developing epilepsy in the first place, rather than seizure recurrence in people who already have it — is examined in GLP-1s and epilepsy risk, and the pattern inside those numbers points the same way.

One sentence we cannot reconcile

To the authors' credit

They describe their own findings as hypothesis-generating and say they warrant confirmation in prospective studies before being translated into practice. That is the correct characterization and it is not a formality — plenty of papers with weaker designs claim more.

What this analysis genuinely establishes is that the question is worth asking. There are mechanistic reasons to wonder whether these drugs affect seizure threshold, and a signal in 8,688 matched pairs is a reasonable basis for designing a trial. It is not a basis for prescribing.