Patients' texts improved an AI seizure detector in replay tests (13-person pilot)

Thirteen adults were admitted to an epilepsy monitoring unit, the ward where doctors record brain activity to catch seizures. For several days they texted the system that was watching them, and their replies went straight back into the software that detects the seizures. The study was published in Nature Medicine on 16 October 2026.

The loop works in steps. Software reads the electroencephalogram (EEG), the recording of electrical activity in the brain, and flags possible seizures on its own. Each flag was followed by a question: the patient marked it as real, as a false alarm or as uncertain, and the team used those labels to retrain the detector. Patients responded to more than 67% of real-time detections. The system started about 69% of the 1,307 messages exchanged, including alerts and surveys sent three times a day. Every message in both directions passed an AI moderation check, and a member of the study team reviewed every exchange each day.

When they asked about their own data, it answered 35 of 38 correctly. NIH's summary of the work puts the rate of accurate and relevant answers about brain activity above 92%, and says the detector, fine-tuned on patient feedback, improved its ability to pinpoint seizures. That gain was measured in a replayed evaluation, not in a rollout. The study ran at one hospital with 13 people, who were already monitored for clinical care, and standard treatment stayed unchanged. Coverage of the paper stresses that the result is not evidence that patients had fewer seizures or safer care.

For people with epilepsy, the change to watch is a detector that learns from the judgment of the person it monitors. If larger studies confirm the gain, the next setting would be outpatient care, where patients are not in a hospital bed. The researchers are planning those studies, but their timing is not yet known. Nature Medicine's own summary points to the harder problem ahead: building AI that responds appropriately, not only a detector that improves.

The pilot's location also needs a clarification. Medical Economics places the work at the Hospital of the University of Pennsylvania, while the paper's authors are listed at the Texas Institute for Restorative Neurotechnologies and McGovern Medical School. The accounts reaching us do not explain the difference, and it does not change the study's size or its limits.