// Contents
Over a few days in hospital, 13 epilepsy patients received seizure alerts on their phones and answered each one with a yes or a no. They responded to more than two thirds of them. The algorithm was retrained on those short answers, and in a simulation its false alarms fell by about 80%. Labels from experienced specialists added only another 2%. But there is a catch: the new version caught real seizures no better than the old one.
All the participants were staying in the epilepsy monitoring unit at the Hospital of the University of Pennsylvania, and each had a phone app with an AI assistant wired to their own EEG. The work by neurologist Brian Litt’s group appeared in Nature Medicine on October 13, 2026. The authors call it a proof of concept: the sample is tiny, and everything happened inside the hospital. Bioengineering doctoral students Zack Goldblum and Haoer Shi from Litt’s lab built the platform.
The ward where everyone waits for a seizure
People are admitted to a unit like this to have a seizure under supervision, not to get treated. A patient spends several days on video with electrodes on their head so that doctors can catch a seizure on the recording and see where it starts. EEG, or electroencephalography, records the brain’s electrical activity, and during a seizure the rhythm on the trace changes in a way an experienced doctor spots at once. Eleven participants had electrodes on the scalp and two had them inside the brain, according to the description of the open dataset .
A detector program watches the recording. It often raises the alarm for nothing, for example when someone scratches their head or turns over in their sleep and the movement leaves noise on the trace. Call a nurse for every one of those and by evening nobody will believe the alarms anymore.
Normally doctors and engineers tune such a detector to a specific person, and that takes months of in-person programming. Litt’s team suggested asking the one who knows for sure whether a seizure happened. The stakes are higher than they look: the WHO estimates that up to 70% of people with epilepsy could live seizure-free if properly diagnosed and treated. An accurate record of seizures is where that diagnosis begins.
Three buttons instead of months of tuning
The main finding hides in the humblest action. Whenever the detector flagged a possible seizure, the phone asked: did this really happen? The patient could answer yes, no or not sure. Patients responded to more than two thirds of those questions, reports MedicalXpress .
Those answers became hints for the algorithm, much like the captions under photos that teach a neural network to recognise cats. Once retrained on them, the detector produced about 80% fewer unnecessary alarms over the hospital stay. That figure comes from a simulation: the team ran the recordings back through the retrained model after the fact and counted how many alarms it would have raised. So how the system behaves live is still an open question.
The new version became calmer, not sharper. It stopped making noise for nothing, yet it missed just as many real seizures as before.
What people ask their own brain
The app works both ways. A patient types a question, and the AI agent answers from the live EEG recording, sometimes with a chart. Behind the conversation was a version of OpenAI’s GPT-4o-mini, reached through a private channel that is contractually barred from keeping patient data or training on it. Every message in either direction went through an automatic safety check, writes Medical Economics .
Most conversations were about seizures, sleep, mood and energy. A typical question: did I have a seizure last night? The system itself did most of the talking, about 69% of the messages: alerts, surveys three times a day and replies. A simple answer came back in about 2 seconds, faster than a nurse could walk to the room. A question that sent the system digging through the EEG record, such as “how much did I sleep?”, took about 9 seconds. Five of the thirteen asked on their own whether they could take the system home. The team’s answer was not yet, but that is where the work is heading, Goldblum recalls in an interview with Medical Economics. Only five participants filled in the usability questionnaire, and on average they gave the app 83 points out of 100.
// The trial in numbers
Why medical devices stay silent
Litt puts the idea simply: patients often understand aspects of their disease that sensors cannot measure directly. Those are his words in the University of Pennsylvania announcement . A person knows they slept badly before a seizure or skipped a pill. The device sees only a line on a trace.
The channel in the other direction barely exists either. In the US, three implanted devices are approved for epilepsy: vagus nerve stimulation, responsive neurostimulation (RNS) and deep brain stimulation. None of them can share information with patients in real time, the authors write. In other words, a person can learn something from their own device only through a doctor. Litt says so directly: the data hide many clues about how behaviour affects health, but patients currently have no way of learning them in the moment, as NIH quotes him .
How the system is built
The system has three parts. Software continuously captures the signal from the EEG electrodes. A secure cloud environment that meets US medical privacy rules analyses the recording in near real time. The third part is a smartphone app through which the patient messages the AI agent, which answers questions and sends notifications of its own. Recordings from the participants are openly available to other researchers. The work is funded by a $5.6 million NIH Pioneer Award, and an early version appeared in January 2026 as a preprint on medRxiv under an even bolder title: “An Implantable Device that Converses with Patients and Learns to Co-Manage Epilepsy”.
What the trial has not shown yet
About 50 million people worldwide have epilepsy, according to the WHO . So it is tempting to read this work as news about treatment. It is about something else. Thirteen patients spent a few days in one hospital under staff supervision, and the trial was not designed to show whether people would have fewer seizures.
That is for the next, outpatient trials with minimally invasive EEG and wearable sensors to test. They will count what matters in everyday life: how often seizures happen and how severe they are, how many ambulance calls and hospital visits people need. After that, the authors want to use this kind of dialogue to tune deep brain stimulation for Parkinson’s disease automatically. They believe it could also help people with pacemakers and insulin pumps. They expect regulators to approve such systems in stages, starting with ones that only give advice and decide nothing on their own.
Devices like these spend years collecting data about the body in silence. In this trial, a device asked its owner what was going on, and the owner’s answers made it more accurate.
