Patients with epilepsy texted questions to their brain monitors in Penn pilot study

Thirteen adults admitted to the epilepsy monitoring unit (EMU) at the Hospital of the University of Pennsylvania spent several days texting their own brain monitors.

Through a smartphone app, they asked whether they'd had a seizure, how they slept and what their brain activity looked like, and an artificial intelligence (AI) agent with live access to their electroencephalogram (EEG) data answered, sometimes with charts.

Patients could also correct the system. When the app flagged a possible seizure, they marked it as real, a false alarm or uncertain, and researchers used those labels to retrain the seizure detector.

In a simulation run after the fact, false alarms fell 80% across patients' hospital stays, according to the study published Tuesday in Nature Medicine by researchers at Penn's Perelman School of Medicine and its Center for Neuroengineering and Therapeutics.

EEG devices record brain activity continuously, but the data generally reach expert reviewers on a daily or multiday basis, according to a news release from the National Institutes of Health (NIH), which funded the work.

None of the three devices the U.S. Food and Drug Administration has approved to treat epilepsy can share information with patients in real time, the authors wrote.

"These devices are made first and foremost for doctors, however, they do see a lot of information that patients would benefit from knowing. There are insights buried there about how their behavior impacts health, but they currently have no way of learning them in the moment," senior author Brian Litt, M.D., a professor of neurology, neurosurgery and bioengineering at Penn, said in the NIH release.

How the platform worked

The system streamed each patient's EEG from the hospital network into Penn's HIPAA-compliant cloud, where existing models scored the data for seizures, sleep stages and abnormal electrical spikes between seizures.

A version of OpenAI's GPT-4o-mini handled the conversation through a private endpoint contractually barred from training on or keeping patient data.

Patients and the system exchanged 1,307 messages, about 69% of them started by the system, including alerts and surveys sent three times a day.

When patients asked about their own data, the agent answered 35 of 38 questions correctly. Simple chat replies took about two seconds; data questions that required a database search took about nine.

Every message in both directions passed through an AI moderation check, and a member of the study team reviewed every exchange daily. Patients were told their care team wasn't monitoring the app and to press the unit's seizure alarm button for any suspected seizure.

Penn Medicine's AI Governance Committee, set up for this project and since made a standing body, reviewed how the AI interacted with patients.

Patient feedback and false alarms

The app sent 58 real-time seizure alerts. Patients responded to about two-thirds and rejected 22 as false alarms.

For safety, only the original detector ran during patients' stays. Researchers later replayed the data, retraining the model every three hours on patients' labels.

The median false alarm rate dropped from 0.74 an hour to 0.10, roughly the same gain researchers saw when they retrained on presses of the EMU's standard seizure alarm button.

The retrained model caught seizures no more often than the original. The remaining rate, roughly two to four false alarms a day, may still burden some patients and needs further work before real-world use, the authors wrote.

A second approach brought clinicians back in. Once a day, the system picked the 10 stretches of EEG, each 90 seconds long, it was least sure about and asked experts to label them.

In simulation, that cut false alarms 82.6%. Once trained and validated, a system like this may need only occasional expert review, the authors wrote, which would greatly reduce the resources needed to run it outside the hospital.

What patients said

Five of the 13 patients asked, unprompted, to keep using the platform after discharge.

"Completely unprompted, several patients asked if they could take the system home with them, which was one of the best signals we could receive," co-first author Zack Goldblum, a bioengineering doctoral student in Litt's lab, said in the NIH release. "We had to tell them, 'No, not yet, but we'll get there.'"

Five patients completed an exit survey. Three said the app made them more aware of links between their behavior, feelings and health, and one wrote that it helped "solidify what I felt to be true between some of my episodes and emotions." Another said it largely "reinforced what I already knew." The five rated the app's usability 83 out of 100 on a standard usability scale.

Patients were paid $25 a day. Several people who were approached declined because they weren't interested in research or were already in other studies, which the authors said may have tilted the group toward people interested in AI and technology.

The EMU is also an artificial setting where medication is reduced to provoke seizures for diagnosis, so the results show that patients will engage with such a system; whether it helps them manage epilepsy at home is the question a planned outpatient study is designed to answer, the authors wrote.

What's next

The team is preparing a study with Lauren Hammer, M.D., Ph.D., an assistant professor of neurology and bioengineering, that would use the platform to tune deep brain stimulation devices for Parkinson's disease.

The authors describe the current approach as monthly office visits after implant for expert programming, and wrote that access to the therapy is limited in part by access to the movement disorder specialists who program the devices.

In their proposal, the device would ask a patient whether their walking feels adequately treated, then adjust settings while the patient goes for a walk.

"Patients often understand aspects of their disease that sensors cannot measure directly," Litt said in a Penn Medicine news release. "Imagine an insulin pump that can answer questions about how a meal will affect blood sugar, or a Parkinson's device that adjusts treatment based on how a patient feels during a walk or writing a letter."

Outpatient studies using minimally invasive EEG and wearables will measure whether the platform reduces seizure frequency and severity, emergency department visits and health care use, according to Penn. The work is supported by a $5.6 million NIH Pioneer Award.

The authors expect regulatory approval for devices like this to come in steps, starting with low-level advisory systems.