Adults admitted to a specialized epilepsy unit were able to chat with their own brain monitors using a smartphone application powered by artificial intelligence, according to a study published Tuesday in the journal Nature Medicine.
The pilot trial, conducted at the Hospital of the University of Pennsylvania, included 13 patients who texted questions about their brain activity, sleep quality and whether they had experienced a seizure. In response, an AI agent connected to their live electroencephalogram, or EEG, feeds provided real-time answers and visual charts, Medical Economics reported.
Patient feedback cuts false alarms
Standard hospital EEG devices record electrical brain activity continuously, but the National Institutes of Health noted that data typically reaches doctors only on a daily or multiday schedule. Furthermore, the researchers pointed out that none of the three epilepsy devices approved by the U.S. Food and Drug Administration can deliver real-time data directly to patients.
During their hospital stays, participants and the AI system traded 1,307 messages, with the system initiating nearly 70 percent of interactions through scheduled surveys and alerts. When patients posed direct questions about their data, the AI answered 35 out of 38 queries correctly, Medical Economics reported. Simple responses arrived in about two seconds, while complex inquiries requiring database queries took roughly nine seconds.
The application sent 58 real-time seizure warnings to participants, who responded to two-thirds of them and dismissed 22 as false alarms. For clinical safety, the live system relied solely on the original detection model during patient stays. However, researchers later replayed the data in computer simulations, retraining the algorithms every three hours using patient responses. That retraining drove down median false alarm rates by 80 percent, from 0.74 per hour to 0.10 per hour.
A separate simulation tested a hybrid approach in which clinicians reviewed the 10 most ambiguous 90-second EEG segments once a day. That method reduced false alarms by 82.6 percent, indicating that future systems might operate outside medical facilities with only minimal expert supervision.
Cloud architecture and privacy safeguards
To run the platform, EEG signals were streamed from bedside monitors into a secure, HIPAA-compliant cloud environment. The conversational tool relied on OpenAI’s GPT-4o-mini through a private setup contractually barred from retaining or training on participant data.
Safety oversight was handled by the Penn Medicine AI Governance Committee, which required continuous automated moderation of messages alongside daily manual reviews by study personnel. Patients were clearly instructed that hospital clinicians were not monitoring the chat tool and that they must use the standard hospital alarm button if they suspected a seizure occurred.
Participants, who received $25 daily for taking part, showed high enthusiasm. Five of the 13 patients asked on their own whether they could continue using the software after leaving the hospital.
“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 said in an NIH release.
Future trials beyond the hospital
Researchers cautioned that the hospital setting differs from daily life because anti-seizure medications are deliberately lowered during monitoring to prompt episodes for diagnosis. Even after simulated retraining, the system still generated two to four false alarms per day, a rate authors noted could burden users at home.
The team is now preparing to adapt the conversational platform for other neurological conditions. Working with Dr. Lauren Hammer, researchers plan to evaluate the system for fine-tuning deep brain stimulation hardware used to treat Parkinson’s disease. Rather than requiring frequent clinic visits for manual adjustments, the platform could ask patients about their mobility while they walk and automatically adjust device settings.
Supported by a $5.6 million NIH Pioneer Award, upcoming outpatient studies will use wearable and minimally invasive sensors to evaluate whether real-time feedback can reduce emergency room visits and seizure frequency.
