The National Institutes of Health (NIH) announced that an NIH-funded clinical study found a conversational artificial intelligence interface can let epilepsy patients communicate directly with their monitoring devices in real time, improving seizure detection, according to the official release. The study, conducted at the University of Pennsylvania and published in Nature Medicine on October 13, 2026, reported that patient feedback helped reduce the system's false seizure alarms by about 80%.

Key takeaways

  • A small NIH-funded clinical study at the University of Pennsylvania tested a conversational AI (SI) interface letting epilepsy patients query real-time EEG data via smartphone.
  • The system answered more than 92% of patient questions about their brain activity accurately, while patients responded to over 67% of real-time seizure alerts.
  • Patient feedback helped the system cut its false alarm rate by about 80%, with expert feedback adding a further 2% reduction.
  • Researchers, led by corresponding author Brian Litt, M.D., plan larger outpatient studies and a parallel platform for Parkinson's disease deep brain stimulation devices.

What the Study Measured

The research, funded by NIH's National Institute of Neurological Disorders and Stroke (NINDS), tested a smartphone-based super intelligence (SI) platform with 13 patients staying in the epilepsy monitoring unit at the Hospital of the University of Pennsylvania over several days. Participants received alerts, surveys, and notifications tied to their electroencephalography (EEG) data, and could ask the conversational SI direct questions, such as whether a seizure occurred while they slept.

Across the study period, patients and the SI exchanged more than 1,300 messages. The system responded to over 92% of patient questions about their brain activity with accurate and relevant information, according to NIH. Patients, in turn, responded to more than 67% of real-time seizure detections, marking each alert as true, false, or uncertain.

That two-way exchange is the core finding: patients could extract usable information from a continuous stream of EEG data that, under current practice, is typically reviewed by experts only once a day or every few days, per the release.

The Numbers on Alarm Reduction

The headline figure in the release is the drop in false alarms. NIH states that the seizure-detecting model fine-tuned itself using patient feedback, which improved its overall ability to pinpoint seizures and reduced its false alarm rate by 80% across the length of the study. Allowing the SI to also incorporate feedback from expert reviewers produced a further 2% decrease in alarms.

MeasureFigure reportedComparisonChange
Patients enrolled13Not statedNot stated
Total messages exchangedMore than 1,300Not statedNot stated
Accurate answers to patient questionsOver 92% of questionsNot statedNot stated
Patient responses to seizure alertsMore than 67% of detectionsNot statedNot stated
False alarm rate reduction (patient feedback)Reduced by about 80%Baseline before feedback-80%
Additional reduction from expert feedbackFurther 2% decreaseAfter patient-feedback adjustment-2%

NIH did not provide a numeric baseline false-alarm rate before the intervention, only the percentage reduction achieved through the feedback loop.

Why the Alarm Rate Matters for Device Design

According to the release, EEG devices used to diagnose epilepsy place electrodes on or beneath the scalp to record brain activity signaling an active seizure or elevated seizure risk. These devices continuously capture data, but patients typically see only a fraction of it because information is shared with expert reviewers on a daily or multi-day basis, NIH said.

Brian Litt, M.D., a professor of neurology and bioengineering at the University of Pennsylvania and the study's corresponding author, said the devices are built first for doctors even though they contain insights patients could use in the moment. At the same time, onboard algorithms that flag suspicious signals frequently mislabel benign activity, requiring experts to manually disregard false positives, according to the release. The study's co-first authors, graduate researchers Zack Goldblum and Haoer Shi, built the SI interface specifically to let non-expert patients supply the real-time correction that algorithms were missing.

Amy Bany Adams, Ph.D., acting director of NINDS, described the work as a proof-of-concept with potential to shift epilepsy management, and possibly management of other diseases, toward a more adaptive, patient-involved model, per the release.

How NIH Funds This Kind of Research

NINDS, part of NIH, describes itself as the nation's leading funder of research on the brain and nervous system, with a mission to build fundamental knowledge of neurological disease and apply it to reduce disease burden. NIH, the federal government's medical research agency within the Department of Health and Human Services, comprises 27 Institutes and Centers and funds basic, clinical, and translational research nationwide, including this project through seven separate NINDS grants listed in the release.

This study illustrates the standard NIH funding model: a university research team, in this case based at the University of Pennsylvania, receives federal grant support to run a clinical proof-of-concept study, with results published in a peer-reviewed journal, here Nature Medicine, before scaling to larger trials.

What Comes Next

NIH said the research team is now planning larger-scale studies for outpatient use of the interface, moving beyond the hospital monitoring-unit setting used in this proof-of-concept. The same group is simultaneously preparing to deploy a parallel platform enabling communication between deep brain stimulation devices and patients with Parkinson's disease.

The release notes patient enthusiasm during the study: several participants asked, unprompted, whether they could take the system home with them. Goldblum said researchers told them, "No, not yet, but we'll get there," which the team took as a strong signal for future recruitment, according to NIH.

Frequently asked questions

How many patients were involved in the study?

The clinical study tested the interface with 13 patients who stayed in the epilepsy monitoring unit at the Hospital of the University of Pennsylvania over several days, according to NIH.

What exactly improved by 80%?

NIH reported that the system's false seizure alarm rate was reduced by about 80% after the model fine-tuned itself using real-time patient feedback on whether detected events were true, false, or uncertain.

Is this technology available to patients now?

No. NIH describes this as a proof-of-concept study, and researchers are now planning larger outpatient trials; several participants asked to take the system home but were told it was not yet available.

Will this technology be used for other conditions?

The research team is preparing to deploy a similar platform to enable communication between deep brain stimulation devices and patients with Parkinson's disease, per the release.