An AI Epilepsy Monitor Learned From Patients. False Alarms Fell 77.8% in a Replay

Thirteen adults with epilepsy spent a few days in a Philadelphia hospital texting with software that could read their brain waves. A University of Pennsylvania team streamed each patient's EEG to a secure cloud, ran seizure and sleep detectors on it, and placed a large language model between those numbers and a smartphone app. The result published in Nature Medicine on October 13, 2026: feedback from patients alone retrained the seizure detector and cut its false alarms by a median of 77.8%. The fine print is that the retraining was tested in a replay of the data, not live.

What the patients actually used

The study ran in an epilepsy monitoring unit, the ward where patients are wired to EEG, with medication reduced, so doctors can record seizures for a surgical evaluation. Eleven patients had scalp EEG and two had electrodes inside the skull. Median participation was 4.87 days.

The app sent seizure and spike alerts, three surveys a day and a morning sleep report. Patients could also ask about their own data, and a "data agent" wrote database queries to answer. The model behind it was a HIPAA-compliant instance of GPT-4o-mini, chosen, the authors write, because at the start of the study it was the only model in the hospital's compliant cloud tenancy with reliable enough instruction following and tool use.

Over the study, 1,307 messages went back and forth. The system started 900 of them and patients started 407, mostly ordinary conversation. When patients asked about their data, the data agent answered correctly in 35 of 38 checked cases (92.1%). A moderation layer blocked 6 patient messages and none of the 900 AI-generated ones, and investigators reviewed every conversation at least daily.

The false alarm number, read closely

The starting point was not impressive. The pretrained detectors produced a median of 0.47 false alarms per hour. Of 59 seizures confirmed by clinicians in the 11 patients who had the annotation feature, the app's alerts flagged 8, and only 8 of its 58 alerts were real seizures.

Patients answered 67.2% of those alerts, labeling 9 as real seizures, 22 as false alarms and 8 as uncertain. The question was whether labels like these could retrain the detector every three hours with no expert involved.

Two label sources were tested. With the hospital's standard push-button alarms, available for all 13 patients, retraining cut false alarms by a median of 77.8%. Replayed across the whole stay, the rate fell from 0.469 to 0.097 per hour while median sensitivity stayed at 1.00. With the labels patients entered in the app, available for only 6 patients, false alarms fell by 80.9%.

The detector learned from patients, but during the real sessions patients only ever saw alerts from the unchanged baseline model.

That was a deliberate safety choice, and it shapes how far the result goes. The analysis was pseudoprospective: retrained models were run on later data after the fact, mirroring live use without being live. Sensitivity did not improve. The authors also note that the false alarms left after retraining, 2 to 4 per day, may still burden some patients, and that the lightweight detectors they used are not state of the art.

How big is this, really

Thirteen people over a few days is a feasibility study, and the authors frame it that way: prospective, single center, observational, no control group, and a sample set by feasibility rather than a power calculation. Usability scores were high, a mean of 83 out of 100 on the System Usability Scale, but only 5 patients completed that survey. Five of the 13 asked for access after discharge. Participants were paid $25 per day, and the authors flag a possible tilt toward people who like technology.

Nothing here measures what matters most to patients: fewer seizures, fewer emergency visits, better days. The authors say showing that will require those outcomes and that the burden of evidence is high. Their next steps are an outpatient deployment and a trial that uses the same platform to program deep brain stimulation devices in Parkinson's disease.

One disclosure is worth knowing. Three authors are named inventors on a pending provisional patent for the system, filed by the University of Pennsylvania. The work was funded by NIH awards and private foundations.

A device that asks "was that a seizure?" is a clever way to collect labels. Will people still answer it at home, after weeks of false alarms, with no nurse down the hall?