Patient feedback improves AI seizure detection in simulation, small study finds

An AI epilepsy monitor that talks with patients offers a promising feedback loop—but the strongest finding is narrower than better treatment. In a Nature Medicine study published October 13, 2026, researchers tested a smartphone platform with 13 hospitalized adults. Patient feedback improved seizure-detector precision in a simulation, without expert intervention; the study left standard treatment unchanged.

The research addresses a gap between what monitoring devices record and what patients can learn from them immediately. The authors describe existing epilepsy devices as unable to share their information with patients in real time. Their prototype creates a two-way channel: the system explains data and sends alerts, while patients contribute symptoms and assessments that can help refine detection.

A chat window connected to live brain data

The prospective, single-center observational study enrolled adults at the Hospital of the University of Pennsylvania’s epilepsy monitoring unit. Eleven had scalp electroencephalography, or EEG, which records brain electrical activity; two had intracranial recordings. Participants were undergoing continuous monitoring to characterize seizures for evaluation before possible surgery, and used the app during their hospital stays.

Hospital software sent live EEG data to a secure cloud environment. Specialized models extracted measures including seizure probability, sleep stages and brief electrical spikes between seizures. The platform also incorporated medication information. These measurements entered a structured database that the conversational AI could query when patients asked about their data.

OpenAI’s GPT-4o-mini powered the conversation through HIPAA-compliant instances. One mode handled general conversation and symptom logging; another retrieved clinical measurements to answer data questions. The language model was therefore the patient-facing interface, while separate algorithms processed the incoming brain recordings.

Incoming messages and outgoing responses passed through moderation checks for safety and clinical compliance. Data questions were also checked for scope and ambiguity before the system retrieved information. Within the chat, patients could receive visualizations such as seizure heatmaps and sleep charts, alongside alerts and scheduled surveys.

Patients supplied more than questions

Across a median participation period of 4.87 days, the patients and system exchanged 1,307 messages. The system initiated 68.9% of them, including conversations, surveys and clinical notifications. Patients initiated 407 messages, mostly general conversation or subjective logging. Their smaller set of data queries concerned seizures, sleep and electrical spikes between seizures.

Patients could mark seizure alerts as true, false or uncertain. According to NIH’s summary, they responded to more than 67% of real-time seizure detections. That gave the researchers patient-generated labels, rather than relying exclusively on expert review, to test whether a detector could learn from the person being monitored.

NIH put the reduction in false alarms at about 80%. The paper identifies the detector-improvement evaluation as a pseudoprospective simulation: patient feedback alone improved precision without expert intervention. That distinction matters when interpreting the result—the demonstrated gain concerns the detector’s performance, not proof that patients experienced fewer seizures.

This proof-of-concept showcases a technology with the exciting potential to shift management of epilepsy and possibly other diseases in a more adaptive direction that puts patients in the driver’s seat

Amy Bany Adams, acting director of NIH’s National Institute of Neurological Disorders and Stroke

The next test is outside the hospital

This was a small inpatient study, not an evaluation of routine use at home. The team is planning larger outpatient studies, according to NIH. Those would move the platform beyond the hospital setting in which this prototype was tested; their results remain ahead.

There were signs of demand before that next step. Co-first author Zack Goldblum told NIH that several patients, unprompted, asked to take the system home. The researchers had to decline. For now, patient enthusiasm and a simulated improvement in detection support further testing—not a take-home treatment.