For several days at a time, people with hard-to-treat epilepsy lie in a hospital ward wired to an EEG, a recording of the brain’s electrical activity, so doctors can capture their seizures, often ahead of surgery. At the University of Pennsylvania, 13 such patients could also text the system reading their brain, ask whether they had a seizure in their sleep, and tell it when its alarms were wrong. Brian Litt, the senior author, compares the idea to “your own personal Jiminy Cricket on your shoulder”.
The study was published in Nature Medicine on 13 October. Penn holds a provisional patent on the platform, has named no commercial partner, and is preparing to try it next with deep brain stimulation for Parkinson’s disease, in which an implant sends electrical pulses into the brain.
What the system did
Eleven of the 13 patients wore scalp electrodes. Two had temporary electrodes placed inside the skull for their stay; no permanent implant was involved. Their EEG streamed to Penn’s secure cloud, where software estimated the chance of a seizure, sleep stages and abnormal electrical spikes. A smartphone app with a conversational AI, built on OpenAI’s GPT-4o-mini, sent alerts and surveys, logged symptoms and answered questions using those measures.
Over a median of just under five days per patient, patients and the system exchanged 1,307 messages. Most of what patients wrote was general conversation. When they asked about their own data, mostly seizures and sleep, the AI pulled the right information 35 times out of the 38 requests the team checked. Five of the 13 asked to keep using it after discharge.
What was simulated
For safety, only the original seizure detector ran during the study, and its app alerts were limited to high-confidence detections and muted overnight. The app sent 58 alerts, which caught 8 of the 59 seizures clinicians confirmed. Patients answered about two-thirds of the alerts, marking each as a real seizure, a false alarm or uncertain.
Retraining was tested afterwards, in a simulation that replayed each patient’s recordings. Using patients’ app answers alone, from the six who gave them, false alarms fell by 80 per cent. Using the alarm buttons patients press when they feel a seizure, across all 13, they fell by 77 per cent. The retrained detector caught no more real seizures, and it would still raise roughly two to four false alarms a day.
The group was small and may have leant towards people interested in technology. The hospital ward is an artificial setting, where medication is reduced to provoke seizures. The in-app surveys are not validated clinical measures.
Why it matters
None of the three devices the FDA has approved to treat epilepsy shares its data with patients in real time. A chat layer could give device makers a service patients use every day between clinic visits, and labels supplied by patients could reduce the specialist time needed to keep detection software accurate.
The Parkinson’s plan carries the commercial logic. Deep brain stimulation settings are adjusted in clinic visits, and access is limited partly by the number of specialists who programme them. Penn wants patients to report how they feel, during a walk for example, so that settings can be tuned more often with less clinic time.
Penn built safeguards into the study. A person reviewed every patient conversation daily, the language model ran through a private endpoint barred from training on or keeping patient data, and a Penn AI governance committee set up for the project approved the deployment. The study leaves open who is clinically responsible when a patient acts on an AI answer.
Litt and co-first authors Zack Goldblum and Haoer Shi are named inventors on a provisional US patent Penn filed in December 2025. The patient app code is posted on GitHub under a Penn licence limited to research and non-commercial use. The work is funded by the NIH, including a $5.6 million Pioneer Award to Litt.
Not disclosed
Penn has not named a commercial partner or licensee, a timeline for an outpatient version, or the deep brain stimulation device the Parkinson’s study will use. The study does not report whether the platform changed seizure frequency, quality of life or emergency visits; a planned outpatient study is designed to test those.
What to watch
Whether the Parkinson’s deep brain stimulation study starts, and with which device maker’s implants.
Whether an outpatient version, using minimally invasive or wearable EEG, reduces seizures or emergency visits.
Whether Penn licenses the platform to a device company.
Whether a first FDA submission covers an advisory-only version that informs patients and leaves treatment decisions to clinicians.
Sources
Primary:
- Nature Medicine: Goldblum, Shi and others, Human-artificial intelligence collaboration for comanaging epilepsy in real time, 13 October 2026
- Penn Medicine: AI platform allows real-time conversation between patients and their medical devices, 13 October 2026
- NIH: Patients engage with epilepsy monitor in real time, improving seizure detection, 13 October 2026
- medRxiv via PubMed Central: An Implantable Device that Converses with Patients and Learns to Co-Manage Epilepsy (preprint), 2026
- GitHub: Penn Center for Neuroengineering and Therapeutics, Pioneer chat interface code (research-only licence)
Coverage:
Cross-reference to prior InsideBCI coverage: