Inpatient prototype of a conversational EEG device that retrains its own seizure detector
A prototype linking inpatient scalp and intracranial EEG to large language models let 13 epilepsy patients exchange messages with their monitoring system through a phone app, and the seizure-detection algorithms improved their precision over several days from that feedback without expert reprogramming.
- Model
- Unnamed secure large language models behind an AI agent; no vendor or version is given in the preprint abstract
- Field
- Medicine
- Date
- 2026-01-27
- Human collaborators
- Zack Goldblum, Brian Litt
What was found
Implantable epilepsy devices can detect impending seizures, missed medication and impaired cognition, but have no way to tell the person carrying them in real time, which the authors describe as a limitation common to most medical devices. The prototype links scalp and intracranial EEG to secure large language models that communicate bidirectionally with patients through a smartphone app, with an AI agent consuming biomarkers of sleep, medication level, cognition and seizure risk extracted from brain activity. Either side can begin a conversation: the agent reports clinical events and physiological trends, records symptoms, responses and behaviours, and retrains itself on what it gathers. The authors demonstrate the platform in 13 patients undergoing inpatient video-EEG monitoring and report that seizure-detection algorithms optimised their precision over several days without expert intervention, which they contrast with the months of iterative in-person physician programming currently required. Patients rated the system highly usable and responded positively to messages about sleep, cognition and seizure risk, and the authors describe safeguards including efficient expert review for further fine-tuning and controls on communication content.
Novelty check
Searched Europe PMC and Crossref on 2026-10-22 for a journal version (Europe PMC TITLE:"Co-Manage Epilepsy", and a Crossref bibliographic search on the full title). Both return only the medRxiv preprint, doi 10.64898/2026.01.26.26344234, posted 27 January 2026 and indexed by Europe PMC as PMID 41646755 and PMC12870640 with its journal field reading medRxiv. Penn Medicine's October 2026 announcement reports publication in Nature Medicine, and no such record exists in either index as of that date, so the published version is unconfirmed. No prior registry entry covers an implantable or conversational seizure-monitoring device, and the registry's other conversational-AI medical entries concern diagnostic dialogue rather than device telemetry.
Caveats and known objections
Graded claimed: the only artifact this entry can point to is a preprint, so nothing here has been through peer review. Penn Medicine's announcement says the work appeared in Nature Medicine, and no journal record was found in Crossref or Europe PMC, both searched on 2026-10-22; confirming that publication would move the grade to peer-reviewed, and nothing else needs to change for it. Thirteen inpatients at a single centre, all of them already admitted to an epilepsy monitoring unit, which measures a prototype under continuous observation and says nothing about use at home. The endpoints are the detector's own precision and usability ratings, both surrogates, with no seizure-control, safety or other patient outcome. The abstract reports the precision gain without numbers, and the roughly 80% reduction in false seizure alarms that press coverage attributes to the study could not be verified for this entry: the NIH and Penn Medicine pages refuse automated fetch and the PubMed Central copy serves a bot check, so that figure is deliberately not stated above. Funded by NIH NINDS grants to the investigators, with no company role recorded in the indexed metadata. Autonomy is ai-assisted: the system messages patients and tunes its own detector while clinicians retain oversight and expert review is built into further fine-tuning. The device is investigational, and nothing here bears on a marketing authorisation.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Claimed and AI-assisted.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded claimed for verification and ai-assisted for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Inpatient prototype of a conversational EEG device that retrains its own seizure detector. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-01-27-conversational-epilepsy-device-prototypeBibTeX
@misc{whataifound-universityofpennsylvania-2026-prototype,
title = {Inpatient prototype of a conversational EEG device that retrains its own seizure detector},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by University of Pennsylvania. Verification: Claimed. Autonomy: AI-assisted.},
url = {https://whataifound.org/finding/2026-01-27-conversational-epilepsy-device-prototype}
}