In the epilepsy unit, specialists still read days of EEG after the recording
About 70 million people worldwide have epilepsy, and roughly one-third have seizures that medication does not control. In an epilepsy monitoring unit, patients spend days connected to scalp or intracranial EEG. Specialists review those recordings after the fact. For decades, researchers have pursued automated seizure detection, but false alarms remain a central challenge for systems intended to monitor people beyond the hospital.
The reported 80% reduction came from patients correcting the detector, not from a new detector running live.
Penn tested an EEG chat app with 13 inpatients
On 13 October 2026, a Penn team led by Brian Litt published a Nature Medicine study of a conversational system connected to patients’ EEG monitors. Thirteen inpatients used it: 11 with scalp EEG and two with intracranial EEG. A smartphone app let patients exchange information with their own monitoring system.
The detector improved in a simulation using patient feedback alone. That shifts attention toward a practical question for future systems: how to collect useful patient labels consistently. The study did not establish that patient feedback replaces specialist review, or that the retrained detector improves care in real time.
Litt has described the aim as a device that can detect changes and alert someone like “your own personal Jiminy Cricket on your shoulder.” The prototype adds a second channel: patients can respond to the system and contribute information used to retrain its detector.
The app connected EEG, cloud analysis and patient data
The prototype joined three parts: software that captured EEG, a secure cloud environment that analyzed incoming data in near real time, and a smartphone application, according to Penn Medicine. The app’s data agent, built with LangGraph, used SQL to answer questions about a patient’s seizure events, sleep reports and spike-rate data. Language-model moderation checked both user messages and AI responses. The research code is available in Penn’s GitHub repository, which says the software is not a medical device and must not be used for clinical decisions outside an approved research protocol.
The app reportedly used OpenAI’s GPT-4o-mini, according to Inside BCI’s account of the study. It processed biomarkers related to sleep, medication, cognition and seizure probability, sent notifications about events and trends, and recorded symptoms and responses. Patients’ feedback supplied annotations for retraining the seizure detector.
Patients and the system exchanged 1,307 messages over a median of just under five days per patient. In 38 checked requests about their own data, the AI retrieved the right information 35 times. Five of the 13 patients asked to keep using the system after discharge. These figures measure use and data retrieval, not clinical benefit.
For safety, only the original detector ran during the study, and app alerts were limited to high-confidence detections, according to the same report. The retrained detector was not tested as a live alerting system.
The 80% figure comes from a pseudoprospective simulation
In a pseudoprospective simulation, seizure detectors improved in precision using patient feedback alone, without expert intervention, the Nature Medicine paper reports. The National Institutes of Health said patient feedback reduced false alarms by about 80%. The result supports the value of patient annotations in that evaluation. It does not show that an available device now delivers an 80% reduction to patients.
Thirteen patients took part in an inpatient prototype; the evaluation was pseudoprospective, not a live trial of a retrained detector. The study therefore shows improved precision in simulation, not that the same result would hold at home or in routine care.
The study’s dataset is available through NEMAR: 69 EEG recordings, one EDF file per recording day, with annotations for seizures, patient button presses, sleep, medication and platform events. That release gives other researchers material to inspect and reuse. Penn reportedly holds a provisional patent and has named no commercial partner, according to Inside BCI.
The NIH described the project as a “super intelligence (SI)-based interface,” a phrase absent from the paper and Penn’s account. The more useful description is a chat interface tied to patient data and a seizure detector. In a Nature Medicine commentary, Samden Lhatoo and Vladimir Vashin identified the harder challenge: building AI that responds appropriately. Retrieving a patient’s data and answering a clinical question safely are distinct tasks.
Device makers will have to prove their feedback loops work
The study points toward a different engineering target for seizure monitoring: collecting patient labels that are frequent and reliable enough to improve a detector. Sensitivity and precision will remain important, but a headline accuracy figure cannot show whether a product can gather useful feedback from the person wearing it.
A detector can flag an event. It cannot establish whether the event matches a patient’s experience without a label from a patient or an expert. Expert annotation takes specialist time; a conversational interface offers another route, though the Penn study did not compare the labor or reliability of those routes. Patient-facing software could become part of the detection system rather than a companion feature, if it produces dependable annotations and patients continue to use it.
Prediction: by October 2028, at least one major EEG or neurostimulation vendor will license or clone a patient-facing chat layer. The commercial logic is that a feedback workflow may make a detector more useful over time. The harder part to reproduce may be the combination of accessible explanations, easy reporting and patient trust, not access to an LLM. The prediction fails if vendors do not adopt such interfaces within that window, or if real-world use produces too few reliable labels to change detector performance.
Penn is preparing to test the platform with deep brain stimulation for Parkinson’s disease, according to Inside BCI. That work could test whether chat-mediated labeling applies beyond epilepsy. Success would support a broader approach for devices that collect signals requiring interpretation; failure would suggest that the epilepsy monitoring setting supplies conditions the method needs.
There is also a regulatory question, not yet an established consequence of this study. If a patient’s message changes labels used to retrain an algorithm in a regulated device, regulators and manufacturers would need to define how those inputs are recorded, reviewed and monitored after deployment. The study did not test a deployed, changing medical device, so that framework remains a future issue rather than a current finding.
The strongest counterargument is the setting. Patients wired to video-EEG in a hospital may report events differently from people at home using ambulatory monitors or implants. Their labels may be more consistent, and the research team’s supervision may matter. If home users do not keep engaging, or their reports prove too noisy, the feedback loop may offer little benefit outside the unit. That uncertainty does not make the interface irrelevant; it makes label quality and sustained use the next claims vendors would need to demonstrate.
For vendors whose pitch rests on trusting a black-box detector, Penn’s simulation raises a challenge: product performance may depend partly on whether patients can correct the system. The model still matters. So does the route by which it receives evidence from the person it monitors.
Penn’s prototype gives patients a way to answer the monitor
For people whose seizures resist medication, the prototype offers a way to report what they experienced and ask about data from their monitoring. The study does not show how well that exchange would work at home. It does show that patients can contribute annotations to a detector’s training process.
That contribution also raises questions about consent, data ownership and whether patient reports count as clinically admissible labels. Patients risk becoming unpaid annotators if their time improves a system without a clear benefit to them. Five patients asking to continue after discharge is a sign of interest, not evidence of lasting engagement.
The monitoring unit still holds the same scene: a patient connected to EEG for days, with a specialist reviewing recordings afterward. Penn’s prototype adds a question the machine can ask while the recording is underway, and a patient’s answer that may change what the detector learns.
