AI Chat Platform Lets Epilepsy Patients Query Their Brain Data and Cuts False Seizure Alarms

In a prospective, single-center observational study reported in Nature Medicine, researchers led by Brian Litt at the University of Pennsylvania, with first author Zack Goldblum, tested a platform that links scalp or intracranial EEG to a secure, large language model-based AI that talks with patients through a smartphone app. Thirteen adults undergoing video-EEG monitoring in the epilepsy monitoring unit used it during their stay—11 with scalp EEG and 2 with intracranial EEG. The goal was not to treat seizures but to see whether patients and an AI could meaningfully communicate about brain data in real time, and whether that conversation could improve seizure-detection algorithms.

The 13 patients exchanged 1,307 messages with the system. The AI initiated most interactions (68.9%), including scheduled surveys and clinical alerts; patients initiated 31.1%, mostly general chat, with smaller numbers querying seizure, sleep, and spike data. Seizure alerts drew the fastest responses: 44.8% within 10 minutes and 60.3% within 30 minutes. Survey response was 55.0% before a 3-hour expiration. Patients responded to 67.2% of real-time seizure detections, labeling them as true seizures, false positives, or uncertain. Usability was rated excellent on average (System Usability Scale 83/100; n=5), and 5 of 13 patients asked for outpatient access.

The platform's main technical test was whether patient feedback—imperfect labels—could retrain pretrained seizure detectors. In simulations of prospective use, fine-tuning with patient push-button alarms reduced false alarms per hour by a median of 77.8% and improved F1 score, while preserving baseline sensitivity. With annotations collected directly through the app, false alarms fell by about 81% and F1 improved; across the monitoring stay, false alarm rate decreased 80.0%. Clinician-in-the-loop fine-tuning performed even better in simulation, reducing false alarms by 82.6%. The authors emphasize that these are pseudoprospective simulations on a small inpatient sample, not proof that the system improves clinical outcomes.

For epilepsy, the appeal is bidirectional feedback. Current implantable devices can detect some events but generally cannot tell patients what they know in real time or learn from patient-reported context. A device that can explain seizure risk, flag missed medication or sleep loss, and let patients annotate events could make monitoring more useful and personal. The authors also see the platform as a template for deep brain stimulation in Parkinson's disease, cardiac devices, insulin pumps, and other conditions where patient symptoms and device data need to meet.

Key limits: only 13 patients at one center, in an artificial inpatient setting where medications are often reduced to provoke seizures; short participation; possible selection bias toward technology-interested patients. Surveys were engagement tools, not validated patient-reported outcomes. After fine-tuning, false alarms still occurred roughly 2–4 times per day in some analyses, which could burden patients. Sensitivity did not improve, the detectors used were lightweight rather than state-of-the-art, and no clinical benefit was measured.

The next step is outpatient testing—longer-term data from implanted systems, rigorous user-experience studies, and evidence that the approach reduces seizures or improves daily life. The study suggests the conversation between patients and devices is feasible; whether it changes care remains to be shown.

Reference: Goldblum, Z., Shi, H., Xu, Z. et al. Human–artificial intelligence collaboration for comanaging epilepsy in real time. Nature Medicine (2026). https://doi.org/10.1038/s41591-026-04680-4