AI Epilepsy Platform Cuts False Alarms 80% in 13-Patient Real-Time Validation
Updated
Updated · Nature.com · Oct 14
AI Epilepsy Platform Cuts False Alarms 80% in 13-Patient Real-Time Validation
3 articles · Updated · Nature.com · Oct 14
Summary
A University of Pennsylvania prototype linked real-time scalp or intracranial EEG to a smartphone chat app and was prospectively tested in 13 epilepsy inpatients, letting the system alert patients, answer questions and collect symptom feedback.
Patient feedback then retrained seizure detectors without expert input in pseudoprospective simulations, cutting false alarms by about 77% to 81% while improving F1 scores and preserving baseline sensitivity.
Across 1,307 messages, the data-query agent answered retrieval tasks with 92.1% accuracy, the message router reached 89.7%, and usability scored 83 out of 100 among five surveyed patients.
The study targets a gap in current epilepsy devices, which can detect biomarkers but do not share them with patients in real time; about one-third of the world’s 70 million epilepsy patients still have uncontrolled seizures.
Researchers called the inpatient study a proof of concept and said outpatient trials are next, with possible extensions to Parkinson’s disease, cardiac devices and insulin pumps.