Summary
Researchers tested a smartphone-based AI platform that let 13 hospitalised adults with epilepsy receive EEG alerts, query personal data and label seizure detections. In simulations, fine-tuning seizure detectors with patient feedback reduced false alarms, but the study was a small inpatient proof of concept.
Researchers have tested a smartphone platform that lets people with epilepsy receive alerts from their brain recordings, ask questions about processed EEG data and label seizure detections. In a prospective, single-centre observational study of 13 adults undergoing hospital monitoring, patient feedback was also used to fine-tune seizure-detection models in a simulation, where false alarms fell. The work is an inpatient proof of concept, not a test of the system as routine care.
The platform linked scalp or intracranial electroencephalography (EEG) to cloud-based analysis and a chat app. EEG is a recording of electrical activity in the brain; the study’s software applied separate models to extract features such as seizure probability, sleep stages and interictal spikes—brief discharges between seizures. A conversational AI system could use these processed measures to answer data queries, while also sending alerts and surveys. The researchers used a HIPAA-compliant version of GPT-4o-mini for the chat and data-retrieval functions; the AI was not designed to make autonomous clinical decisions.
A two-way interface for EEG monitoring
The 13 participants were adults admitted to the Hospital of the University of Pennsylvania’s epilepsy monitoring unit for continuous video-EEG monitoring as part of presurgical evaluation. Eleven had scalp EEG and two had intracranial EEG. Their standard treatment was not changed for the study. Over a median participation period of 4.87 days, patients and the system exchanged 1,307 messages.
Patients could ask about their own recorded data and respond to seizure alerts with an annotation such as “yes”, “no” or “uncertain”. Those responses could then serve as labels for detector training: in principle, a patient’s feedback can help a model distinguish a genuine event from a false alarm. The study also used surveys about sleep, mood and other experiences, but these were engagement measures rather than validated epilepsy-specific clinical outcomes.
What the detector simulations found
The researchers tested model updates using patient labels and evaluated the updated detectors on later, unseen recordings from the hospital stays. In a simulation using annotations gathered through the app, the median false-alarm rate across the monitored stays fell from 0.736 to 0.104 per hour, while the median F1 score—a measure that combines precision and sensitivity—rose from 0.021 to 0.143. The analysis involved six patients who provided app annotations.
The detector’s alerts were not a reliable count of all seizures in this small study: app alerts captured 8 of 59 seizures confirmed by clinicians, and patients provided annotations confirming five of those 59. That context matters when interpreting the improvement in false alarms. The results show that patient feedback could help refine a detector in this simulation; they do not show that the platform was ready to identify seizures reliably in everyday life.
A separate simulation using patients’ clinical push-button seizure alarms as training labels included 13 patients. Across the simulated hospital stays, the median false-alarm rate declined from 0.469 to 0.097 per hour and the F1 score increased from 0.063 to 0.160, with median sensitivity preserved. This was a distinct analysis from fine-tuning with app annotations.
The researchers describe the system as a proof of concept. Participants were in a supervised hospital setting for only a few days, and the models were evaluated through simulations rather than a trial of outpatient care. The authors say that testing the platform during longer-term use outside hospital is a next step. Establishing whether it can provide dependable alerts and useful support in that setting will be important before its role in routine epilepsy management can be assessed.