← Back to Home

Giving EEG Monitoring a Voice: Patients with Epilepsy Use Phone Feedback to Help AI Detect Seizures

An inpatient study of 13 people connected EEG data to a conversational interface, allowing patients to ask about changes in their bodies and flag false alarms. Preliminary results support humans and machines working together to improve detection, but whether this can reduce seizures or improve daily life remains unanswered.

By SURL BioNews

EEG monitoring can record electrical activity in the brains of patients with epilepsy, but it may not answer their questions in real time: Was that recent discomfort accompanied by abnormal signals? How has their sleep changed? A University of Pennsylvania team connected monitoring data to a conversational interface on a phone, aiming to help patients understand their own data and use their firsthand experiences to help the system learn. This proof-of-concept study was published in *Nature Medicine* on October 13.

The study enrolled 13 adults undergoing continuous video EEG monitoring in an epilepsy monitoring unit, including 11 using scalp electrodes and 2 undergoing intracranial monitoring. This was a single-center observational study that did not change standard treatment. The platform connected hospital EEG data, secure cloud analysis, and a mobile application, with a large language model supporting two-way communication. The system tested was therefore a prototype for inpatient monitoring; long-term home use remains to be validated.

Patients could ask about EEG and sleep trends, view charts, or record symptoms and how they felt; the system could send alerts and questionnaires. According to Penn Medicine, the team analyzed 1,307 messages exchanged, most concerning seizures, sleep, mood, and mental state. The conversations served not only to present data but also to capture subjective experiences that sensors cannot obtain directly.

This feedback also became material for refining the seizure detection model. Patients could mark which alerts matched their experiences and which might be false alarms, while clinicians could also review events. The paper reported that, in an analysis simulating model updates in chronological order, fine-tuning with patient annotations collected through the platform reduced false alarms by 80.9%. This supports the idea that feedback can improve detection, but it cannot be interpreted directly as a reduction in the number of seizures patients experienced.

There are still clear gaps in detection performance. Among the 11 patients who had access to the annotation feature, clinicians confirmed 59 seizures, of which the original system's alerts correctly captured only 8. Reducing false alarms and avoiding missed actual seizures must both be assessed; improvements in a small inpatient study are also insufficient to demonstrate that the system can reliably provide alerts in daily life.

Another foundation of the study is data available for scrutiny. The public dataset on Pennsieve Discover covers all 13 participants and lists 597 files totaling 285.96 GB. It includes de-identified continuous EEG recordings, seizure timelines, patient reports, sleep and medication records, and model outputs and evaluation metrics before and after fine-tuning. The dataset was updated to its second version on August 13 and is licensed under CC-BY-NC-SA-4.0, providing material for other researchers to reproduce and examine the results.

Moving from the hospital ward to the home will require longer testing with more patients. The paper noted that staff reviewed patient interactions with the AI every day during the study; whether this level of oversight can extend beyond the hospital is an important question for practical deployment. The team will next assess seizure frequency and severity, quality of life, and healthcare use. At this stage, the clearer achievement is the creation of a channel for patients to participate in monitoring and model refinement, while clinical benefits still await direct evidence.

References

  1. Penn Medicine
  2. Pennsieve Discover