← Back to Home

Letting Patients Respond to EEG Alerts: Small Study of an AI Epilepsy Platform Reduces False Alarms by 80%

Thirteen hospitalized patients used their phones to ask about EEG information and help label suspected seizures. This two-way exchange improved detection performance, but longer-term validation remains necessary before use in everyday life.

By SURL BioNews

EEG monitoring can capture signals of suspected epileptic seizures, but it can also flag harmless activity, leaving large numbers of records for experts to review. A team at the University of Pennsylvania in the United States tried incorporating patients’ experiences into this process: having a monitoring platform communicate with patients through their phones, then using their responses to improve interpretation. In a proof-of-concept study involving 13 participants, patient feedback reduced the detection model’s false-alarm rate by approximately 80%.

The study was published in *Nature Medicine* on October 13. All participants were adults with epilepsy undergoing presurgical evaluation at the Hospital of the University of Pennsylvania; 11 underwent scalp EEG monitoring and 2 underwent intracranial EEG monitoring. This was a single-center observational study, and standard treatment was unchanged during testing.

The platform connects EEG analysis results to a large language model interface, allowing patients to receive notifications, record symptoms, and ask questions themselves, such as whether suspected seizures occurred during sleep. According to the US National Institutes of Health, more than 1,300 messages were exchanged during the study; more than 92% of the system’s answers to questions about brain activity were rated accurate and relevant. This result measures the quality of information responses and cannot be directly treated as seizure detection accuracy.

Patients also responded to more than 67% of real-time seizure prompts, labeling events as real, false alarms, or uncertain. The research team then used these labels to train the detection model. The paper states that the improvement in false alarms was validated in an analysis that simulated actual operation in chronological order, while maintaining the original detection sensitivity; the 80% reduction therefore reflects model performance and does not mean that patients experienced fewer seizures.

Patients’ experiences can supplement signal interpretation, but they also have blind spots: some seizures go unnoticed, and post-seizure memory impairment or delayed responses may also affect labeling. With only 13 participants, a short observation period, and an inpatient setting with continuous supervision, the current results primarily support technical feasibility. They are not yet sufficient to demonstrate that long-term home use can improve health or quality of life.

The team is planning a larger outpatient study. The next questions are whether patients can remain engaged after leaving the hospital ward, whether the system can maintain reliable interpretation, and how experts should review incomplete feedback. This platform offers a concrete direction: allowing patients to participate in refining monitoring tools as well as receive monitoring results.

References

  1. Medical Xpress / National Institutes of Health