Neuroscience and Medical AI · us
Letting Patient Feedback Teach AI to Detect Seizures: A 13-Person Study Explores Two-Way EEG Monitoring
Through conversations on their phones, patients can check EEG data, report symptoms, and indicate whether alerts are accurate. A small inpatient study suggests this feedback can help reduce false alarms in simulations; whether it can improve everyday care still requires further validation.
Electroencephalography (EEG) can record electrical activity during seizures, but it cannot necessarily reveal what patients are feeling at the time. If a monitoring system could follow an alert by asking, “Did you just experience any symptoms?”, could patients’ answers help it learn to become more accurate? A small study published in *Nature Medicine* has begun bringing this kind of two-way interaction into seizure monitoring.
A team at the University of Pennsylvania in the United States recruited 13 adult inpatients undergoing continuous video EEG monitoring: 11 used scalp EEG, and 2 underwent intracranial EEG monitoring. This was a single-center observational study that did not alter standard treatment. The median participation period was approximately 4.9 days, primarily to test whether the system could function and how patients used it.
The platform connects EEG acquisition, secure cloud analysis, and conversational AI on a smartphone. Specialized algorithms first analyze signals such as seizure probability and sleep, while the conversational interface lets patients query data, view trends, or report symptoms. The study analyzed 1,307 messages exchanged. According to the University of Pennsylvania’s medical school, the interactions covered seizures, sleep, mood, and mental state, offering an opportunity to supplement physiological records with patients’ subjective experiences.
The key test was whether these responses could improve the detector. Using the collected data, the researchers simulated updating the model in chronological order, then evaluated its performance on subsequent data that had not yet been used. Among 6 patients who provided annotations for smartphone alerts, simulated results after feedback-based fine-tuning showed a lower false alarm rate and an increase in the F1 score, a combined measure of detection performance. This supports the feasibility of patient participation in personalized monitoring, but the model improvements were validated in a simulation analysis and cannot be directly regarded as long-term clinical outcomes.
Fewer false alarms alone do not establish that the system is reliable. The paper also reported that, among real-time smartphone alerts that could be evaluated, only 8 of 58 alerts corresponded to clinically confirmed seizures, while there were 59 confirmed seizures during the same period. Patients may fail to respond, respond late, or be unaware of some seizures. Patient annotations can therefore provide signals for learning, but they still cannot replace professional interpretation.
The publicly available preview of a *Nature Medicine* commentary published on October 16 endorsed the direction of patients helping to fine-tune detectors through conversation, while also noting that the more difficult challenge is getting AI to respond appropriately. When a system discusses EEG data and symptoms, identifying signals is only one part of the process. How it explains uncertain results and when healthcare professionals need to intervene also affect the safety of real-world use.
The next step is to move from short-term inpatient testing into everyday life. The University of Pennsylvania team said it will subsequently evaluate outpatient use and outcomes including seizure frequency, quality of life, and emergency department visits. For now, this study offers a path that can be tested: making patients’ experiences part of the data from which monitoring systems learn. Whether it can reduce the burden of disease remains a question for larger, longer studies.