Patients in Penn epilepsy unit texted their EEG system and changed how it responded
Thirteen adults in the epilepsy monitoring unit at the Hospital of the University of Pennsylvania spent several days doing something that would have sounded fanciful not long ago: they texted their own brain monitors. Through a smartphone app linked to live EEG data, they asked whether they had had a seizure, how they slept, and what their brain activity looked like. The system answered in plain language and, when appropriate, with charts that translated raw neurophysiology into something a patient could actually use at the bedside.
The pilot was small, but the design cut against a familiar problem in epilepsy care. Continuous EEG produces a flood of information, yet patients usually do not see it in real time, and clinicians often review it on a daily or multiday delay. In this setting, the gap is not only informational. It also leaves the seizure detector vulnerable to false alarms that can confuse patients, burden staff, and make the technology less useful than it could be.
A simulated replay after the fact found that false alarms fell 80% across hospital stays after patients labeled the alerts they received.
Live EEG, conversational AI, and a feedback loop patients could control
The platform streamed each patient’s EEG from the hospital network into Penn’s HIPAA-compliant cloud, where existing models scored the data for seizures, sleep stages, and abnormal electrical spikes between seizures. A version of OpenAI’s GPT-4o-mini handled the conversation through a private endpoint contractually barred from training on or retaining patient data. That separation mattered because the app was not simply chatting for convenience; it was acting as a front end to physiologic data that could alter how the underlying detector behaved.
Patients and the system exchanged 1,307 messages, and about 69% began with the system. Those messages included alerts and surveys sent three times a day. When patients asked about their own data, the agent answered 35 of 38 questions correctly. Simple replies took about two seconds. Questions that required a database search took about nine.
Every exchange passed through an AI moderation check, and a study team member reviewed the conversation daily. Patients were told their care team was not monitoring the app and were instructed to use the unit’s seizure alarm button if they suspected a seizure. Penn Medicine’s AI Governance Committee reviewed the interaction model and later became a standing body. In practical terms, the study had a human backstop at every step, which is exactly what a real-world neurologic monitoring platform needs when patient safety and automated interpretation overlap.
Patient labels sharply reduced simulated false seizure alarms
The app delivered 58 real-time seizure alerts. Patients responded to about two-thirds of them and rejected 22 as false alarms. Those labels were not passive feedback; researchers used them to retrain the seizure detector in simulation after the hospital stay. The median false alarm rate dropped from 0.74 per hour to 0.10 per hour when the model was retrained every three hours on patient input.
The same pattern appeared when the system was trained on presses of the unit’s standard seizure alarm button. In both cases, the detector improved because it was taught to distinguish true ictal events from the many physiologic signals that can mimic them. Seizure detection is rarely a pure signal problem. It is a context problem, and context is exactly what patients possess in abundance: how they felt, what they noticed, and whether the episode matched their own experience.
The retrained model did not catch seizures more often than the original. The residual false alarm burden remained roughly two to four alerts a day, a level that may still frustrate patients and needs more work before routine clinical use. In this pilot, the technology moved decisively in the right direction without yet crossing the threshold into hands-off deployment.
Clinician review still mattered when the system was least certain
A second approach brought experts back into the loop. Once a day, the system selected the 10 stretches of EEG it judged least certain about, each 90 seconds long, and asked clinicians to label them. In simulation, that strategy cut false alarms 82.6%. The logic is straightforward: when the model is unsure, expert annotation can supply high-value labels exactly where the decision boundary is most fragile.
That workflow is attractive because it could reduce the amount of clinician time needed to maintain the detector outside the hospital. A system that can learn from targeted review rather than exhaustive chart-level oversight would be easier to scale, especially for outpatient settings where human bandwidth is limited. The study stops short of showing that this can be done safely at home, but it does show that sparse expert correction can meaningfully improve the detector’s behavior.
Patients wanted to keep using the platform
Five of the 13 patients asked, without prompting, to keep using the platform after discharge. Five completed the exit survey. Three said the app made them more aware of links between behavior, emotions, and health. One wrote that it helped “solidify what I felt to be true between some of my episodes and emotions.” Another said it largely reinforced what they already knew. On a standard usability scale, the five respondents rated the app 83 out of 100.
Those responses point to an important bedside effect. Real-time access to personal neurologic data can turn a hidden process into something patients can interrogate, compare, and remember. For some people with epilepsy, that may sharpen pattern recognition around sleep, stress, or other triggers. For others, the value may be simpler: feeling that the monitor is not just watching them but also teaching them.
Several approached patients declined participation because they were not interested in research or were already enrolled in other studies, suggesting the pilot may have drawn people more open to AI and technology than the broader epilepsy population. The setting also mattered. The epilepsy monitoring unit is an artificial environment in which antiseizure medication is reduced to provoke seizures for diagnostic purposes. Engagement in that setting does not prove benefit at home.
Home monitoring, Parkinson’s programming, and the next regulatory step
The group is preparing an outpatient study to test whether the same platform can help people manage epilepsy outside the hospital. That work will use minimally invasive EEG and wearables and will measure seizure frequency and severity, emergency department visits, and overall health care use. The broader ambition extends beyond epilepsy. Penn is also preparing a study with Lauren Hammer, M.D., Ph.D., to use the platform for deep brain stimulation programming in Parkinson’s disease, where the device could ask whether walking feels adequately treated and then adjust settings during a walk.
The appeal is obvious. Many implanted and wearable systems are already rich in sensor data, but they still miss the lived experience that only patients can describe. A device that can ask a direct question and react in real time could narrow the gap between physiologic measurement and functional status. The same principle could even apply to other closed-loop therapies, where symptom burden does not map neatly onto a single numeric signal.
The work is supported by a $5.6 million NIH Pioneer Award, and the authors expect regulatory approval for this kind of system to arrive in steps, beginning with low-level advisory tools. That is the right sequence for a platform that sits at the intersection of neurology, machine learning, and real-time patient interaction. The remaining questions are not theoretical: whether the false alarm rate can be lowered further, whether home users will sustain engagement, and whether the same feedback loop can hold up once the artificial structure of the monitoring unit is gone.