Most medical devices are good at collecting data and bad at talking about it. A new Penn Medicine system tries to change that by letting patients message their device like a colleague, and the early results from a small epilepsy study are worth watching.

Researchers at the Perelman School of Medicine and the Center for Neuroengineering and Therapeutics reported the proof-of-concept work in Nature Medicine. Thirteen adults in Penn’s Epilepsy Monitoring Unit, all on continuous EEG, used a smartphone app connected to an AI conversational agent built on technology similar to ChatGPT.

How the system works

Three pieces make it run:

• Software that streams EEG data from electrodes in the brain
• A HIPAA-compliant cloud environment that analyzes the signal in near real time
• A phone app with the conversational agent, built by Penn Bioengineering PhD students

Patients could ask about their brain activity, request charts of trends, log symptoms, and answer surveys three times a day on sleep, anxiety and mood. The device could also reach out when it spotted something relevant.

What patients and clinicians taught it

The team analyzed 1,307 messages, mostly about seizures, sleep, mood and energy. Patients also flagged which detected events were real seizures and which were false alarms, and the system’s accuracy improved substantially. Clinicians could do the same, and an “expert in the loop” algorithm limited their input to the most valuable annotations. That matters, because specialist time is the bottleneck for any long-term outpatient use.

Participants responded well. One said the platform helped “solidify what I felt to be true between some of my episodes and emotions.” Others valued the sleep data and the prospect of warnings before major seizures.

Where it goes next

Epilepsy is only the starting point. Senior author Brian Litt, MD, sees uses in cardiac devices, insulin pumps, and neuromodulation for Parkinson’s, depression, OCD and addiction. His team is already planning work with Lauren Hammer, MD, PhD, to tune deep brain stimulation for Parkinson’s automatically, and is adding wearables for insomnia and vital signs monitoring.

The caveats are real. This was a small, inpatient study, and it measured usability and detection, not outcomes. Future trials will test whether the approach reduces seizure frequency and severity, improves quality of life, and cuts emergency visits. Outpatient testing with minimally invasive EEG and wearables is also planned. The work is backed by a $5.6 million NIH Pioneer Award.

Still, the direction is clear. Devices that listen to patients, not just their physiology, could make chronic disease management far more personal.