West Virginia Univ. Researchers Using AI to Enhance Epilepsy Care

Artificial intelligence is allowing clinicians to search EHR data in new ways to glean insights into care gaps for patients with complex conditions. West Virginia University (WVU) pediatric neurosurgeon P. David Adelson, M.D., recently spoke with Healthcare Innovation about a partnership to improve outcomes for people with epilepsy, including identifying those who might be candidates for surgery or are missing test results.

Adelson, who is Steve A. Antoline Endowed Chair for Children's Neurosciences and vice chair of West Virginia University’s Rockefeller Neuroscience Institute, explained that epilepsy is a very complex disease process, so being able to understand it requires specialization. 

“Not only are the patients undergoing regular evaluations for assessing their seizure syndrome, but they're also undergoing imaging, neuro-cognitive testing and evaluation, and if the medicines aren't working, it involves trials of different therapies, so you've got a lot of interchangeable parts,” he said. “If you think about the healthcare climate these days, especially with documentation and electronic health records, we often get deluged with paperwork, and, the demands on clinicians for access have grown so that we don't have doctors spending an hour or two with a patient each visit.”

To address this patient data challenge, Adelson is working with a company called Sephos AI on integrating the company’s Autonomous Registry and Analytics (AURA) platform, which utilizes AI, into WVU Medicine’s electronic health record. The registry includes 3,348 individuals with epilepsy (1,176 children and 2,172 adults).


During a 90-day prospective evaluation, AURA assessed clinicians’ unstructured notes, imaging reports, assessments and documentation for 820 scheduled patient visits with neurologists. AURA identified 88 patients (11%) who met the criteria for drug-resistant epilepsy but who had not been referred for surgery. (American Academy of Neurology guidelines recommend that all people with drug-resistant epilepsy be evaluated for surgery, although ultimately not everyone is a candidate.)

AURA also highlighted care gaps in patients’ records, including:
 
• Outdated or missing magnetic resonance imaging (MRI) tests (54%).
• Missing neuropsychological evaluations (91%).
• Missing electroencephalographs, or EEGs (35%).
 
Across the full registry, folic acid deficiencies were identified in 81% of women of childbearing age. Researchers also determined that the technology increased post-surgical outcome documentation from 1% to 70.9%. Healthcare teams reported more comprehensive evaluations, better collaboration in care planning and faster recognition of patients who might benefit from surgery. 

Adelson spoke about how AI tools and registries can not only identify gaps in care, but also start to fill some of those gaps. “It can make inferences to be able to better phenotype the patient. It can give us a characterization of what's going on, and then serve as a tool to offer reminders to the provider,” he explained. “The AI agent is really like an artificial fellow. It’s that workhorse behind you. You still need the expertise that gives the nuance to care, but now that we have identified this patient doesn't have an MRI, we’ll make a reminder for that. They need an EEG; we’ll make a reminder. They've never had neuropsychological evaluation; we’ll schedule it. This takes the busy work out of the physician’s hands and allows them to focus on the optimal care for this patient.”

Besides creating its own epilepsy registry, WVU is also involved in other consortia, such as the Epilepsy Learning Healthcare System (ELHS) and the Pediatric Epilepsy Research Consortium. “What's nice about our registry is that with this AI, we can pull the fields for each of these consortia directly from our EHR and then export that data in a much more simple way, whereas in the past, we would have to have a nurse coordinator or medical students pulling all this data manually,” Adelson said. “Now we can define our fields, and the AI will pull all of that — even from unstructured data, and be able to place it into the fields, to be able to be exported into a consortium.”

Adelson also described how AURA employs a multi-agent framework to respond to prompts, rather than relying on a single model. Each AI agent is trained to perform a specific task, such as identifying surgical candidates or checking for safety concerns. A separate AI judge then reviews the outputs and compares them to the source data, providing feedback until the information is accurate and fully grounded in the medical record. “As we fine-tune the prompts, this will get even better as it goes,” he said.

“The great thing about this platform is that we could use the same model to go after other areas like stroke or movement disorders or other neurology  diagnoses or even non-neuro diagnoses — heart failure, diabetes, things like that,” Adelson said. 

The work was presented at the recent annual meeting of the American Epilepsy Society. “Research like this demonstrates the growing potential of responsible artificial intelligence to enhance epilepsy care,” said Howard P. Goodkin, M.D., Ph.D., president of the American Epilepsy Society, in a statement. “Instead of replacing the epilepsy specialist, AI acts as a partner, enhancing human expertise in medicine by tracking complex medical information, identifying gaps and prompting action to ensure care remains on course over time.”