http://icbssr.comEpilepsy is one of the most prevalent neurological disorders, affecting over 70 million people globally. In the U.S. alone, approximately 3.4 million individuals live with this condition. While many can manage their seizures with medication, about one-third of patients do not respond to drug therapy. For these individuals, surgical removal of the epileptogenic zone (EZ)—the specific area of the brain where seizures originate can offer a path to seizure freedom.

However, the success rate of such resective surgeries remains modest, hovering around 50–60%. A major reason for this is the challenge of accurately identifying the EZ. To locate it, patients undergo multiple diagnostic tests such as MRI, electroencephalography (EEG), and intracranial EEG. Epileptologists analyze this data to describe seizure semiology, the set of symptoms and behaviors observed during seizures, which helps in estimating the EZ’s location.

Yet, the terminology used to describe seizure semiology can vary significantly across epilepsy centers. “Different epilepsy centers may use different terms for the same seizure presentation,” explains Dr. Feng Liu, Assistant Professor in the Department of Systems and Enterprises at Stevens Institute of Technology. For instance, terms like “asymmetric posturing” and “asymmetric tonic activity” might refer to the same clinical observation. This inconsistency can create challenges for surgeons attempting to interpret the data.

Given the text-based nature of seizure descriptions, Large Language Models (LLMs) like ChatGPT trained on vast public datasets offer potential for improving EZ identification. Dr. Liu and his collaborators, including clinicians from institutions such as Case Western Reserve, Rutgers, UCSF, and Goethe University, explored the use of ChatGPT in interpreting seizure semiology to predict EZ locations.

In their study, five board-certified epileptologists completed a survey with 100 questions focused on EZ localization based on seizure descriptions. ChatGPT was given the same task. Results showed that ChatGPT performed on par with or even outperformed human experts in identifying EZs in commonly affected areas like the frontal and temporal lobes. However, epileptologists were more accurate when dealing with rarer EZ regions, such as the insula and cingulate cortex. These findings were published in the Journal of Medical Internet Research on May 12.

To enhance accuracy further, Liu's team developed EpiSemoLLM, the first LLM tailored specifically for seizure semiology interpretation. Hosted on Stevens’ GPU servers, this model aims to assist neurosurgeons and epileptologists during the presurgical evaluation phase.

“Our findings suggest that LLMs, especially when fine-tuned, can serve as valuable tools in preoperative assessments for epilepsy surgery,” says Liu. “But the most effective approach is a collaborative one where humans and AI work together.”

Source: https://www.stevens.edu/news/chatgpt-helps-pinpoint-precise-locations-of-seizures-in-the-brain-aiding