AI tools may help doctors better locate seizure origins for epilepsy surgery.
Epilepsy affects millions of people around the world, and in the United States alone, more than three million people live with the condition. Some can manage it with medication, but about one-third of them don’t respond well to drugs. For these patients, epilepsy surgery may be an option. Surgeons can remove the part of the brain where seizures start, which in some cases can stop the seizures entirely or make them much less frequent. But finding that exact spot in the brain, called the epileptogenic zone, is not always easy. Success depends heavily on whether doctors can pinpoint that area correctly before surgery. Right now, surgery helps about half to 60% of patients, meaning there’s a lot of room for improvement.
To figure out where seizures begin, doctors use a mix of brain scans, EEGs (which track brain activity), and detailed notes describing what happens to the person during a seizure. These observations—things like body movements, eye shifts, or sudden changes in posture—are known as seizure semiology. Doctors use these signs to guess which part of the brain might be the problem. However, different doctors or hospitals might describe the same thing in completely different ways. One hospital might say “asymmetric posturing” while another says “asymmetric tonic activity,” even though both mean the same thing (one side of the body stiffens while the other stays relaxed). This kind of inconsistency makes it harder to compare notes or come to a solid conclusion about where the seizures are coming from.
This is where researchers believe large language models, like the one behind popular AI chat programs, can help. These programs are trained to read and understand huge amounts of text, and they can pick up on patterns that humans might miss. A team at Stevens Institute of Technology recently tested this idea. They had five experienced epilepsy doctors read 100 descriptions of seizure symptoms and decide where in the brain the seizures were probably starting. Then, they had ChatGPT do the same thing. In common problem areas like the frontal and temporal lobes, the AI either matched or outperformed the doctors. But in rare cases involving less common parts of the brain, like the insula or cingulate cortex, the human doctors did better.
Based on this, the team took things a step further and built a version of the AI that was trained only on seizure-related information. They called it EpiSemoLLM and ran it on a dedicated server. Their goal is not to replace doctors, but to give them a reliable assistant. Especially in tough cases where seizure signs are unclear or described differently, this tool could help doctors make more confident decisions before surgery. The team believes that the best outcomes will come from combining the knowledge and experience of human doctors with the speed and pattern recognition of AI. By working together, both could help more patients find real relief from seizures and possibly avoid unnecessary or ineffective epilepsy surgery.
Sources:
Large language models help predict epileptogenic zones for surgery
Join the conversation!