Neural Noise

AI-assisted localization of seizure foci from iEEG for presurgical planning

2 min read

  • AI (e.g. ChatGPT) can help pinpoint precise locations of seizures in the brain from intracranial/clinical data, aiding neurosurgeons in presurgical planning.1
  • The application combines neural data analysis and AI for clinical neurophysiology and is relevant to iEEG and the presurgical workflow.1 1

Gardner updates

  • Functional connectivity features from interictal intracranial EEG support prediction of seizure onset zones, reducing the need for ictal captures. 2
  • The approach is transferable to presurgical planning and closed-loop neuromodulation. 2

Weekly enrichment (2026-07-20)

  • The underlying study, published in the Journal of Medical Internet Research (JMIR, 2025;e69173) on May 12, 2025, systematically evaluated ChatGPT’s ability to interpret seizure semiology text and localize the epileptogenic zone (EZ) for drug-resistant focal epilepsy.3 4
  • Two cohorts were used: a public cohort of 852 semiology–EZ pairs drawn from 193 peer-reviewed publications, and a private cohort of 184 pairs from Far Eastern Memorial Hospital (FEMH) in Taiwan.4
  • ChatGPT was tested with two strategies, zero-shot prompting (ZSP) and few-shot prompting (FSP), and benchmarked against 8 epileptologists who each interpreted 100 randomly selected semiology records.4
  • Performance was scored with regional sensitivity (RSens), weighted sensitivity (WSens), and net positive inference rate (NPIR); RSens reached 80–90% for the frontal and temporal lobes but fell to 20–40% for parietal, occipital, and insular cortex and only about 3% for the cingulate cortex.4
  • Weighted sensitivity consistently exceeded 67% and mean NPIR stayed near 0, and ChatGPT-4 significantly outperformed epileptologists on RSens for the commonly represented frontal and temporal lobes (p < 0.001).4
  • Epileptologists remained more accurate than ChatGPT for rare EZ sites such as the insula and cingulate cortex, supporting an assistive, human-in-the-loop role rather than autonomous localization.5
  • Motivating the work, current resective epilepsy surgery has only a roughly 50–60% success rate, partly because EZs are not accurately identified from MRI, EEG, and iEEG workups.5
  • The team (led by Y. Liu at Stevens Institute of Technology, with collaborators at Case Western Reserve, Rutgers, UCSF, and Goethe University) subsequently built a domain-specific model, EpiSemoLLM, hosted on a Stevens GPU server for semiology interpretation.5

Footnotes

  1. https://news.google.com/rss/articles/CBMiiAFBVV95cUxOSW4zTFoxYXItUHhQdWJEYzh1TExPZHY0V0VpSTdHZXVKWlRFdkg2SlNMc3NPWFhIZVUwQWhmZ2pRb1g4NWlybTdKSXY5RmM3bXFTR0w2R3lQY25CTE5uZ1dJcXc5ZUVhX2tXUktaRUxHcUdITDhlQkdGUTJzeVU4RElha2xSV3Fa?oc=5 ↩ ↩2 ↩3

  2. https://news.google.com/rss/articles/CBMiX0FVX3lxTFBUeTF1SzZmRTVxM003bTdhRGMwc2VjUmotSDJTYjVmZXkxMlpEVDdJQ25vWFpFRnVOaE5oTjFfOUo4MlNKQ2x2eU9YR1Y0SVRHOTVhQjlrQUFSWTE0NGVr?oc=5 ↩ ↩2

  3. https://www.jmir.org/2025/1/e69173/ ↩

  4. https://pubmed.ncbi.nlm.nih.gov/39974103/ ↩ ↩2 ↩3 ↩4 ↩5

  5. https://www.stevens.edu/news/chatgpt-helps-pinpoint-precise-locations-of-seizures-in-the-brain-aiding ↩ ↩2 ↩3

Graph View