Can large language models aid pre-surgical epileptogenic zone localization? A multi-source text analysis performance...

Area of Science:

• Artificial Intelligence in Medicine
• Computational Neuroscience
• Medical Informatics

Background:

• Large language models (LLMs) offer potential for analyzing complex biomedical text.
• Localizing the epileptogenic zone (EZ) is crucial for epilepsy surgery but remains challenging.
• The utility of LLMs for EZ localization from pre-surgical data is largely unexplored.

Purpose of the Study:

• To evaluate the performance of leading LLMs in predicting the surgically validated EZ.
• To assess LLMs' ability in laterality classification, lobar localization, and SEEG stratification.
• To determine the value of LLMs as clinical decision-support tools for epilepsy.

Main Methods:

• Three leading LLMs were evaluated on predicting the EZ from unstructured clinical text of 154 drug-resistant epilepsy patients.
• Performance was benchmarked against resected lobes for laterality, lobar localization, and SEEG stratification.
• A modality ablation analysis assessed the impact of different text sources on LLM performance.

Main Results:

• LLMs achieved high accuracy in EZ laterality classification (98.3%–100%).
• GPT-4.1 and Claude 3.7 Sonnet demonstrated superior lobar localization compared to Deepseek-R1 in one cohort.
• LLM predictions showed excellent test-retest reliability (ICC=0.951), with higher SEEG scores for patients undergoing invasive monitoring.

Conclusions:

• LLMs can accurately infer the surgically validated EZ from raw preoperative text.
• LLMs show promise as valuable clinical decision-support tools for epilepsy surgery.
• Further research can optimize LLM application in neurosurgical contexts.