EpiSemoGPT: A Fine-tuned Large Language Model for Epileptogenic Zone Localization Based on Seizure Semiology with a Performance Comparable to Epileptologists

SignificanceSeizure semiology, the study of signs and clinical manifestations during seizure episodes, provides crucial information for inferring the location of epileptogenic zone (EZ). Given the descriptive nature of seizure semiology and recent advancements in large language models (LLMs), there is a potential to improve the localization accuracy of EZ by leveraging LLMs for interpreting the seizure semiology and mapping its descriptions to the corresponding EZs. This study introduces the Epilepsy Semiology Large Language Model, or EpiSemoLLM, the first fine-tuned LLM designed specifically for this purpose, built upon the Mistral-7B foundational model. MethodA total of 865 cases, each containing seizure semiology descriptions paired with validated EZs via intracranial EEG recording and postoperative surgery outcome, were collected from 189 publications. These collected data cohort of seizure semiology descriptions and EZs, as the high-quality domain specific data, is used to fine-tune the foundational LLM to improve its ability to predict the most likely EZs. To evaluate the performance of the fine-tuned EpiSemoLLM, 100 well-defined cases were tested by comparing the responses from EpiSemoLLM with those from a panel of 5 epileptologists. The responses were graded using the rectified reliability score (rRS) and regional accuracy rate (RAR). Additionally, the performance of EpiSemoLLM was compared with its foundational model, Mistral-7B, and various versions of ChatGPT, Llama as other representative LLMs. ResultIn the comparison with a panel of epileptologists, EpiSemoLLM achieved the following score for regional accuracy rates (RAR) with zero-shot prompts: 60.71% for the frontal lobe, 83.33% for the temporal lobe, 63.16% for the occipital lobe, 45.83% for the parietal lobe, 33.33% for the insular cortex, and 28.57% for the cingulate cortex; and mean rectified reliability score (rRS) 0.291. In comparison, the epileptologists averaged RAR scores were 64.83% for the frontal lobe, 52.22% for the temporal lobe, 60.00% for the occipital lobe, 42.50% for the parietal lobe, 46.00% for the insular cortex, and 8.57% for the cingulate cortex; and rectified reliability score (rRS) with mean of 0.148. Notably, the fine-tuned EpiSemoLLM outperformed its foundational LLM, Mistral-7B-instruct, and various versions of ChatGPT and Llama, particularly in localizing EZs in the insular and cingulate cortex. EpiSemoLLM offers valuable information for presurgical evaluations by identifying the most likely EZ location based on seizure semiology. ConclusionEpiSemoLLM demonstrates comparable performance to epileptologists in inferring EZs from patients seizure semiology, highlighting its value in epilepsy presurgical assessment. EpiSemoLLM outperformed epileptologists in interpreting seizure semiology with EZs originating from the temporal and parietal lobes, as well as the insular cortex. Conversely, epileptologists outperformed EpiSemoLLM regarding EZ localizations in the frontal and occipital lobes and the cingulate cortex. The models superior performance compared to the foundational model underscores the effectiveness of fine-tuning LLMs with high-quality, domain-specific samples.