Epilepsy domain adapted large language models enhance epileptogenic zone prediction from seizure semiology

Seizure semiology provides clinically meaningful clues for localizing the epileptogenic zone (EZ) in drug-resistant focal epilepsy, but its interpretation varies across clinicians and institutions. We aimed to develop and externally validate a domain-adapted large language model for EZ localization from narrative seizure descriptions. We developed EpiSemoLLM using 1,372 surgically verified semiology–EZ pairs collected from 392 published studies and a retrospective cohort from Far Eastern Memorial Hospital. Of these cases, 1,272 were used for model development and 100 were reserved for evaluation. The model combined parameter-efficient fine-tuning with structured epilepsy knowledge and literature-based evidence retrieval. We compared its performance with general-purpose and biomedical language models and with five epileptologists. External validation was conducted in an independent cohort of 83 cases from Taipei Veterans General Hospital evaluated by six epileptologists. Performance was assessed using case-level and region-level localization measures, with uncertainty estimated by paired bootstrap resampling. Here we show that EpiSemoLLM achieves the strongest overall localization performance among the evaluated language models and performs better than the median epileptologist in the held-out cohort. Similar findings are observed in the independent external cohort. The model shows relative strengths in temporal, parietal, occipital, and cingulate localization, whereas clinicians perform better in frontal and insular localization. These findings support the use of domain adaptation and literature-based evidence retrieval for interpreting seizure semiology. EpiSemoLLM may complement clinical reasoning during early presurgical evaluation and help improve the consistency of EZ localization across clinical settings. Prospective validation is needed before routine clinical use. People with epilepsy that does not improve with medicine may be considered for brain surgery. Before surgery, doctors need to identify the part of the brain that is responsible for the seizures and should be targeted during treatment. Signs and symptoms during a seizure can help, but different doctors may interpret them differently. We trained a language model using published reports and hospital records from patients whose seizure-related brain regions had been confirmed by surgery and follow-up. We compared its answers with those of epileptologists and tested it on cases from another hospital. Our model outperformed all baselines and was as accurate as or more accurate than the epileptologists. It is designed to complement clinical expertise and standard presurgical tests, supporting more consistent early assessments. ED SUM: Yang et al. develop and externally validate EpiSemoLLM, a domain-adapted language model that localizes epileptogenic zones from seizure semiology. The model outperforms evaluated language models and performs comparably to or better than epileptologists across internal and external cohorts.