Epilepsy-specific language model improves seizure semiology localization from narrative descriptions
A domain-adapted large language model built for epilepsy has outperformed general-purpose and biomedical models in localizing the epileptogenic zone from seizure semiology, a task that sits at the center of presurgical decision-making for drug-resistant focal epilepsy. The model, EpiSemoLLM, was trained on surgically verified semiology–epileptogenic zone pairs and then tested against both expert clinicians and independent validation data. In a field where narrative symptom descriptions often guide the next diagnostic step, even small gains in consistency can matter at the bedside.
The study addresses a familiar problem in epilepsy surgery workups: seizure semiology carries useful localization clues, but interpretation is variable. A brief description of eye deviation, automatisms, speech arrest, head version, or sensory aura may point toward a specific cortical network, yet the same narrative can trigger different localization judgments across institutions. A model that can standardize that reasoning, while retaining access to literature-based evidence, has obvious appeal for early presurgical review.
Training on surgically verified semiology–EZ pairs
EpiSemoLLM was developed using 1,372 surgically verified semiology–epileptogenic zone pairs collected from 392 published studies and a retrospective cohort from Far Eastern Memorial Hospital. Of these, 1,272 cases were used for model development and 100 were set aside for evaluation. The architecture combined parameter-efficient fine-tuning with structured epilepsy knowledge and literature-based evidence retrieval, a design meant to keep the model grounded in domain-specific facts rather than broad, non-specialized language patterns.
That choice matters. In epilepsy, semiology is not just text; it is an encoded clinical phenotype shaped by seizure onset, spread through functional networks, and the patient’s ability to report or recall events. A model trained only on generic medical text would be expected to miss some of those subtleties. By pairing structured epilepsy knowledge with retrieval from published evidence, the investigators built a system that could connect symptom language to localization logic more like an epileptologist than a general chatbot.
Better overall localization than baseline models and median experts
Across the held-out cohort, EpiSemoLLM achieved the strongest overall localization performance among the evaluated language models and performed better than the median epileptologist. The same pattern held in the independent external cohort of 83 cases from Taipei Veterans General Hospital, where six epileptologists also participated in evaluation. The model was assessed using case-level and region-level localization measures, with uncertainty estimated through paired bootstrap resampling.
One sentence captures the clinical meaning: the model was not simply fluent, it was useful. In presurgical epilepsy care, the difference between a vague and a plausible localization can change which tests are ordered next, how quickly a surgical conference converges, and whether a patient moves toward intracranial monitoring or a more conservative path. A tool that narrows the field early could reduce ambiguity before high-stakes procedural decisions are made.
The strongest performance appeared in temporal, parietal, occipital, and cingulate localization. Clinicians, by contrast, performed better in frontal and insular localization. This split is clinically intuitive. Temporal lobe semiology often contains recognizable patterns that recur across cases and literature, while frontal and insular seizures can be more protean, rapidly spreading, and harder to infer from narrative alone. The human advantage in those regions likely reflects the value of tacit pattern recognition, bedside nuance, and awareness of syndromic outliers.
Why domain adaptation changed the result
The design of EpiSemoLLM helps explain why it outperformed broader models. Seizure semiology is compact but dense clinical language, and a successful system must interpret both the explicit wording and the implied neuroanatomy. Parameter-efficient fine-tuning likely helped the model specialize without losing general language competence, while evidence retrieval anchored responses in published surgical experience. Together, those elements reduce the risk that the model will overgeneralize from misleading phrasing or rare symptom combinations.
This is especially relevant in epilepsy, where localizing value often depends on the sequence of symptoms rather than any single sign. An aura followed by speech arrest, automatisms, or version may carry different localization implications depending on timing and laterality. A domain-adapted model can encode those patterns as probabilistic associations rather than rigid rules, which is closer to how clinicians actually reason under uncertainty.
Presurgical use, but not bedside replacement
The investigators frame EpiSemoLLM as a complement to clinical reasoning during early presurgical evaluation, not a replacement for specialist judgment. That distinction is important because seizure semiology is only one piece of the localization puzzle. Scalp EEG, imaging, neuropsychology, and sometimes intracranial recordings still shape the final decision. A model that improves the first pass through the history may still fail when the phenotype is incomplete, atypical, or confounded by medication effects, mixed seizure types, or poor recall.
The study also makes clear where automation may be most useful and where expert review remains essential. Frontal and insular localization remain more difficult for the model than for clinicians, which argues against uncritical deployment. In practical terms, an algorithm that performs well in one cortical region but not another is best used as decision support, with transparent uncertainty and direct comparison against the rest of the presurgical dataset.
External validation strengthens the signal, prospectively unanswered questions remain
The independent validation cohort from Taipei Veterans General Hospital provides an important second look beyond the development set. Replication in a separate center reduces the chance that the result is only a feature of one hospital’s documentation style or surgical referral pattern. Even so, the study remains retrospective and retrospective validation cannot establish how the model will perform inside a live presurgical conference, where incomplete histories and time pressure are routine.
Prospective validation is still needed before routine clinical use. Future work will need to test whether the model changes physician behavior, reduces disagreement, improves localization efficiency, or helps identify patients who should proceed to more invasive evaluation. It will also need to address how the system handles atypical semiology, mixed language descriptions, and real-world cases where the epileptogenic zone is not as surgically clean as in curated datasets. For now, the study shows that an epilepsy-specific language model can localize seizure narratives better than broad general models and at least as well as many clinicians in selected settings, while also exposing the cortical territories where human expertise remains harder to automate.