Foundation Models for Epileptogenic Zone Identification in Drug-Resistant Epilepsy

Identifying the epileptogenic zone (EZ)—the specific brain region responsible for generating seizures—is the primary goal of resective surgery for drug-resistant epilepsy. However, current methods often rely on the seizure onset zone (SOZ) or surgical resection boundaries, which are imprecise and contribute to low long-term seizure-freedom rates. This paper introduces EpiiSLM, a dual-foundation model system designed to improve the accuracy and interpretability of EZ localization using stereo-electroencephalography (sEEG).

EpiiSLM consists of two integrated components: a Signal Foundation Model (SFM) and a Language Foundation Model (LFM). The SFM is trained on over 100,000 minutes of sEEG data using unsupervised masked similarity learning, allowing it to extract robust biomarkers from interictal (non-seizure) recordings. By anchoring the model on a well-defined "non-epileptic" class—contacts that were not resected in patients who achieved long-term seizure freedom—the researchers avoid the ambiguity of traditional positive-label training. The LFM then synthesizes these signal-derived probabilities with clinical data (such as MRI findings and SOZ locations) to provide final, interpretable predictions.

EpiiSLM significantly outperforms the standard SOZ-as-EZ baseline. In leave-one-patient-out evaluations, the model achieved a contact-level positive predictive value (PPV) of 0.978 when allowing for a small spatial margin, representing a 15.1% improvement over the baseline. The model also demonstrated strong performance on an external validation cohort. Crucially, because the SFM learns from interictal sleep data, the system could potentially reduce the duration of invasive sEEG monitoring from weeks to a single overnight session, minimizing patient risk and hospital burden.