Foundation Models for Epileptogenic Zone Identification in Drug-Resistant Epilepsy

Researchers developed EpiiSLM, a dual foundation model system that significantly improves identification of epileptogenic zones in drug-resistant epilepsy patients using stereo-electroencephalography data. The system achieved 97.8% contact-level accuracy and requires only one night of monitoring, potentially reducing invasive procedures and improving surgical outcomes where current seizure freedom rates remain below 50%.

EpiiSLM represents a meaningful advancement in clinical neurology by applying modern machine learning architectures to a persistent medical challenge. Drug-resistant epilepsy affects approximately 30% of epilepsy patients, and surgical intervention offers the best path to seizure freedom, yet current methods struggle to precisely locate the epileptogenic zone, resulting in failed surgeries and continued seizures. The development of this dual foundation model system addresses this gap by combining specialized signal processing with multimodal clinical integration.

The technical approach demonstrates sophisticated engineering: a signal foundation model trained on over 100,000 minutes of real-world sEEG recordings learns to identify epileptogenic biomarkers by anchoring on non-epileptic signals as reference points, avoiding bias toward seizure-specific patterns. A language foundation model then synthesizes these signals with clinical metadata to produce interpretable predictions. This architecture reflects broader trends in AI healthcare applications, where foundation models trained on large, diverse datasets increasingly outperform narrowly specialized systems.

The clinical implications are substantial. Reducing monitoring duration from multiple nights to a single night of interictal sleep data could decrease patient morbidity, infection risk, and hospitalization costs while enabling faster surgical planning. External validation showing 85.7% contact-level positive predictive value suggests genuine generalization capability, though the performance gap between internal (97.8%) and external validation warrants careful consideration in clinical deployment.

Success in this domain could accelerate foundation model adoption across neurosurgery and neurology, establishing new standards for AI-assisted surgical planning. Future developments should focus on prospective clinical trials and integration with existing surgical navigation systems.

• →EpiiSLM achieves 97.8% contact-level accuracy in identifying epileptogenic zones, outperforming baseline methods by 15.1%
• →The system requires only one night of interictal sleep monitoring instead of multiple nights, reducing patient burden and procedural risk
• →Dual foundation model architecture combines signal processing with multimodal clinical data integration for interpretable predictions
• →External validation shows 85.7% accuracy, indicating genuine generalization beyond the training dataset
• →Success could accelerate AI adoption in neurosurgical planning and establish new clinical standards for drug-resistant epilepsy treatment