EpilepsyFM: A domain-specific foundation model for epileptic representation learning using EEG signals.

Area of Science:

• * Neurology
• * Artificial Intelligence
• * Medical Informatics

Background:

• * Epilepsy diagnosis and treatment present significant challenges due to complex seizure mechanisms and diverse clinical presentations.
• * Electroencephalography (EEG) is indispensable for epilepsy diagnosis, yet general AI models struggle with specialized data and domain-specific features.
• * Existing AI approaches face limitations in capturing the nuanced characteristics of epilepsy from EEG data.

Purpose of the Study:

• * To introduce EpilepsyFM, a domain-specific foundation model tailored for epilepsy research and clinical applications.
• * To enhance the representation capacity of AI models for epilepsy by integrating diverse EEG datasets.
• * To improve the accuracy and applicability of AI in diagnosing and managing epilepsy.

Main Methods:

• * Developed EpilepsyFM using self-supervised pre-training on extensive clinical EEG data and public datasets (e.g., TUH EEG Corpus).
• * Employed a discrete neural tokenizer to create a domain-specific neural codebook for epilepsy.
• * Incorporated a novel brain region masking strategy and integrated temporal, spectral, and spatial encoding modules to capture spatiotemporal seizure features.

Main Results:

• * EpilepsyFM achieved state-of-the-art performance across six diverse downstream tasks.
• * Demonstrated superior generalization ability in seizure detection, seizure type classification, and signal forecasting.
• * Showcased effectiveness in analyzing anti-seizure medication efficacy and surgical outcomes.

Conclusions:

• * EpilepsyFM represents a significant advancement in AI for epilepsy, offering robust performance and broad clinical utility.
• * The domain-specific approach effectively addresses limitations of general AI models in specialized medical fields.
• * This model holds substantial potential for improving epilepsy diagnosis, treatment monitoring, and personalized medicine.