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

Epilepsy with its complex seizure mechanisms and diverse clinical manifestations, presents numerous challenges for clinical diagnosis and treatment, while electroencephalography (EEG) plays a crucial and irreplaceable role in its diagnosis. Although general-purpose foundation models have demonstrated some capability in knowledge processing, they still face challenges in capturing specific disease features and dealing with data scarcity in highly specialized domains such as epilepsy. To address these issues, we propose a domain-specific foundation model for epilepsy-EpilepsyFM, designed to learn generalized representations of epilepsy to support various downstream tasks. EpilepsyFM utilizes self-supervised pre-training, integrating clinical EEG data from top-tier hospital neurosurgery departments with large-scale public datasets such as TUH EEG Corpus, covering a variety of patient conditions to enhance the model's representation capacity. The model employs a discrete neural tokenizer to construct a domain-specific neural codebook for epilepsy and proposes a brain region masking strategy based on the mechanisms of clustered neuronal discharges during seizures, allowing for more effective capture of the spatiotemporal features of seizures. Furthermore, EpilepsyFM integrates temporal, spectral, and spatial encoding modules to fully exploit the multidimensional propagation patterns of epilepsy. Experimental results show that EpilepsyFM achieves state-of-the-art performance in six downstream tasks, including seizure detection, seizure type detection, short- and long-term signal forecasting, frequency-phase forecasting, anti-seizure medication efficacy analysis, and radiofrequency thermocoagulation surgery analysis, demonstrating outstanding generalization ability and broad clinical application potential.