M4CEA: A Knowledge-guided Foundation Model for Childhood Epilepsy Analysis.

Existing electroencephalogram (EEG)-based deep learning models are mainly designed for single or several specific tasks in childhood epilepsy analysis, which limits the perceptual capabilities and generalisability of the model. Recently, Foundation Models (FMs) achieved significant success in medical analysis, motivating us to explore the capability of FMs in childhood epilepsy analysis. The objective is to construct a FM with strong generalization capability on multi-tasking childhood epilepsy analysis. To this end, we propose a knowledge-guided foundation model for childhood epilepsy analysis (M4CEA) in this paper. The main contributions of the M4CEA are using the knowledge-guided mask strategy and the temporal embedding of the temporal encoder, which allow the model to effectively capture multi-domain representations of childhood EEG signals. Through pre-training on an EEG dataset with more than 1,000 hours childhood EEG recording, and performance fine-tuning, the developed M4CEA model can achieve promising performance on 8 downstream tasks in childhood epilepsy analysis, including artifact detection, onset detection, seizure type classification, childhood epilepsy syndrome classification, hypoxic-ischaemic encephalopathy (HIE) grading, sleep stage classification, epileptiform activity detection and spike-wave index (SWI) quantification. Taking HUH (Helsinki University Hospital) seizure detection task as an example, our model shows 9.42% improvement over LaBraM (a state-of-the-art Large Brain foundation Model for EEG analysis) in Balanced Accuracy. The source code and pre-trained weight are available at: https://github.com/Evigouse/M4CEA Project.