MULTIMODAL EEG WORLD MODEL: SELF-SUPERVISED LATENT TRANSITION LEARNING FOR WEARABLE EEG SEIZURE DETECTION
Journal:
medRxiv
Published Date:
Oct 21, 2026
Abstract
Existing supervised and self-supervised EEG models mainly learn within-segment discriminative or reconstructive representations, leaving transitions between adjacent EEG segments underexplored. We propose a self-supervised Multimodal EEG World Model for EEG seizure detection that predicts the latent representation of the next EEG segment from the current EEG state, conditioned on synchronized ECG, EMG, and movement signals. A learnable query-based fusion module aggregates the multimodal synchronized signals into a compact physiological condition. Pretraining uses no seizure labels; downstream detection retains only the frozen EEG encoder and a linear probe. On the SeizeIT2 dataset with a strict patient-wise split, the model achieved an AUPRC of 0.3748, ROC-AUC of 0.8025, and balanced accuracy of 0.7308, outperforming official database baselines (SVM with hand-crafted features and ChronoNet), an external baseline (DSAINet), and a foundation model (NeuroLM). Ablations show a 14.4% relative AUPRC gain over a random-initialized encoder, supporting multimodal-conditioned latent transition learning for wearable EEG representation learning.
