RECTOR Advances EEG/sEEG Representation Learning for Affective, Cognitive Disorders

This research introduces RECTOR, a novel self-supervised framework designed to learn robust representations from electroencephalography (EEG) and stereo-EEG (sEEG) data. The goal is to improve the diagnosis of affective and cognitive disorders, which manifest as complex, time-varying brain network dynamics across different regions, channels, and time. RECTOR-SA, a core component, employs hierarchical, block-sparse self-attention, which is guided by Adaptive Functional Partitioning. This allows the model to evolve region structures from static anatomical definitions to more adaptive, functional regions. The self-supervision mechanism is driven by Masked Topology and Representation Learning, which optimizes three objectives: Masked Predictive Modeling, Topological Structure Modeling, and Cross-View Consistency. Empirical results show RECTOR achieving state-of-the-art performance in EEG emotion recognition and sEEG task-engagement classification, demonstrating strong robustness to missing channels and cross-montage generalization, which is crucial for large-scale pre-training on diverse EEG/sEEG datasets.