Hyperbolic Relation-Aware Graph Pooling For EEG Foundation Models

Electroencephalography (EEG) is widely used in brain-computer interfaces, and large-scale pretraining has made EEG foundation models increasingly useful for downstream tasks. However, common readout heads pool pretrained tokens without explicitly considering which tokens convey similar information or which encode more specific semantic content. We propose Hyperbolic Relation-Aware Graph Pooling (HRAGP), a lightweight plug-and-play readout module that models these relations for downstream classification. HRAGP first maps EEG tokens into hyperbolic space to capture their geometric structure. Then, it constructs a weighted directed graph. Wherein, geodesic distance provides similarity-based edge weights, while entailment-cone compatibility, directional preference, and radial ordering determine which edges are retained and how they are oriented. Finally, a directional graph summarizes incoming and outgoing relations and combines the resulting representation with the average-pooled backbone features. Experiments with four EEG foundation models across five downstream benchmarks show that HRAGP improves over average pooling in most settings while adding relatively few trainable parameters.