DANCE: Domain-Aware Network-Coupling Encoder for Coordinate-Free Intracranial EEG Representation Learning

The brain is a network of interacting regions whose coupling changes with behavioural state. Existing foundation models for intracranial EEG (iEEG) either ignore this structure or encode the static spatial location of each electrode, leaving the interactions between electrodes to be learned implicitly. We propose DANCE, a Domain-Aware Network-Coupling Encoder, which places functional connectivity in the pretraining objective: a channel-agnostic encoder is trained to align its embeddings with measured connectivity, estimated over short intervals to reflect the dynamic coupling of the network as brain state changes. We evaluate DANCE on two distinct datasets. On Brain Treebank, a public language-decoding benchmark, DANCE matches the strongest published model, improves speech-decoding ROC-AUC by 16% on the subjects with the most electrodes and the fewest labels, and loses no performance when the evaluated patient is excluded from pretraining. On a chronic ambulatory cohort recorded by an implanted device from four channels, DANCE outperforms state-of-the-art models adapted to the cohort, reaching 0.845 macro-F1 on sleep stage classification and 0.967 ROC-AUC on seizure detection. Strong performance on two datasets that differ in recording setting, electrode configuration and task demonstrates the broad applicability of DANCE and its potential for wider use in the analysis of brain activity.