Intracranial electroencephalography (iEEG) offers high-resolution access to neural activity in both the spatial and temporal domains. Despite this resolution being well-suited to deep learning, iEEG remains underexplored in the context of large-scale supervised models and foundation models (FMs). This is mainly due to data heterogeneity, limited labelled data, and restricted accessibility. To address these challenges, we developed a foundation model based on a graph neural network with positional encoding, which we trained on a clinical dataset sourced from Utrecht University Medical Centre, using contrastive learning. Graph representations lend themselves well to the structure of intracranial electrode grids and the self-supervised approach of contrastive learning alleviates the need for labelled data. All evaluations used a binary rest-versus-move classification task over five participants. As a comparative baseline, we took a classical preprocessing approach with high frequency band features. We showed that the FM outperformed the baseline in data-scarce environments, with 10%, 30% and 50% data availability and with comparable performance at 70% data availability. This difference was confirmed with the Wilcoxon signed-rank test showing significance at
|
