Seizure localisation with attention-based graph neural networks.

In this paper, we introduce a machine learning methodology for localising the seizure onset zone in subjects with epilepsy. We represent brain states as functional networks obtained from intracranial electroencephalography recordings, using correlation and the phase-locking value to quantify the coupling between different brain areas. Our method is based on graph neural networks (GNNs) and the attention mechanism, two of the most significant advances in artificial intelligence in recent years. Specifically, we train a GNN to distinguish between functional networks associated with interictal and ictal phases. The GNN is equipped with an attention-based layer that automatically learns to identify those regions of the brain (associated with individual electrodes) that are most important for a correct classification. The localisation of these regions does not require any prior information regarding the seizure onset zone. We show that the regions of interest identified by the GNN strongly correlate with the localisation of the seizure onset zone reported by electroencephalographers. We report results both for human patients and for simulators of brain activity. We also show that our GNN exhibits uncertainty for those patients for which the clinical localisation was unsuccessful, highlighting the robustness of the proposed approach. • We present a new method for seizure localisation in subjects with epilepsy. • Our method is based on graph neural networks (GNNs) with an attentional layer. • The attention scores of the GNN predict the clinically-identified seizure onset zones. • In uncertain clinical outcomes, our GNN also shows uncertainty in the localisation. • Our method works well even in the case of small training datasets available. [ABSTRACT FROM AUTHOR]