Abstract
Graph neural networks (GNNs) and the attention mechanism are two of the most significant advances in artificial intelligence methods over the past few years. The former are neural networks able to process graph-structured data, while the latter learns to selectively focus on those parts of the input that are more relevant for the task at hand. In this paper, we propose a methodology for seizure localisation which combines the two approaches.
Our method is composed of several blocks. First, we represent brain states in a compact way by computing functional networks from intracranial electroencephalography recordings, using metrics to quantify the coupling between the activity of different brain areas. Then, we train a GNN to correctly distinguish between functional networks associated with interictal and ictal phases. The GNN is equipped with an attention-based layer which 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 is fully unsupervised, meaning that it does not use any prior information regarding the seizure onset zone.
We report results both for human patients and for simulators of brain activity. 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 also show that our GNN exhibits uncertainty on those patients for which the clinical localisation was also unsuccessful, highlighting the robustness of the proposed approach.
Competing Interest Statement
The authors have declared no competing interest.
