Electroencephalography (EEG) is currently the mostused way to accurately diagnose epilepsy given its ability tomeasure hypersinchronized periods of brain activity known asseizures. However, EEG recordings are noisy and require trainedpractitioners for meaningful information to be extracted. Mostimportantly, further post hoc analyses are inherently time-consuming and subjective. Recent advances in artificial intelli-gence have paved the way to develop automated workflows easingthe task of preprocessing and detecting epileptic activity fromEEG. Yet, these models are ubiquitously difficult to interpretthus posing a challenge for its wide acceptance in clinicalscenarios. Here, we propose a graph neural network enhancedwith attention layers able to accurately and robustly identifypathological brain activity. We provide both feature and graphexplanations for each prediction of the trained model. Crucially,we show how graph neural networks capture non-trivial depen-dencies between cortical regions that agree with the currentclinical consensus. Altogether, these results highlight the factthat explainable artificial intelligence need not compromise itsperformance and represents an improvement in the applicabilityof artificial intelligence networks in clinical practice.

Explainable graph neural networks for EEG classification and seizure detection in epileptic patients