Explainable Graph Neural Networks for EEG Classification and Seizure Detection in Epileptic Patients

Electroencephalography (EEG) is currently the most used way to accurately diagnose epilepsy given its ability to measure hypersinchronized periods of brain activity known as seizures. However, EEG recordings are noisy and require trained practitioners for meaningful information to be extracted. Most importantly, further post hoc analyses are inherently time-consuming and subjective. Recent advances in artificial intelligence have paved the way to develop automated workflows easing the task of preprocessing and detecting epileptic activity from EEG. Yet, these models are ubiquitously difficult to interpret thus posing a challenge for its wide acceptance in clinical scenarios. Here, we propose a graph neural network enhanced with attention layers able to accurately and robustly identify pathological brain activity. We provide both feature and graph explanations for each prediction of the trained model. Crucially, we show how graph neural networks capture non-trivial dependencies between cortical regions that agree with the current clinical consensus. Altogether, these results highlight the fact that explainable artificial intelligence need not compromise its performance and represent an improvement in the applicability of artificial intelligence networks in clinical practice.