An Interpretable Graph Based Model for Classification Of EEG using Directional Functional Connectivity

In this study, we introduce a method for classification of Electroencephalography (EEG) signals. Our focus is on deep learning methods to study how connectivity patterns can be integrated to classify the underlying EEG signals. We use Graph Neural Networks model to classify the EEG signals, and additionally explore the usage of Granger causality as a metric for directed connectivity. We show that directed connectivity increases the performance of predictions on several benchmark datasets. Our method achieves approximately 4.7% and 4% increases in accuracy than the best achieving models on two different benchmark datasets. Additionally, the importance of the connectivity edges in the prediction of the model is studied through saliency and integrated gradient maps to interpret the model locally and globally.