Use of some variants of Graph Neural Networks for the classification of Electroencephalograms of schizophrenia

Graph Neural Networks (GNN) are a form of neural network that have shown their addedvalue on non-Euclidean data such as electroencephalograms (EEG). The aim of this work was to evaluatesome GNN variants for the classification of schizophrenia EEGs; namely Graph Convolution Network(GCN) and Graph of Attention (GAT). For this purpose, the Pycaret tool and a Convolutional NeuralNetwork (CNN) were used for the classification of the obtained graphs. This made it possible to compareGCN and GAT associated with Pycaret on the one hand, and GCN and GCN coupled with a ConvolutionalNeural Network (CNN) for classification on the other. It was found that GCN and GAT offer accuracyof 85% and 80% respectively. GCN coupled with a CNN offers 80% accuracy. However, despite the factthat GAT is resource intensive, its confusion matrix shows that it offers better sensitivity and specificitythan the other methods. In other words, the error rate is 5%.