Computing Hierarchical Complexity of the Brain from Electroencephalogram Signals: A Graph Convolutional Network-based Approach

Abstract

Brain structures and their varying connectivity patterns form complex networks that provide rich information to help in understanding high-order cognitive functions and their relationship with low-order sensory-motor processing. The brains with pathological conditions such as Autism Spectrum Disorder (ASD) exhibit diverse modular networks organised in hierarchies with small-world properties. However, much of the network hierarchy has not been carefully examined in ASD. Different machine learning architectures including Convolutional Neural Networks (CNN) have failed to extract related complex neuronal features and to exploit the hierarchical neural connectivity present at different electrode sites of the electroencephalogram (EEG) data. The presented work has addressed the mentioned limitations by developing a two-layered Visible-Graph Convolutional Network (VGCN) which projects each channel's EEG sample onto nodes of a graph with weighted edges formulated as per the hierarchical visibility among nodes. The proposed model has been applied to EEG signals obtained from ASD and Typical Individuals (TD) and has achieved a classification accuracy of 93.78% in comparison to state-of-the-art methods, including support vector machines (89.52%), deep neural network (78.21%), convolutional network (83.88%) and graph network (86.45%). Other performance metrics such as precision, recall, F1-score and Mathews correlation coefficient showed similar results, hence, supporting the proposed model's strengths. This evidence suggests that graph networks can confidently reveal hierarchical imbalances in the brain functioning of ASD.