Autism Spectrum Disorder Classification with EEG Signals Using Dense Graph Convolution Neural Network Based on Brain Regions

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by variations in brain function and structure. This study introduces an innovative approach to predict ASD using electroencephalography (EEG) data targeted at specific brain regions. By employing a Densely Connected Graph Convolutional Neural Network (DenseGCN), we aim to classify EEG signals as either ‘autism’ or ‘control’. The methodology includes extracting and normalizing wavelet, spectral, and autoregressive features from EEG signals, followed by constructing adjacency matrices using Minimum Spanning Tree (MST) based on Euclidean distances. These matrices capture connectivity among EEG channels and serve as the basis for training graph-based neural networks. With an average accuracy of 98% and high precision and recall across diverse brain regions, the DenseGCN demonstrates its potential to advance ASD diagnosis and treatment strategies, including personalized neurorehabilitation.