Graphical attention networks for autism spectrum disorder classification

Autism Spectrum Disorder (ASD) is consider as a complex neurodevelopmental challenge, which is often diagnosed later than desirable, resulting in delayed interventions. But identifying it as early as possible is very important for the patient to receive proper intervention to improve their condition. The purpose of this research is to identify ASD in its earliest stage. In this work, Electroencephalogram (EEG) signals are preprocessed using Discrete Wavelet Transform and Fast Fourier Transform. Correlation graph method is utilized for constructing the EEG graph from both preprocessed signal. To capture spatiotemporal dependency of data, attention-based Graph Neural Network (GNN) are proposed, which incorporate Diffusion Convolutional Recurrent Neural Network and Gated Recurrent Units. Additionally, the model integrates Graph Attention Mechanism to capture significant nodes within the graph signals. As a result, the proposed model with Fast Fourier transform preprocessed data achieves 96.2% accuracy, which is better than the model with Discrete Wavelet Transform accuracy of 90.4%. GNN and attention mechanisms offer a promising avenue for early identification of autism, even with limited data.