AGCN: Adaptive Graph Convolution Network with Hemibrain Differences of Resting-State EEG for Identifying Autism in Children

Early identification of autism in children has become a great concern due to the increasing prevalence of autism. Recently, EEG has become a promising tool to identify autistic children with some advanced signal processing and machine learning methods. However, it is still challenging to achieve accurate and reliable result due to the complexity of EEG signals. To this end, we propose a new adaptive graph neural network model by incorporating hemibrain differences (AGCN) of resting-state EEG to classify autistic children and typical developing (TD) children. First, a temporal feature learning module is designed to learn temporal dynamics of EEG signals. Then, to capture spatial correlations, we design an adaptive graph convolution network combined with spatial attention mechanism to learn the connectivity between nodes from the perspective of developmental abnormalities in the hemiencephalar and hemicerebral of autism. We collected a EEG dataset containing 45 ASD children and 45 typically developing (TD) children to evaluate our proposed method. Experimental results shown that our proposed method is superior to several baseline methods.