Classification of Autism Spectrum Disorder Using Edge-Weight Enhanced Graph Attention Network With Multiple Features of Resting-State fNIRS Signals.

Functional near-infrared spectroscopy (fNIRS), as a noninvasive brain imaging modality, has shown great potential for autism spectrum disorder (ASD) identification combined with machine learning. In this work, we proposed an ASD identification method using edge-weight enhanced graph attention network (EWE-GAT) with multiple features in resting-state fNIRS signals measured from the bilateral temporal lobes on 22 typically developing (TD) children and 25 children with ASD. Seven features were selected for the EWE-GAT model, including five node features: the coupling between oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (Hb) fluctuations, sample entropy for HbO and Hb, and average resting-state functional connectivity (RSFC) for HbO and Hb of each channel, and two edge features: RSFC between each channel pair for both HbO and Hb. With the proposed method, high accurate classification was achieved with 97.92% accuracy, 100% sensitivity, 96.43% precision, and 98.08% F1 score, outperforming usually used traditional machine learning and convolutional neural network models.