FC-GARNet Functional Connectivity Graph with Attention and Recurrence Network for ASD Prediction

Autism Spectrum Disorder (ASD) is a category of complex neurodevelopmental disorders that occur in early childhood. Individuals with ASD experience various challenges in social interaction and communication, and often exhibit repetitive and stereotyped behaviors. Currently, there is no reliable and effective method for accurately diagnosing the disorder. To address existing research challenges related to ASD, this paper proposes a deep learning-based position-aware graph convolutional network model. By incorporating functional connectivity matrices, positional embedding techniques, and graph sparsification, the model effectively mitigates difficulties in modeling neural connections, improves the accuracy of node matching across different brain regions, reduces the dimensionality of fMRI samples, and enhances interpretability to strengthen diagnostic persuasiveness–making it well-suited for ASD diagnosis. Using the publicly available Autism Brain Imaging Data Exchange (ABIDE) dataset, the proposed model achieves an accuracy of 75.4%, a specificity of 78.2%, and a sensitivity of 73.4% under the CC200 atlas. Furthermore, the brain regions identified as most relevant to ASD show strong alignment with established medical theoretical knowledge, suggesting potential biomarkers for clinical diagnosis of ASD.