Multimodal Analysis of Brain Network Abnormalities in Autism Spectrum Disorder: a Graph Neural Network-based Fusion Investigation of EEG and fNIRS.

OBJECTIVE: This study aimed to identify brain network abnormalities of children with Autism Spectrum Disorder (ASD) by analyzing multimodal neuroimaging data through graph neural networks (GNNs), laying a technical foundation for future machine learning-based neuroimaging analysis. METHODS: We employed combined electroencephalogram (EEG) and functional near-infrared spectroscopy (fNIRS) to capture brain activity of 58 ASD children and 63 age- and gender-matched typically developing (TD) children. The concurrent EEG and fNIRS data were fused using an enhanced GNN to identify ASD with high accuracy. Key connections and regions of interest (ROIs) were identified for each group, followed by graph theory metrics (e.g., local efficiency) and topological data analysis (TDA) to quantify group differences. RESULTS: The proposed GNN model achieved 84% accuracy, outperforming the single-modality approach. Analysis of subgraph created by the model showed that the two groups exhibited different neural patterns in motor-language integration and lateralization, a key basis for GNN models to distinguish between the two groups. Individuals with ASD may also have topological defects such as lower average local efficiency and a lack of high-order structures. CONCLUSION: The proposed framework effectively captures quantitative insights ranging from specific connection-level variations to network-level differences, identifying candidate discriminative connectivity patterns within this cohort. SIGNIFICANCE: This study establishes a multimodal deep learning framework that identifies candidate discriminative topological signatures from fused EEG-fNIRS data, providing preliminary within-cohort classification evidence and an exploratory topological analysis framework.