Detection of ASD Using DySTEP-GAT Architecture With EEG-Based Attention-Fused Brain-Net Graph Representation
TL;DR
Findings indicate that the learned brain network representations capture distinct and quantifiable abnormalities in ASD-related neural dynamics, establishing the proposed framework as a robust, interpretable, and measurement-aligned tool for objective ASD assessment.
Abstract
Electroencephalography (EEG)-based assessment of autism spectrum disorder (ASD) remains challenging due to the nonstationary, high-dimensional, and heterogeneous nature of neural signals, as well as the difficulty of jointly modeling spatiotemporal dynamics and functional connectivity patterns in a unified framework. Existing approaches often rely on either single-domain features or static graph representations, limiting their ability to capture dynamic neural interactions associated with ASD. To address these limitations, this work proposes a novel EEG-based functional brain network framework with graph learning for automated ASD detection. The framework comprises two key components: 1) a dual-attention-fused Brain-Net graph representation that models EEG-derived functional connectivity and 2) a dynamic spatiotemporal edge progression graph attention (DySTEP-GAT) network for robust classification. Resting-state EEG signals acquired from multiple cortical regions are transformed into multiband functional connectivity matrices, forming refined Brain-Net graphs that encode node-level spatiotemporal characteristics and edge-level interchannel synchronization. A semantic similarity-driven dual-attention mechanism is then employed to fuse rhythm-specific connectivity information into unified functional brain network representations. The proposed DySTEP-GAT incorporates three measurement-centric modules—dynamic edge selection block (DESB), spatiotemporal GAT block (STGATB), and progressive graph embedding block (PGEB)—to effectively model nonstationary neural coupling patterns associated with ASD. Experimental validation on two publicly available EEG datasets demonstrates that the proposed framework achieves high ASD detection accuracy (96 %–97 %), along with improved precision (0.95) and $F1$ -score (0.95) compared to recent state-of-the-art methods. Ablation studies further confirm the effectiveness of attention-based multiband fusion. These findings indicate that the learned brain network representations capture distinct and quantifiable abnormalities in ASD-related neural dynamics, establishing the proposed framework as a robust, interpretable, and measurement-aligned tool for objective ASD assessment.