Gated edge-node interaction graph convolutional network for ADHD classification.
Authors
Affiliations (3)
Affiliations (3)
- School of Microelectronics and Control Engineering, Changzhou University, Changzhou, People's Republic of China.
- School of Medical Information and Engineering, Xuzhou Medical University, Xuzhou, People's Republic of China.
- College of Information Science and Engineering, Hohai University, Changzhou, People's Republic of China.
Abstract
<i>Objective.</i>Attention deficit hyperactivity disorder (ADHD) remains challenging to diagnose objectively and often relies on subjective clinical assessments. Although graph neural network-based methods have been widely used for ADHD identification, most existing approaches mainly focus on node features while underutilizing edge information, which limits their diagnostic performance.<i>Approach.</i>We propose a gated edge-node interaction graph convolutional network (GENI-GCN) to enhance ADHD classification using neurobiological signals. By leveraging resting-state functional magnetic resonance imaging-derived multi-band amplitudes of low-frequency fluctuations and functional connectivity data across 50 brain regions, the GENI-GCN model employs an effective co-embedding strategy for nodes and edges, fostering improved feature extraction on graphs. In detail, we design refined self-loop adjacent matrices to guide the message passing among nodes and edges, while an adaptive gated mechanism is utilized to mitigate the graph degeneration arising from excessive self-loop influence.<i>Main results.</i>The proposed GENI-GCN is further integrated within the existing binary hypothesis testing framework to form a GENI-GCN-based BHT scheme. Extensive experiments demonstrate remarkable performance, achieving accuracies of 97.9% on ADHD-200 and 96.6% on Autism Brain imaging data exchange I, surpassing current state-of-the-art methods. Moreover, gradient-based interpretability analysis reveals both discriminative brain regions and connectivities, aligning with prior ADHD findings.<i>Significance.</i>The proposed framework not only improves ADHD classification performance, but also provides interpretable evidence for identifying abnormal brain topology associated with ADHD. These findings support the interpretability of the proposed method and potential for advancing neurobiological diagnostics.