An explainable spatio-temporal graph convolutional network for the biomarkers identification of ADHD

Attention deficit hyperactivity disorder (ADHD) is a neurodevelopmental disorder that lacks a diagnostic basis based on connection anomalies. Identifying biomarkers associated with ADHD helps to understand the underlying causes of the disorder and leads to earlier diagnosis and specific treatment. Recently, graph neural networks (GNNs) have been applied to ADHD analysis. However, the lack of spatio-temporal fusion learning of brain networks hinders the characterization of abnormal dynamic changes due to functional connectivity. Additionally, there is a lack of explainability, i.e., the ability to recognize important features that contribute to decision-making. To address these issues, we propose ESTGCN, an explainable spatio-temporal graph convolutional network that combines temporal dependencies and spatial node information to model functional brain networks and describes the dynamic functional connectivity (dFC) between brain regions. Considering explainability, the ESTGCN is constructed based on the gated recurrent unit (GRU) with spatial self-learning to better identify abnormal functional connections. The experimental results based on fMRI data indicate that ESTGCN can identify brain functional abnormalities of ADHD and achieves better performance in detecting dFC. Furthermore, abnormal functional connections are identified mainly between frontal and cingulate gyrus, frontal and parietal lobes, parietal lobes and cingulate gyrus, internal default mode network (DMN), as well as cerebellum-thalamus-prefrontal circuit.