Multi-Modal Graph Construction and Classification for Autism Spectrum Disorder using Stacked GCNs

Multimodal data are essential to improve the precision and effectiveness of diagnosing neurological conditions such as Autism Spectrum Disorder (ASD). This study leverages functional magnetic resonance imaging (fMRI) time series data, processed using Harvard-Oxford (HO) and Automated Anatomical Labeling (AAL) atlases, alongside nonimaging demographic features (e.g., gender, site), to improve ASD classification in complex clinical settings. We propose a multimodal framework that begins with recursive feature elimination (RFE), tailored for ASD tasks involving 1D fMRI time series, to isolate the most significant features contributing to class distinction and reduce dimensionality while preserving critical temporal patterns. The selected attributes are used to formulate a functional connectivity graph of the whole brain. This graph is processed through a hybrid architecture: a graph convolutional network (GCN) with residual connections to mitigate information loss and stabilize gradients, and a Stacked Graph Convolutional Network (GCN) that leverages random walk embeddings to capture higher-order structural relationships. The outputs of the DeepGCN with residual connections and the Stacked GCN are combined and subsequently fed into a Multi-Layer Perceptron (MLP) for binary classification of ASD versus typical controls.