An Ensemble Graph-Based Deep Learning Framework Integrating Phenotypic and fMRI Data for Autism Classification

Autism spectrum disorder (ASD) affects approximately$\mathbf{1}-\mathbf{2. 6 \%}$of the pediatric population globally. The present study introduces an innovative methodology for the classification of autism by incorporating resting-state functional connectivity (FC) data with phenotypic data. Contrasting with conventional methodologies that treat these two data modalities in isolation, we employ machine learning classifiers on phenotypic data and implement a graph-based deep learning technique on dense FC matrices derived from resting-state functional magnetic resonance imaging (fMRI) data. To mitigate the noise and address the high dimensionality inherent in these dense FC matrices, sparse FC matrices are estimated utilizing a log-regularized Laplacian Sparse Functional Connectivity (LSFC) method. Subsequently, an Adaptive-Hierarchical Graph Convolutional Network (AH-GCN) model is used to generate similarity- and weight-adaptive graphs from these sparse matrices, employing hierarchical clustering to augment graph discriminative features further. Finally, an ensemble stacking model synthesizes AH-GCN outputs with three machine learning classifiers. Empirical evaluations using the ABIDE I dataset indicate that this ensemble approach enhances classification accuracy relative to individual classifiers. Also, the AH-GCN ensemble achieves a superior performance-cost tradeoff with lower time and memory usage compared to existing graph-based models. Lastly, an interpretability analysis verifies our model as an individual graph model capable of visualizing salient brain regions pertinent to ASD classification.