Neuroimaging-Based Autism Spectrum Disorder Risk Screening Using Graph-Based Modeling of Resting-State Functional Connectivity

Abstract

Neuroimaging-based artificial intelligence for ASD risk screening holds considerable promise as an objective, scalable complement to behavioral assessment; however, most existing studies are limited by poor generalizability, absence of external validation, and inadequate model interpretability. This study aims to address these limitations by developing a robust, interpretable, and externally validated neuroimaging-only framework for ASD risk screening using restingstate functional magnetic resonance imaging (rs-fMRI) and graph neural networks. Functional connectivity matrices were derived from rs-fMRI data and modelled as undirected brain graphs, in which each region of interest (ROI) is a node and pairwise connectivity strength defines the edge weights. A graph-based connectome encoder with an attention-aware pooling layer was employed to generate subject-specific neurobiological representations, followed by a calibrated classification layer producing probabilistic ASD risk scores. Brain parcellation was performed using the AAL-116 atlas (116 ROIs). Model training and internal validation followed a nested crossvalidation protocol on the ABIDE I dataset (871 subjects after quality control; 403 ASD, 468 TD), with independent external validation on ABIDE II (834 subjects; 378 ASD, 456 TD) without retraining or parameter adjustments. The proposed framework achieved an AUROC of 0.83 (95% CI: 0.80–0.86) on ABIDE I and 0.79 (95% CI: 0.76–0.82) on ABIDE II, outperforming all baseline models including SVM, MLP, and standard GNN architectures. The Expected Calibration Error (ECE) was 0.048 on ABIDE I and 0.061 on ABIDE II, indicating reliable probabilistic risk estimates. Site-wise robustness analysis demonstrated low AUROC variance (SD = 0.018) across acquisition sites, confirming generalizability across heterogeneous multi-site data. Attention-based attribution identified discriminative connectivity patterns within the Default Mode Network, Social Brain Network, and Salience Network, consistent with established ASD neurobiological findings. The minimal performance gap between ABIDE I and ABIDE II validation results demonstrates that the graph-based connectome encoding approach captures intrinsic neurobiological signal rather than dataset-specific artefact. The integration of GNNExplainer-based attribution maps bridges the gap between predictive performance and neurobiological interpretability, addressing a key barrier to clinical adoption of deep learning models in neurodevelopmental screening. Importantly, ABIDE datasets are retrospective and research-grade; the reported metrics reflect statistical generalization across research cohorts and do not constitute clinical diagnostic validation, which would require prospective evaluation in real-world clinical populations. This study presents a robust, interpretable, and externally validated neuroimaging-only framework for ASD risk screening that demonstrates strong generalization across multi-site rs-fMRI datasets. The framework provides calibrated probabilistic risk estimates alongside neurobiologically grounded explanations, positioning it as a promising computational support tool for clinicians. Prospective validation in clinically recruited, early-childhood populations remains a necessary step toward translational deployment.