Multimodal Deep Learning Framework for Autism Spectrum Disorder Detection Using CC200 Functional Connectivity and Clinical Phenotypes

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition defined above all by its profound heterogeneity—no two individuals present quite alike, and no single biological marker has proven sufficient for reliable automated diagnosis across diverse populations. Most existing computational approaches tackle this challenge through single-modality pipelines, analyzing resting-state functional neuroimaging or clinical demographic data in isolation and, in doing so, discard the complementary diagnostic signal available from the other source. A dual-branch deep learning architecture is introduced that processes both data streams in parallel through a late-fusion strategy. One branch ingests functional connectivity features derived from resting-state fMRI using the Craddock 200 (CC200) atlas, compressed to 42 principal components via Principal Component Analysis (PCA). A second branch simultaneously processes twelve standardized clinical phenotypic variables encompassing age, biological sex, and full-scale IQ, among others. Each branch independently learns modality-specific latent representations, which are then concatenated and forwarded to a unified classification head. The model was trained and evaluated on 1,114 participants from the ABIDE- II repository under rigorous five-fold stratified cross-validation. The complete multimodal system achieved a mean accuracy of 94.17% ( ±  0.71%), AUC-ROC of 0.9704, precision of 96.84%, and specificity of 97.48%. Systematic ablation experiments reveal that while the clinical branch alone reaches 94.62% accuracy—a result interrogated carefully through dedicated confound and bias analysis—the full multimodal model delivers more consistent cross-fold performance and markedly reduced false-positive rates. This confirms that the neuroimaging branch contributes genuine physiological regularization beyond what demographics alone can provide. Findings are contextualized through comparison against six baseline systems, including logistic regression, support vector machines, random forests, single- branch MLPs, and an early-fusion variant. Interpretability is addressed through SHAP-based feature attribution, a principled justification for PCA component selection is provided through explained variance profiling, and an honest assessment of limitations arising from the absence of external validation is offered. The goal is a diagnostic framework that is not only high-performing, but transparent and reproducible.