Evaluating Functional-Connectivity Representations for Cross-Site Autism Spectrum Disorder Classification Under Site-Held-Out Validation

This study proposes a leakage-controlled multi-atlas framework for cross-site ASD classification using resting-state fMRI. It systematically compares static Pearson, mean-pooled dynamic, and tangent-space functional connectivity across BASC-118 and CC200 under strict leave-one-site-out validation. A lightweight six-branch ensemble integrates these complementary representations using equal-weight probability averaging. The study also demonstrates how site-mixed validation and diagnosis-informed preprocessing can substantially inflate reported performance, highlighting the importance of rigorous site-held-out evaluation for reliable cross-site generalisation.