Abstract
Autism diagnosis lacks objective context-specific neurobiological markers, as traditional structural and resting-state neuroimaging fails to capture the dynamic social-cognitive processing differences that define the condition. We developed DualPathNet, an interpretable dual-stream deep learning network that simultaneously captures stable trait-like patterns and transient event-specific neural responses during naturalistic movie viewing. Across 555 children (274 ASD, 281 controls), our framework achieved > 70% accuracy using only 2-3 minutes of emotionally challenging stimuli, substantially outperforming 63% resting-state scans. Explainable AI with DualPathNet revealed that autism-related neural signatures were selectively expressed during high-demand social-emotional moments requiring empathy and emotion regulation, rather than uniformly expressed across all contexts. Critically, temporally specific neural responses during emotionally salient events predicted core autism symptoms including repetitive behaviors and social deficits. Our neuro-AI approach demonstrates that autism involves dynamic, context-dependent neural vulnerabilities rather than static disruptions, providing interpretable biomarkers for precision diagnosis and targeted intervention.