Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by challenges in communication, social interaction, and behavioral patterns. While early diagnosis remains crucial, there is a critical need for effective screening methods for school-age children, as many cases with nuanced presentations often go undetected during early childhood. This research presents a facial morphology-based computational screening framework specifically targeting school-age children, developed and evaluated primarily on a specialized dataset collected via webcam during Virtual Reality Continuous Performance Test (VR-CPT) attention assessment tasks. The framework introduces two dual-branch architectures. The Hybrid Feature Fusion Network (HFFN) combines handcrafted SIFT descriptors with deep neural representations through late fusion, while the Geometric-Deep Feature Network (GDFN) integrates clinically-derived facial landmark distances with learned features. On the attention assessment dataset, GDFN achieved the strongest performance with 96.33% accuracy. To assess generalizability, GDFN was additionally evaluated on the Autism Facial Image dataset, used as a secondary benchmark, attaining 89.60% accuracy under cross-validation. Comprehensive explainability analysis reveals distinct attention patterns between ASD and typically developing children, providing insights into the model’s decision-making process. The proposed framework is intended as a preliminary screening aid to flag children who may benefit from comprehensive clinical evaluation, rather than as a standalone diagnostic tool. Requiring only facial images, the approach is low-cost and feasible for integration as a preliminary screening step within educational settings where digital infrastructure already exists, potentially supporting timely referral and improved outcomes for school-age children.