A Screening Method for Children with Autism Spectrum Disorder Based on a Dual-Stream, Multi-Scale, Cross-Modal Attention Mechanism

Abstract

Early screening for autism spectrum disorder (ASD) currently relies on behavioral observation, which suffers from subjectivity and delays. There is an urgent clinical need for objective, quantifiable auxiliary methods. Unimodal approaches based solely on facial expressions or functional near-infrared spectroscopy (fNIRS) have inherent limitations: facial features are susceptible to situational and individual variations, while fNIRS signals are sensitive to data quality and motion artifacts, and unimodal accuracy is inherently bounded. To address these issues, this paper integrates facial expressions and fNIRS and proposes an ASD screening method based on a dual-stream, multi-scale, cross-modal attention mechanism. For unimodal processing, the facial branch uses a Multi-Scale Temporal Interaction Network (MS-TIECN) with a dual-stream BiGRU to fuse multi-scale temporal and interaction features; the fNIRS branch applies IQR-based outlier removal and fold-wise independent standardization, followed by an MLP classifier (denoted FIQ-MLP). At the multimodal level, dual-stream encoders separately encode facial and fNIRS features, followed by multi-scale fusion in the latent space using cross-modal dot-product attention, with enhanced regularization and early stopping. Evaluated on 228 scene-level samples (116 ASD scenes, 112 TD scenes) from 57 parent–child dyads across four interaction scenes, using pair-based 5-fold cross-validation, the proposed Dual-Stream Multi-Scale Cross-Modal Attention (DS-MS-CMA) model achieves an accuracy of 89.03% ± 2.77%. This outperforms the facial-only unimodal baseline (85.73%), FIQ-MLP (83.23% ± 4.07%), and late/early/simple-concatenation intermediate fusion (87.73%), with superior stability. Our model yields an ASD recall of 90.43% ± 5.07% and an AUC of 0.9181 ± 0.0608. These results show that DS-MS-CMA achieves moderate but stable performance gains for auxiliary ASD screening, with reliable robustness and preliminary clinical value.