Multimodal AI for risk stratification in autism spectrum disorder: integrating voice and screening tools

Abstract

Early Autism Spectrum Disorder (ASD) identification is crucial but resource-intensive. This study evaluated a novel two-stage multimodal AI framework for scalable ASD screening using data from 1242 children (18-48 months). A mobile application collected parent-child interaction audio and screening tool data (MCHAT, SCQ-L, SRS). Stage 1 differentiated typically developing from high-risk/ASD children, integrating MCHAT/SCQ-L text with audio features (AUROC 0.942). Stage 2 distinguished high-risk from ASD children by combining task success data with SRS text (AUROC 0.914, Accuracy 0.852). The model's predicted risk categories strongly agreed with gold-standard ADOS-2 assessments (79.59% accuracy) and correlated significantly (Pearson r = 0.830, p < 0.001). Leveraging mobile data and deep learning, this framework demonstrates potential for accurate, scalable early ASD screening and risk stratification, supporting timely interventions.