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

Area of Science:

• Artificial Intelligence
• Developmental Pediatrics
• Machine Learning

Background:

• Early identification of Autism Spectrum Disorder (ASD) is critical for effective intervention but remains resource-intensive.
• Current screening methods often require specialized personnel and significant time.
• Scalable and accessible screening tools are needed to improve early detection rates.

Purpose of the Study:

• To evaluate a novel, two-stage multimodal Artificial Intelligence (AI) framework for scalable Autism Spectrum Disorder (ASD) screening.
• To assess the framework's accuracy in differentiating typically developing children from those at high risk or with ASD.
• To determine the framework's ability to stratify risk and support early diagnostic pathways.

Main Methods:

• A cohort of 1242 children aged 18-48 months participated.
• A mobile application collected parent-child interaction audio and screening tool data (MCHAT, SCQ-L, SRS).
• A two-stage AI model integrated multimodal data (text, audio, task success) for risk assessment, validated against ADOS-2.

Main Results:

• Stage 1 achieved an AUROC of 0.942 in differentiating typically developing from high-risk/ASD children.
• Stage 2 achieved an AUROC of 0.914 and 85.2% accuracy in distinguishing high-risk from ASD children.
• The AI model's risk predictions showed strong agreement (79.59% accuracy) and significant correlation (Pearson r=0.830) with gold-standard ADOS-2 assessments.

Conclusions:

• The developed AI framework offers a promising approach for accurate and scalable early ASD screening.
• Leveraging mobile technology and deep learning enhances the potential for early risk stratification.
• This framework can support timely interventions by facilitating earlier identification of children with ASD.