Safety-Constrained Agentic AI for Autism Screening: A Multimodal, Clinician-Guided Architecture

Area of Science:

• Artificial Intelligence in Healthcare
• Developmental Pediatrics
• Clinical Decision Support Systems

Background:

• Autism spectrum disorder (ASD) diagnosis faces significant delays due to specialist shortages and access issues.
• Delayed diagnosis hinders early intervention, crucial for improving developmental outcomes in children with ASD.
• Current AI adoption in ASD screening is limited by concerns regarding reliability, governance, bias, and clinical trust.

Purpose of the Study:

• To propose a conceptual, governance-driven, clinician-augmented AI framework for ASD screening.
• To design an agentic system prioritizing governance and safety through specialized agents.
• To assist, not replace, clinicians in the ASD screening process.

Main Methods:

• Developed a multi-agent AI framework collecting multimodal data (text, audio, video).
• Integrated specialized agents for consent validation, bias monitoring, model selection, and confidence-based abstention.
• Focused on validating the governance and multi-agent system through Stage 1 testing with over a hundred scenarios.

Main Results:

• Successfully designed and validated a conceptual governance-driven agentic framework at Stage 1.
• Demonstrated a system architecture for managing governance and safety rules via specialized agents.
• Confirmed the framework's potential to provide structured decision-support information to clinicians.

Conclusions:

• The proposed framework offers a novel approach to integrating AI into ASD screening, emphasizing governance and clinician augmentation.
• Stage 1 validation confirms the feasibility of the agentic system for managing safety and governance.
• Future work will focus on clinical data collection, model training, multimodal integration, and real-world validation for a complete clinical solution.