ASD-HI: A Parent-Child Interaction Dataset for Automated Assessment of Home Intervention

Home-based interventions are vital for supporting young children with autism spectrum disorder (ASD), yet many parents struggle to implement strategies effectively due to limited training. While specialists such as educators and speech-language pathologists provide guidance, real-time feedback outside professional settings remains scarce. To bridge this gap, we leverage advances in AI to support parents through automated assessment. However, training such AI systems requires robust data, which is currently limited. To address this, we created ASD-HI (Autism Spectrum Disorder - Home Intervention), a multi-modal dataset comparing 473 real parent-child interaction videos across three families. ASD-HI supports two core tasks: 1) Strategy Detection, identifying the behavioral strategies parents use, and 2) Fidelity Assessment, assessing the fidelity with which these strategies are implemented. We also propose a prompting-based LLM pipeline as a reference approach. It achieves 74% recall and 50% precision for strategy detection and 60% accuracy for fidelity assessment. Our work lays a foundation for developing AI-driven tools to enhance home interventions and improve outcomes for children with special needs.