Abstract
Children with developmental delays such as Autism Spectrum Disorder, Down Syndrome, and Attention-Deficit/Hyperactivity Disorder frequently exhibit problem behaviors including tantrums, aggression, and self-injury. These behaviors affect not only children’s social and developmental trajectories but also the quality of family life. While gold-standard behavioral assessments such as Applied Behavior Analysis and functional interviews remain essential, they depend heavily on subjective recall and labor-intensive human observation. This reliance on retrospective, caregiver-reported data often introduces bias and limits the timeliness of interventions. This dependence constrains timeliness, accuracy, and ecological validity. Against this backdrop, my dissertation addresses the urgent need for objective, continuous, and caregiver-centered approaches that can extend evidence-based support from clinics into everyday home contexts.To investigate this challenge, I develop and evaluate a framework of caregiver–AI collaboration, grounded in multimodal sensing and human-centered design. First, I examine which behavioral markers and contextual cues should be prioritized for automated analysis outside clinics, identifying tensions between unobtrusiveness valued by families and diagnostic fidelity emphasized by clinicians. Building on these insights, I design ATP-CLIP, a few-shot vision–language model that integrates temporal pooling with hierarchical prompts to detect problem behaviors in long-form caregiver–child interaction videos, making outputs both accurate and interpretable. I then introduce ChEMA, an AI-augmented Ecological Momentary Assessment system that fuses passive sensing with active prompts to support just-in-time intervention in real homes; a six-week deployment demonstrated feasibility and stress reduction for caregivers. Finally, I explore long-term positive behavior change through a multi-sensory therapeutic play prototype inspired by drama therapy, showing how caregiver–AI loops can shift from reactive monitoring toward proactive skill development.
Across these studies, the dissertation advances both technical and conceptual foundations for designing adaptive, context-aware systems that embed caregivers as active partners in AI-driven monitoring and intervention. By demonstrating how caregiver–AI collaboration can reduce burden, enhance interpretability, and foster both immediate responsiveness and sustained developmental gains, this work contributes a pathway toward accessible, trustworthy technologies that support families managing childhood problem behaviors.
| Date of Award | 15 Jan 2026 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Xu Sun (Supervisor), Cheng Yao (Supervisor), Xiangjian He (Supervisor) & Qingfeng Wang (Supervisor) |