Explainable Reinforcement Learning for Caregiver Decision Support in Autism: A Human-in-the-Loop Safety Architecture

Abstract

Behavioral escalation in children with autism spectrum disorder (ASD) presents significant challenges for caregivers due to its unpredictability, individualized triggers, and potential safety implications. Artificial intelligence–based decision support systems, particularly those using reinforcement learning, have shown promise in anticipating escalation events and recommending timely interventions. However, the opacity of many reinforcement learning models and the absence of meaningful human oversight undermine caregiver trust, ethical acceptability, and real-world adoption. This study proposes an explainable reinforcement learning framework embedded within a human-in-the-loop safety architecture for caregiver decision support in autism care. The framework integrates interpretable state representations, policy-level explanation mechanisms, and caregiver override and feedback loops to ensure transparency, accountability, and shared decision authority. Using simulated autism care scenarios informed by prior empirical studies, the proposed system is evaluated on prediction accuracy, intervention appropriateness, caregiver trust, and override frequency. Results indicate that explainability-aware reinforcement learning improves caregiver confidence and decision quality while maintaining competitive predictive performance. This research advances trustworthy, human-centered artificial intelligence by operationalizing explainability and safety as core design principles rather than post hoc additions.