A Trust-Calibrated Federated Learning Framework for Predicting Behavioral Escalation in Children with Autism Using Edge IoT Systems

Abstract

Behavioral escalation episodes in children with autism spectrum disorder present complex clinical, ethical, and technological challenges due to their unpredictability, individual variability, and the sensitivity of behavioral data. Artificial intelligence and Internet of Things technologies have demonstrated potential for early escalation detection; however, centralized learning architectures introduce privacy risks, governance challenges, and caregiver mistrust. This study proposes a trust-calibrated federated learning framework deployed on edge IoT systems to predict behavioral escalation while preserving data locality and embedding human trust directly into the learning process. The framework integrates reinforcement learning for temporal behavior modeling, confidence-weighted federated aggregation, and caregiver-driven trust calibration aligned with AI risk management principles. Experimental evaluation using simulated autism care scenarios demonstrates improved prediction accuracy, reduced false alerts, and measurable gains in caregiver trust compared to centralized and non-trust-aware baselines. The proposed approach advances privacy-preserving, human-centered artificial intelligence for autism care and provides a scalable pathway for ethically aligned deployment in real-world environments.