Privacy-by-Design in AI-Assisted Systems for Caregivers of Children with Autism: A Secure Multi-Agent Architecture

Caregivers of children with Autism Spectrum Disorder (ASD) frequently experience chronic psychological stress, thereby necessitating accessible support. Although artificial intelligence (AI)-based assisted technologies have the potential to reduce caregiver workload, most existing solutions lack robust privacy control and clinical interoperability, which significantly limits their adoption in regulated healthcare environments. To address these challenges, this paper proposes a Privacy-by-Design (PbD) multi-agent architecture that enables consent-aware, auditable, and privacy-preserving AI-assisted support for caregivers of children with ASD. The effectiveness of the proposed architecture was evaluated using two datasets: one focusing on clinically grounded autism-related knowledge and another reflecting naturalistic caregiver observation language. System performance was assessed using a Retrieval-Augmented Generation Assessment (RAGAs)-based framework with a Large Language Model (LLM)-as-a-Judge approach implemented via a locally deployed Llama 3 8B model. The system achieved answer relevancy scores of 0.767 for the clinical dataset and 0.750 for the observational dataset, with corresponding Recall@K values of 0.400 and 0.742, respectively. Context precision ranged from 0.599 to 0.631, and no harmful content was detected. Overall, the proposed architecture demonstrates secure caregiver–specialist collaboration through consent-aware routing, anonymised data storage, and controlled data reconstruction, providing a regulation-aligned design option for privacy-preserving AI integration in assisted care platforms.