Development of a generative AI agent for family support in implementing family-based treatment for children and adolescents with anorexia nervosa

Introduction Family-based treatment (FBT) is a first-line psychotherapy for children and adolescents with anorexia nervosa (AN). However, families must understand the principles of FBT, provide meal support, and manage their children's pathological behaviors. Difficulties occur outside clinic hours when it is impossible to consult professionals. This “support gap” increases caregivers’ psychological distress and threatens their treatment continuity. To the best of our knowledge, this is the first domain-specific generative artificial intelligence (AI) agent designed to provide situation-specific, FBT-concordant advice and psychological support. Methods The system integrates three components: (1) an FBT-specific knowledge base constructed from treatment manuals, family guides, guideline-compliant resources, and a clinical Q&A corpus; (2) a multistage natural language processing pipeline using Retrieval-Augmented Generation (RAG), with intent and sentiment analyses; and (3) safety guardrails that prohibit unsolicited numerical goals or direct hospitalization recommendations and standardized escalation to clinicians. When strong negative emotions are detected, empowerment messages are dynamically incorporated to maintain caregivers’ confidence. Six clinicians with expertise with pediatric mental health authored queries that simulated common FBT-related concerns and evaluated each response for clinical appropriateness and safety, and classified problems as information insufficiency, not FBT concordant, or escalation insufficiency. Results Of the 477 queries, 57.0% were FBT-related, 24.5% were general AN, 16.5% were parental psychological distress, and 1.8% were related to other topics. The clinically appropriate response rate was 91.6% (437/477), including 92.3% for FBT-related questions, 88.0% for general knowledge, 93.7% for psychological distress, and 100.0% for other questions. Clinically inappropriate responses (8.4%) were mainly attributable to information insufficiency; not FBT concordant (1.8% of FBT-related responses) and escalation insufficiency (0.6% of all dialogs) rarely occurred. Discussion In this expert review, the safety-gated RAG system predominantly generated FBT-concordant responses that provided meal-level guidance and empathic empowerment-oriented support to families. By proceduralizing complex FBT concepts and presenting multiple response options for pathological behaviors, the system translates FBT principles into practical guidance supporting refeeding adherence, preserving family self-efficacy, and suggesting that domain-specific AI may help bridge structural limitations in FBT. Usability studies and randomized controlled trials are warranted to determine their impact on caregiver burden, self-efficacy, treatment adherence, and clinical outcomes.