LLM-Based Conversation Agent for Trauma Understanding and Structured Assessments

This proposal presents TRUST (TRauma Understanding and STructured Assessments), an LLM-based dialogue framework for supporting formal PTSD assessment. PTSD diagnosis often depends on structured clinical interviews such as the CAPS-5, but these assessments are time-intensive, require specialized expertise, and are difficult to scale. TRUST addresses this gap through a dual-model architecture that combines structured diagnostic assessment with adaptive interview management. The Assessment Model will predict clinically relevant outputs, including symptom severity scores, duration and frequency information, symptom mappings, and evidence-based reasoning. The Conversation Model will guide the interview process by asking protocol-aligned, context-sensitive, and non-leading follow-up questions to fill diagnostic information gaps. This proposal will evaluate TRUST through task-specific assessment metrics, interactive dialogue analysis, clinician feedback, and a web-based pilot study. The goal is to develop a transparent, evidence-aware, and scalable framework for trauma-informed PTSD diagnostic interviewing.