[2025F] Winston Zeng (PhD)

TRUST: An LLM-Based Dialogue System for Trauma Understanding and Structured Assessments

Winston Zeng

Date: 2025-10-31 / 3:00 - 4:00 PM

Location: White Hall 100

Abstract

This presentation introduces TRUST (Trauma Understanding and Structured Assessments), an LLM-based dialogue system designed to support clinicians in conducting and evaluating structured diagnostic interviews for PTSD. The work addresses the growing shortage of mental health professionals and the time-intensive nature of formal diagnostic interviews. Our research contributes two key components: (1) an automated assessment pipeline that uses instruction-tuned large language models to perform PTSD diagnostics from clinician-administered interview transcripts, and (2) an interactive dialogue system capable of simulating clinician–patient interviews following established clinical protocols. Using over 700 hours of real-world diagnostic data from 336 participants, the system demonstrates strong potential for enhancing diagnostic accuracy and accessibility. We introduce a novel dialogue-act schema for clinician utterances, a scalable patient simulation framework, and a multi-layered architecture combining conversation and assessment modules with LangServe-based deployment for real-time evaluation. Early expert reviews indicate that TRUST can approximate human clinician performance while offering reproducibility and scalability across mental health applications. Future work focuses on fine-tuning emerging models (e.g., GPT-5, Llama 3.3, Qwen 2.5-VL) and integrating retrieval-augmented generation (RAG) for adaptive assessment and deployment in clinical and research contexts.

Link

Presentation