Scaffolded Turns and Logical Conversations: Designing Humanized LLM-Powered Conversational Agents for Hospital Admission Interviews
Published in CHI' 25, 2025
This project develops a more humanized LLM-powered conversational agent for hospital admission interviews that addresses the challenges of nurse shortages and time constraints. The system implements clinician-derived communication strategies through dynamic topic management with graph-based conversation flows and context-aware scaffolding using few-shot prompt tuning. Technical evaluation showed performance comparable to or exceeding human-written ground truth while outperforming prompt-engineered baselines, and a between-subject study (N=44) demonstrated significant improvements in user experience and data collection accuracy compared to existing solutions. The work contributes a framework for translating clinician expertise into algorithmic strategies for medical conversational agents, along with insights for balancing efficiency and empathy in healthcare interactions.
Recommended citation: Dingdong Liu, Yujing Zhang, Bolin Zhao, Shuai Ma, Chuhan Shi, Xiaojuan Ma
Recommended citation: Dingdong Liu, Yujing Zhang, Bolin Zhao, Shuai Ma, Chuhan Shi, and Xiaojuan Ma. 2025. Scaffolded Turns and Logical Conversations: Designing Humanized LLM-Powered Conversational Agents for Hospital Admission Interviews. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25). Association for Computing Machinery, New York, NY, USA, Article 643, 1–23. https://doi.org/10.1145/3706598.3714196
Download Paper