A two-stage proactive dialogue generator for efficient clinical information collection using large language model

Xueshen Li, Xinlong Hou, Nirupama Ravi, Ziyi Huang, Yu Gan

Abstract: Highlights•We propose a proactive dialogue system with a critical doctor agent model to automatically collect diagnostic information (e.g., symptoms, test results). Unlike QA tasks, our doctor agent actively asks questions to guide information flow.•We develop a two-stage recommendation framework with query generation and candidate ranking to enhance diagnostic query flexibility and depth. Using real-world history, we design a ranking strategy to boost reasoning accuracy.•We conduct comprehensive experiments on real medical dialogue datasets. Our model better mimics real doctors in professionalism, safety, and style. It also gathers diagnostic data effectively in multi-round clinical scenarios.

External IDs:dblp:journals/eswa/LiHRHG25