Advancing Chinese Conversation-based Patient Guidance with a Benchmark and Knowledge-Evolvable Assistant

Abstract

Chinese Conversation-based Patient Guidance (CCPG) helps patients reach the correct hospital department through natural-language exchanges with medical staff. Despite the rapid success of large language models (LLMs) in other healthcare tasks, CCPG remains under-explored and lacks dedicated benchmarks. We address this gap with PG-Bench, the first comprehensive CCPG benchmark, spanning five subsets, 19,814 annotated dialogues, and 98 clinical departments. We evaluate 25 representative LLMs on PG-Bench and observe uniformly poor performance, even the latest models such as GPT-4 and DeepSeek-V3 fail to meet practical requirements. To close this gap, we introduce the Knowledge-Evolvable Assistant (KEA), a novel framework that augments any LLM with (i) an experience bank of validated, successful CCPG cases for analogy-based reasoning; (ii) a reflection bank that records previously misclassified cases together with their corrections and self-summarized error analyses; and (iii) an external medical knowledge base. KEA employs retrieval-augmented generation to evolve its guidance knowledge iteratively. Experiments show that KEA consistently and significantly boosts the CCPG performance of all tested LLMs on PG-Bench. However, current best results still fall short of clinical expectations, underscoring the difficulty of CCPG and the need for further research. PG-Bench and KEA together establish a rigorous foundation and strong baseline for future work on conversation-driven patient guidance in Chinese healthcare settings.

Ähnliche Arbeiten