📝 SummarXiv

Abstract

Large Language Models (LLMs) perform strongly in static and single-turn medical Question Answer (QA) benchmarks, yet such settings diverge from the iterative information gathering process required in practical clinical consultations. The MEDIQ framework addresses this mismatch by recasting the diagnosis as an interactive dialogue between a patient and an expert system, but the reliability of LLMs drops dramatically when forced to reason with dialogue logs, where clinical facts appear in sentences without clear links. To bridge this gap, we introduce TriMediQ, a triplet-structured approach that summarises patient responses into triplets and integrates them into a Knowledge Graph (KG), enabling multi-hop reasoning. We introduce a frozen triplet generator that extracts clinically relevant triplets, using prompts designed to ensure factual consistency. In parallel, a trainable projection module, comprising a graph encoder and a projector, captures relational information from the KG to enhance expert reasoning. TriMediQ operates in two steps: (i) the projection module fine-tuning with all LLM weights frozen; and (ii) using the fine-tuned module to guide multi-hop reasoning during inference. We evaluate TriMediQ on two interactive QA benchmarks, showing that it achieves up to 10.4\% improvement in accuracy over five baselines on the iMedQA dataset. These results demonstrate that converting patient responses into structured triplet-based graphs enables more accurate clinical reasoning in multi-turn settings, providing a solution for the deployment of LLM-based medical assistants.