RagMedge: Dense Retrieval Augmented Medical Dialogue Generation Towards Personalised Medical Knowledge Grounded Diagnosis

To develop a medical generative system, current research has primarily focused on integrating medical knowledge into language models (LMs) by training them on medical corpora or utilising external medical resources, such as knowledge graphs and terminologies. However, these methods only enable LMs to implicitly integrate medical information from textual narratives or standardized knowledge sources like the UMLS system, resulting in the integrated knowledge being invisible and uncontrollable during inference. This invisibility makes it challenging to track the knowledge accumulated from historical training data. To address these issues, we propose constructing a personalised medical knowledge base, which extracts discrete knowledge points from real doctor-patient dialogues and represents them as semantically condensed vectors. We also introduce a novel dense retrieval-augmented medical dialogue generation framework, called \textbfRagMedge, designed to retrieve relevant semantic embeddings of stored knowledge points that match the input context, and utilise them in generating diagnostic responses. Experiments demonstrate that RagMedge achieves state-of-the-art (SOTA) performance by effectively leveraging its personalised medical knowledge.