Aligning Multimodal Patient Evidence with Biomedical Knowledge Graphs for Clinical LLMs
MM-KG’s claim is that clinical LLMs improve when patient data and biomedical knowledge are linked explicitly, not merely supplied as context.
The paper builds typed graphs connecting EHR text, imaging, genomics, and biospecimen observations to UMLS concepts and biomedical knowledge. In MIMIC-IV and ADNI tests, combining patient evidence with knowledge produced stronger results than either source alone on questions needing both. Removing the single answer-bearing relation dropped performance back to the no-knowledge baseline. The authors argue the value is in retrievable, traceable links tied to the question.
ArXiv · AI/CL/LG's note
The paper builds typed graphs connecting EHR text, imaging, genomics, and biospecimen observations to UMLS concepts and biomedical knowledge. In MIMIC-IV and ADNI tests, combining patient evidence with knowledge produced stronger results than either source alone on questions needing both. Removing the single answer-bearing relation dropped performance back to the no-knowledge baseline. The authors argue the value is in retrievable, traceable links tied to the question.
ArXiv · AI/CL/LG's note
score 4