Large Language Models and Medical Knowledge Grounding for Diagnosis Prediction

Abstract

AbstractWhile Large Language Models (LLMs) have showcased their potential in diverse language tasks, their application in the healthcare arena needs to ensure the minimization of diagnostic errors and the prevention of patient harm. A Medical Knowledge Graph (KG) houses a wealth of structured medical concept relations sourced from authoritative references, such as UMLS, making it a valuable resource to ground LLMs’ diagnostic process in knowledge. In this paper, we examine the synergistic potential of LLMs and medical KG in predicting diagnoses given electronic health records (EHR), under the framework of Retrieval-augmented generation (RAG). We proposed a novel graph model: Dr.Knows, that selects the most relevant pathology knowledge paths based on the medical problem descriptions. In order to evaluate Dr.Knows, we developed the first comprehensive human evaluation approach to assess the performance of LLMs for diagnosis prediction and examine the rationale behind their decision-making processes, aimed at improving diagnostic safety. Using real-world hospital datasets, our study serves to enrich the discourse on the role of medical KGs in grounding medical knowledge into LLMs, revealing both challenges and opportunities in harnessing external knowledge for explainable diagnostic pathway and the realization of AI-augmented diagnostic decision support systems.