MedKGent: LLM Agent Framework for Temporal Medical Knowledge Graphs
A new framework called MedKGent has been developed by researchers, featuring a Large Language Model (LLM) agent aimed at creating medical knowledge graphs (KGs) that evolve over time. This innovative system analyzes more than 10 million PubMed abstracts spanning from 1975 to 2023, continuously updating the KG each day. It utilizes two distinct agents: the Extractor Agent, which detects knowledge triples and assigns confidence levels, and the Constructor Agent, which weaves these triples into a temporal graph, enhancing consistent knowledge and addressing discrepancies. The KG comprises 156,275 entities and 2,971,384 triples, making it the largest medical KG derived from an LLM to date. Both automated and expert evaluations indicate triple-validity rates nearing 90%, tackling the challenges of generalizability and temporal dynamics in existing KG construction approaches.
Key facts
- MedKGent is an LLM agent framework for building temporally evolving medical KGs.
- It uses over 10 million PubMed abstracts from 1975 to 2023.
- The KG is constructed incrementally on a daily basis.
- Two agents: Extractor Agent (identifies triples and assigns confidence) and Constructor Agent (integrates triples into temporal graph).
- The resulting KG has 156,275 entities and 2,971,384 triples.
- It is the largest LLM-derived medical KG to date.
- Triple-validity rates approach 90% based on automated and expert assessments.
- The framework addresses generalizability and temporal dynamics issues in current KG construction.
Entities
Institutions
- arXiv
- PubMed