New AI System Builds Evolving Medical Knowledge Graph from Decades of Research

Keeping pace with medical knowledge has become increasingly difficult as the scientific literature expands exponentially. A team of researchers has developed MedKGent, an AI-powered framework that automatically organizes medical knowledge into a structured graph that evolves with the field.

The Challenge of Scaling Medical Knowledge#

The medical research landscape produces enormous volumes of new findings annually. Traditional approaches to capturing and organizing this knowledge fail to scale, and existing knowledge graph construction methods don't account for how medical understanding changes and improves over time. MedKGent was designed to address both limitations.

How MedKGent Works#

The system operates through a two-agent approach working together to build and maintain the knowledge graph. The Extractor Agent reads through research literature and identifies key relationships—extracting what researchers call "knowledge triples" (connections between concepts) and assigning confidence scores to each one based on reliability. The Constructor Agent then takes these extracted triples and integrates them into a temporal graph structure, reinforcing knowledge that appears repeatedly across papers and resolving contradictions when they arise.

Using over 10 million abstracts from PubMed spanning nearly 50 years (1975–2023), the framework incrementally builds the graph day by day, creating a historical record of how medical knowledge has developed and changed.

Scale and Accuracy#

The resulting knowledge graph contains approximately 156,000 entities connected by nearly 3 million relationships. According to the researchers, this represents the largest LLM-derived medical knowledge graph created to date. Both automated testing and expert review found that roughly 90 percent of the extracted relationships were valid and accurate.

Real-World Impact#

In practical tests, integrating MedKGent's knowledge graph improved how five different large language models answered medical questions across seven separate benchmark tests. The system was particularly valuable for retrieval-augmented generation—a technique where AI systems pull relevant background information to answer questions more accurately.