Researchers build biomedical AI models trained on 3.4 million medical concepts from UMLS, comparing two approaches to make language models understand medical knowledge

Researchers created BERTUMLS and BioBERTUMLS — language models (AI systems that understand and generate text) trained on structured medical knowledge from UMLS Metathesaurus, a database of 3.4 million medical concepts and 34.2 million relationships. They tested two methods: embedding medical knowledge directly into the model during training, and using a knowledge graph (a map of how medical terms connect) that the AI consults when answering questions.

The key difference: continual pretraining bakes medical knowledge into the model's 'brain' permanently, making it heavier but self-contained; GraphRAG keeps knowledge separate and looks it up live, like consulting a reference book during a test. For biomedical researchers, this means choosing between a model that 'knows' medicine versus one that can 'look up' accurate answers without hallucinating medical facts.

For healthcare companies, drug discovery teams, and hospitals using AI to read medical literature or analyze patient records, this research offers a roadmap to make AI systems more trustworthy in high-stakes settings — medical AI that admits what it doesn't know (via lookup) versus AI that confidently states wrong facts. The UMLS knowledge graph is publicly available, so any organization can build on this approach.