UMLS Metathesaurus

The Metathesaurus is organized by concept (meaning), not by term (string). Each unique clinical meaning is assigned a permanent Concept Unique Identifier (CUI). All synonymous terms from different source vocabularies—such as 'Hypertension' from SNOMED CT and 'High Blood Pressure' from ICD-10-CM—are linked to the same CUI.

• CUI: Permanent, meaningless identifier for a concept
• Atom: Each occurrence of a term in a source vocabulary
• String: The actual textual representation of a term
• This design enables cross-vocabulary mapping and semantic interoperability

The LVG engine normalizes biomedical terms to their canonical forms, dramatically improving recall during concept mapping. It applies a cascade of linguistic transformations to generate all known morphological variants of a term.

• Inflectional Morphology: Handles pluralization ('fracture' → 'fractures')
• Derivational Morphology: Converts between parts of speech ('inflammation' → 'inflammatory')
• Spelling Normalization: Maps British to American English ('oesophagus' → 'esophagus')
• Acronym Expansion: Resolves 'MI' to 'Myocardial Infarction' using context
• The output is a set of normalized strings used for dictionary-based entity recognition

Every Metathesaurus concept is assigned one or more Semantic Types from the UMLS Semantic Network, providing a high-level categorical framework. These types are linked by Semantic Relations, forming a graph that enables hierarchical and associative reasoning.

• 133 Semantic Types: Including 'Disease or Syndrome', 'Pharmacologic Substance', 'Laboratory Procedure'
• 54 Semantic Relations: Such as 'isa' (hierarchical), 'causes', 'treats', 'location_of'
• This network allows systems to infer that 'Aspirin' treats 'Pain' even if that explicit link is not in the source data
• Enables validation rules like 'A medication cannot be prescribed for a body part'

The Metathesaurus preserves full provenance for every atom, maintaining an unbroken chain back to its originating vocabulary. This is critical for clinical audit trails and regulatory compliance.

• Source Abbreviation (SAB): Identifies the originating vocabulary (e.g., 'SNOMEDCT_US', 'RXNORM')
• Source-UI: The original identifier in the source system
• Suppressibility Flags: Indicate if a term is obsolete or not recommended for production use
• Term Status: Tracks if a term is current, retired, or erroneous
• This design allows systems to filter by trusted sources and trace every mapped concept back to its authoritative origin

The MRREL file is the core relational engine of the Metathesaurus, storing millions of directed relationships between concepts. These relationships are harvested from source vocabularies and enriched by UMLS editors.

• REL: The relationship type (e.g., 'PAR' for parent, 'CHD' for child, 'RB' for broader)
• RELA: A more specific label when available (e.g., 'has_ingredient', 'dose_form_of')
• Mapping Relationships: Explicit cross-walks between vocabularies, such as ICD-10-CM to SNOMED CT maps
• Co-occurrence Data: Statistical relationships derived from MEDLINE and clinical corpora
• These tables power concept expansion, hierarchical querying, and automated terminology translation

The Metathesaurus is distributed as a set of normalized, pipe-delimited Rich Release Format (RRF) files, each representing a logical data table. This columnar, relational design enables efficient bulk loading into databases and graph stores.

• MRCONSO: The core concept table, containing all atoms and their CUIs
• MRSTY: Maps CUIs to Semantic Types
• MRREL: Stores all relationships between concepts
• MRDEF: Contains textual definitions for concepts
• MRSAT: Stores flexible attribute-value pairs for extensibility
• This file-based architecture allows implementers to load only the subsets relevant to their use case