Battling Scylla and Charybdis: The Search for Redundancy and Ambiguity in the 2001 UMLS Metathesuarus James J. Cimino Department of Medical Informatics Columbia University
2001 Metathesaurus 99 sources (92 in 2000) 1,734,707 strings (1,598,176 in 2000) 797,360 concepts (730,155 in 2000)
Lumping vs. Splitting Cold (infection) Cold (temperature) COLD (COPD) COLD (temperature) Cold (infection) Cold (temperature) COLD (COPD) COLD (temperature) Ambiguity! Redundancy!
Three Auditing Methods Ambiguity through of multiple semantic types Redundancy through semantic string matching Inconsistency in parent-child semantic types
Previous Results: 1995 Possible ambiguity 1,817 Possible redundancy 5,031 Actually redundancy 3,274 Parent-Child problems 544 * Cimino JJ. Auditing the Unified Medical Language System with semantic methods. Journal of the American Medical Informatics Association; 1998;5: *
Tools and Rules Simple Metathesaurus data model Normalized word index “Mutually exclusive semantic types” “Mutual concept subsumption”
Simple Metathesaurus Data Model L : S : “COLD ” C : Chronic Obstructive Airway Disease L : S : “Chronic Obstructive Airway Disease” L : S : “Chronic Obstructive Lung Disease” Semantic type: T04: Disease or Syndrome S : “COLD”
Simple Metathesaurus Data Model S : “COLD ” C : Chronic Obstructive Airway Disease S : “Chronic Obstructive Airway Disease” S : “Chronic Obstructive Lung Disease” Semantic type: T04: Disease or Syndrome S : “COLD”
Simple Metathesaurus Data Model “COLD ” C : Chronic Obstructive Airway Disease “Chronic Obstructive Airway Disease” “Chronic Obstructive Lung Disease” Semantic type: T04: Disease or Syndrome “COLD”
Simple Metathesaurus Data Model C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome COLD Chronic Obstructive Airway Disease Chronic Obstructive Lung Disease COLD C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome Parent-Child (is-a)
UMLS Semantic Types Physical Object Organism Substance Food Alga Plant Invertebrate Animal
Mutually Inclusive Semantic Types Physical Object Organism Animal Invertebrate Plant Alga Substance Food
Mutually Exclusive Semantic Types Physical Object Organism Animal Invertebrate Plant Alga Substance Food
Rules for Multiple Semantic Types 3. Concepts can have two Substance types, except: a) Element, Ion or Isotope and Chemicals Viewed Structurally b) Inorganic Chemical and Organic Chemicals 5. Concepts can have two Conceptual Entity types, except: Molecular Sequence and Geographic Area Molecular Sequence and Body Location or Region Geographic Area and Body Location or Region 7. Concepts can have two Event types, except: Diagnostic Procedure and Laboratory Procedure 8. Concepts can have two types that ancestors/descendants
Detection of Ambiguity by Mutually Exclusive Semantic Types If a concept has multiple semantic types And if any pair of the types are mutually exclusive Then the concept may have multiple meanings (ambiguity) Or the semantic type assignment is incorrect
Ambiguity Examples C : Euglena gracilis Alga and Invertebrate C : Fourth lumbar vertebra Body Part, Organ, or Organ Component and Disease or Syndrome C : Toxicodendron Plant and Disease or Syndrome C : Crown-Rump Length Organism Attribute and Diagnostic Procedure C : Cell Movement Cell Function and Biomedical Occupation or Discipline C : Lice Infestations Invertebrate and Disease or Syndrome C : Chronically Ill Disease or Syndrome and Patient or Disabled Group
Normalized Word Index UMLS Normalized Word Index –e.g., “lungs” “lung” –293,004 words Keyword synonyms –e.g., “lung” “pulmonary” –9,650 mappings Translated strings Built word index
Word Normalization C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome COLD Chronic Obstructive Airway Disease Chronic Obstructive Lung Disease COLD C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome Parent-Child (is-a)
Word Normalization Parent-Child (is-a) C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome cold 3 chronic obstructive airway disease chronic obstructive lung disease cold
Word Normalization Parent-Child (is-a) C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome cold 3 chronic obstructive airway disease chronic obstructive pulmonary disease cold
Word Normalization Parent-Child (is-a) C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome cold 3 chronic obstructive airway disorder chronic obstructive pulmonary disorder cold
Word Normalization Parent-Child (is-a) C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome cold three chronic obstructive airway disorder chronic obstructive pulmonary disorder cold
Word Index airway chronic cold disorder obstructive pulmonary three Parent-Child (is-a) C : Respiratory Tract Diseases Semantic type: T04: Disease or Syndrome C : Chronic Obstructive Airway Disease Semantic type: T04: Disease or Syndrome cold three chronic obstructive airway disorder chronic obstructive pulmonary disorder cold
Mutual String Subsumption 1) If Concept A has String A1 And all words in A1 are in Concept B’s word list Then B subsumes A1 2) If B subsumes any string in A And A subsumes any string in B Then A and B are mutually subsumptive
Mutual String Subsumption common cold cold two cold C : Common Cold T04: Disease or Syndrome cold common two C : cold temperature cold temperature cold one cold T070: Natural Phenomenon or Process cold one temperature C : Chronic Obstructive Airway Disease chronic obstructive airway disorder chronic obstructive pulmonary disorder cold three cold T04: Disease or Syndrome airway chronic cold disorder obstructive pulmonary three
Mutual String Subsumption C : cold temperature cold temperature cold one cold T070: Natural Phenomenon or Process common cold cold two cold C : Common Cold T04: Disease or Syndrome C : Chronic Obstructive Airway Disease chronic obstructive airway disorder chronic obstructive pulmonary disorder cold three cold T04: Disease or Syndrome cold common two cold one temperature airway chronic cold disorder obstructive pulmonary three
Mutual String Subsumption C : cold temperature cold temperature cold one cold T070: Natural Phenomenon or Process common cold cold two cold C : Common Cold T04: Disease or Syndrome C : Chronic Obstructive Airway Disease chronic obstructive airway disorder chronic obstructive pulmonary disorder cold three cold T04: Disease or Syndrome cold common two cold one temperature airway chronic cold disorder obstructive pulmonary three
Mutual String Subsumption C : cold temperature cold temperature cold one cold T070: Natural Phenomenon or Process common cold cold two cold C : Common Cold T04: Disease or Syndrome C : Chronic Obstructive Airway Disease chronic obstructive airway disorder chronic obstructive pulmonary disorder cold three cold T04: Disease or Syndrome cold common two cold one temperature airway chronic cold disorder obstructive pulmonary three
Detection of Redundancy by String Subsumption If A and B are mutually subsumptive And semantic types of A and B are mutually inclusive Then A and B may be redundant
Detection of Redundancy by String Subsumption C : cold temperature cold temperature cold one cold T070: Natural Phenomenon or Process common cold cold two cold C : Common Cold T04: Disease or Syndrome C : Chronic Obstructive Airway Disease chronic obstructive airway disorder chronic obstructive pulmonary disorder cold three cold T04: Disease or Syndrome cold common two cold one temperature airway chronic cold disorder obstructive pulmonary three
Redundancy Examples C : NPS-R-467 (Organic Chemical) C : NPS R-467 (Organic Chemical) C : des-Arg(10)-(Leu(9))kallidin (Amino Acid, Peptide or Protein) C : kallidin, des-Arg(10)-(Leu(9))-) (Amino Acid, Peptide or Protein) C : Congenital diverticulum of esophagus (Congenital Abnormality) C : Congenital esophageal pouch (Congenital Abnormality)
Incorrect synonymy (MeSH translations) C : Dolphins has synonyms “ORCA” (Span.) and "FALSA BALEIA ASSASSINA“ (Port.) so it is mutually subsumptive with C : Whale, False Killer which has synonym "FALSA ORCA" (Span.) Redundancy False Positives Partial names as synonyms: C : Central Diabetes Insipidus has “Diabetes Insipidus” as synonym so it is mutually subsumptive with C : Diabetes Insipidus
Detecting Semantic Type Problems through Parent-Child Relations If Concept A is Parent of Concept B And Concept A has semantic type X And Concept B has semantic type Y And if X and Y are different And X is not an ancestor of Y (in Semantic Net) Then one (or both) semantic types are wrong Or the parent-child relation is wrong
Detecting Semantic Type Problems through Parent-Child Relations OK Wrong Type or Wrong Concept OK Nonspecific Semantic Type Cartilaginous Fish (vertebrate) Shark (vertebrate) Dogfish (fish) Stingray (animal) Skate (manufactured object)
Parent-Child Examples C : Elbow has type Body Location or Regions which is in the Conceptual Entity hierarchy Is parent of: C : Right elbow has type Body Part, Organ, or Organ Component which is in the Physical Object hierarchy
Results: 1995 VS Possible ambiguity 1,817 Possible redundancy 5,031 Actually redundant 3,274 Parent-Child problems 544 Number of concepts:222,927797,359 (3.6x) Parent-Child relations100,586607,043 (6.0x) 8,082 38,140 not done 2,868
Results: 1995 VS Possible ambiguity 1,817 (0.82%) 8,082 (1.01%) Possible redundancy 5,031 (2.26%) 38,140 (4.78%) Actually redundant 3,274 (1.47%) not done Parent-Child problems 544 (0.54%) 2,868 (0.47%) Number of concepts:222,927797,359 (3.6x) Parent-Child relations100,586607,043 (6.0x)
Discussion: Ambiguity Detection Small number (1.01%) is a good sign Allows focusing manual review Semantic type definitions need to be clarified Semantic type assignment rules need to be clarified
Discussion: Redundancy Detection Specificity is worse, without improved sensitivity Normalized string index is part of the reason “Incomplete” names are a bigger part of the reason Manual review will be relatively inefficient Incorrect mappings detected, especially foreign language
Discussion: Parent-Child Relations Mostly detects errors in semantic type assignment Strict hierarchy in Semantic Net causes problems
Conclusions Specific “answers” not possible –Domain expertise needed for assessment of chemical names –Assessments are necessarily subjective –NLM gets to make the rules –NLM hasn’t finished making the rules Methods provide focus for manual review Methods highlight where clearer definitions are needed The results show the UMLS is doing well at a difficult task
Acknowledgments NLM: Bill Hole, Alexa McCray and Betsy Humphreys Home: Rachel and Rebecca Cimino