Consistency between Metathesaurus and Semantic Network Workshop on The Future of the UMLS Semantic Network NLM, April 8, 2005 Olivier Bodenreider Lister Hill National Center for Biomedical Communications Bethesda, Maryland - USA

2 Overview u Defining consistency u What does inconsistency mean? u Testing consistency l Comparing Metathesaurus relations to SN relations l Aligning Metathesaurus concepts and semantic types l Semantic type distribution of sets of descendants of Metathesaurus concepts u Suggestions l Enforcement mechanism l Ontology of relationships l CVF

Two levels in the UMLS

4 The UMLS: a two-level structure Concept 1 Metathesaurus Semantic Network Semantic Type a Semantic Type b Semantic Type c Concept 2

Heart Concepts Metathesaurus Esophagus Left Phrenic Nerve Heart Valves Fetal Heart Medias- tinum Saccular Viscus Angina Pectoris Cardiotonic Agents Tissue Donors Anatomical Structure Fully Formed Anatomical Structure Embryonic Structure Body Part, Organ or Organ Component Pharmacologic Substance Disease or Syndrome Population Group Semantic Types Semantic Network

6 Relationships can inherit semantics Semantic Network Metathesaurus Adrenal Cortex Adrenal Cortical hypofunction Disease or Syndrome Body Part, Organ, or Organ Component Pathologic Function isa Biologic Function isa Fully Formed Anatomical Structure isa location of

Defining consistency

8 The consistency “square” Concept 1 Metathesaurus Semantic Network Semantic Type a Semantic Type b Concept 2

9 The categorization link Semantic Network Professional Society Metathesaurus Salmonella American Medical Association Organism Bacterium isa is an instance of

10 Semantic Network relations u 54 types of relationships u 558 asserted relations (SRSTR) u 6703 fully expanded relations (SRSTRE*) Semantic Network Disease or Syndrome Body Part, Organ, or Organ Component Pathologic Function isa Biologic Function isa Fully Formed Anatomical Structure isa location of

11 Metathesaurus relations u REL vs. RELA u Not always labeled u 106 additional types of relationships u ~7 M symbolic relations Heart Concepts Metathesaurus Esophagus Left Phrenic Nerve Heart Valves Fetal Heart Medias- tinum Saccular Viscus Angina Pectoris Cardiotonic Agents Tissue Donors

12 Metathesaurus relations u Recorded l at the term level: from source vocabularies l at the concept level: from Metathesaurus editors u Aggregated at the concept level Oat cell carcinoma of lung Carcinoma, Small Cell SCLC Lung structure Lung Pulmonary has_finding_site Oat cell carcinoma of lung Carcinoma, Small Cell SCLC Lung structure Lung Pulmonary has_finding_site

13 Not all relationships in hierarchies are isa (1) Autoimmune Diseases Addison’s disease due to autoimmunity Tuberculous Addison’s disease is generally a

14 Not all relationships in hierarchies are isa (2) Environment and Public Health [G03] Public Health [G03.850] Accidents [G ] Accident Prevention [G ] + Accidental Falls [G ] Accidents, Aviation [G ] […] Drowning [G ] +

15 Defining consistency u SN rel. and Meta rel. must have the same direction u SN rel. and Meta rel. must be of the same type (both hierarchical or associative) u Meta rel. must be the same as SN rel. or one of its descendants Concept 1 Metathesaurus Semantic Network Semantic Type a Semantic Type b Concept 2

16 Examples of consistent relations Lung Body Part, Organ, or Organ Component Disease or Syndrome Pneumonia has_location

17 Examples of consistent relations Concept 1 Metathesaurus Semantic Network Semantic Type a Semantic Type b Concept 2 affects treats

What does inconsistency mean?

19 The consistency “square” revisited Concept 1 Metathesaurus Semantic Network Semantic Type a Semantic Type b Concept 2Concept 1 Metathesaurus Semantic Network Semantic Type a Semantic Type b Concept 2 ? ? ? ?

20 What does inconsistency mean? u Inaccurate/missing Semantic Network relation u Inaccurate (/missing?) categorization u Inaccurate Metathesaurus relation

Testing consistency

22 (A) Consistency of associative relations [McCray & Bodenreider, 2002]

23 Results u 6894 pairs of related concepts l 4496 (65%): a SN relation can be inferred unambiguously n Validity confirmed in 1981 cases n 2515 not labeled in the Metathesaurus l 1491 (22%): multiple possible SN relationships n multiple possible Metathesaurus relationships l 907 (13%): inconsistency SN/Meta relationships n 372: no SN relationship between the STs n 415: inconsistent SN/Meta relationship type (REL) n 120: inconsistent SN/Meta relationship attribute (RELA)

24 (B) Consistency of hierarchical relations u Relations used l SN: isa l Categorization: isa l Metathesaurus: PAR/CHD + RB/RN u Hypothesis l For a pair of (ST, C), the concepts categorized by ST (and its descendants) correspond to the descendants of the concept C l In the set of descendants of C, expected STs are the ST of C (and its descendants) ➊ ➋

25 ST-based classes vs. descendants u Semantic type l List of all concepts having this semantic type u Concept l List of all descendants u Comparing the 2 sets l Intersection of the 2 sets ➊ [Bodenreider & Burgun, 2004]

26 Analyzing inconsistencies AmphibianAmphibia 1126 descendants 1135 concepts 1124 in common Tadpole Invertebrate Toad licking Pharmacologic Substance Miscategor- ization (?) Wrong hierarchical relation Missing hierarchical relation Miscategor- ization Rana unclassified Class Reptilia Amphibians and Reptiles

27 Semantic types of descendants u Concept l Set of all descendants u Distribution of semantic types in the set l Allowable STs: ST of C and its descendants (strict) or ST from the same semantic group (loose) [Mougin & Bodenreider, 2005] ➋

28 Analyzing inconsistencies u 26,584 concepts studied u 59% of their descendants have a semantic type incompatible with that of the original concept Reaction belligerent Finding Hostility Mental Process

29 # # C Neoplasm of placenta (disorder) (neop) # * B: 190 C |ST|acab| 5.50|incp C |ST|anab| 1.50|incp C |ST|cgab| 76.50|incp C |ST|dsyn| 27.50|incp C |ST|inpo| 1.00|incp C |ST|neop| 76.50|comp C |ST|patf| 1.50|incp C |SG|DISO| |comp # Analyzing inconsistencies

Suggestions

31 Aligning SN and Meta relationships u 54 types of SN relationships u 106 additional types of Metathesaurus relationships u Some are simply synonymous (caused_by / due_to; follows / temporally_follows) u Some are specialized relationships (manifestation_of / definitional_manifestation_of) u Many types of mapping relationships, not in SN

32 Add classification information to SN u Explicit classificatory principles (in addition to textual definition and examples) u Abandon economy principle and return to JEPD (jointly exhaustive/pairwise disjoint) approach

33 Metathesaurus editing environment u Use SN/Meta relation consistency as a constraint for assigning semantic types u Use SN relations to suggest labels for unspecified Meta relations u Use SN/Meta relation consistency to guide the review by the Metathesaurus editors l Inaccurate categorization? l Inaccurate Metathesaurus relation?

Conclusions

35 Conclusions u Simultaneously l Improve SN l Improve categorization u ST assignment can be automated in part

36 Some references u McCray AT, Bodenreider O. A conceptual framework for the biomedical domain. In: Green R, Bean CA, Myaeng SH, editors. The semantics of relationships: an interdisciplinary perspective. Boston: Kluwer Academic Publishers; p u Bodenreider O, Burgun A. Aligning knowledge sources in the UMLS: Methods, quantitative results, and applications. Medinfo 2004:

37 Some references u Burgun A, Bodenreider O. Aspects of the taxonomic relation in the biomedical domain. In: Welty C, Smith B, editors. Collected papers from the Second International Conference "Formal Ontology in Information Systems": ACM Press; p u Mougin F, Bodenreider O. Approaches to eliminating cycles in the UMLS Metathesaurus: Naive vs. formal. Proceedings of AMIA Annual Symposium 2005:(submitted).

Medical Ontology Research Olivier Bodenreider Lister Hill National Center for Biomedical Communications Bethesda, Maryland - USA