Ontology-based Error Detection in SNOMED-CT

1
Ontology-based Error Detectionin SNOMED-CT

• Werner Ceusters
• European Centre for Ontological Research
• Universität ds Saarlandes
• Barry Smith, Anand Kumar,
• Christoffel Dhaen

2
Presentation overview

• Why this effort on SNOMED-CT ?
• Error detection algorithms
• Combining lexical and terminological relations
• Mistakes detected
• Automated post-coordination
• Mistakes detected
• Conclusion
• Future Work

3
Snomed-CT in Europe ?
1st Recommendation Do conduct at least two
major real life learning experiments with SNOMED.
Make on the basis of the outcome a decision on
SNOMED as a national standard.
4
Description of Algorithms
Ceusters W, Smith B, Kumar A, Dhaen C. Mistakes
in Medical Ontologies Where Do They Come From
and How Can They Be Detected? in Pisanelli DM
(ed) IOS Press, Studies in Health Technology and
Informatics, vol 102, 2004.
5
Lexical, semantic and conceptual relations
First algorithm
6
Mistake indicator
Conceptual search
Term search
7
Missed subsumption (1)
Concept search heart structure AND tumor
Term search heart tumor
8
Missed subsumption (2)
Missing ISA neoplasm of heart
9
Undetected synonymy (1)
10
Undetected synonymy (2)
subcutaneous mastectomy
isa
mastectomy for gynecomastia
isa
subcutaneous mastectomy for gynecomastia
11
Inadequate homonymy
leg repair
lower leg structure
Procedure-site
leg reconstruction
lower limb structure
Procedure-site
12
Mistakes due to inappropriatelexical mapping
Vulval vein
Vein of head
isa
isa
structure of labial vein
13
Total / partial inconsistencies
terms
14
Errors in direct (manual ?) subsumption assignment
mediastinoscope
isa
diagnostic endoscopic examination of mediastinum
NOS
Inguinal hernia
isa
Lichtenstien repair of inguinal hernia
15
Automatedpost-coordinationusing version space
theory
Second algorithm

• Guided creation of
• Most specific generalisations
• If there is a concept with properties P1 and P2,
  then create a concept that is defined by these
  properties
• Most generic specialisations
• Reclassify all existing concepts having the
  properties P1 and P2
• Process guided by existing subsumption hierarchy

16
Generated post-coordinations
17
Suspicious configurations

• The presence of only one generated concept in a
  list of the concepts subsumed by a given concept

The presence of only one existing subsumed
concept next to a list of generated concepts for
the same subsumer
The presence of a pre-existing concept that is
subsumed by a generated concept without any other
additional relationships from the pre-existing
concept to another one.
(xxx)
(xxx)


XXX
XXX
XXX


XXX


XXX

18
Reclassification example
19
Mereological errors(or peculiar meaning of part
of)
Thigh part
Lower leg structure
isa
isa
Structure of tibial nerve
20
Improper negation handling
contracture of palmar fascia
isa
Dupuytrens disease of palm, nodules with no
contracture
21
Conclusion

• SNOMED-CT is a great resource ...
• ... but must be improved
• ... By removing inconsistencies at three levels
• Lexical
• Semantic
• Conceptual
• ... Using adequate tools
• ... to make it a better system
• ... for clinical code assignment
• ... for post hoc decision support on the basis of
  the codes

22
Required future work

• Concept system ontology
• Concepts are unique units of thought .
• Disorders are concepts in which there is an
  explicit or implicit pathological process causing
  a state of disease which tends to exist for a
  significant length of time under ordinary
  circumstances.
• How ?