Extraction of SNOMED CT Associative Relations to Improve Grouping of the Related WHO-ART Adverse Drug Reactions Terms

1
Extraction of SNOMED CT Associative Relations to
Improve Grouping of the Related WHO-ART Adverse
Drug Reactions Terms

• Iulian Alecu
• Cédric Bousquet
• Marie-Christine Jaulent

2
Pharmacovigilance

• Definition
• the science and activities relating to the
  detection, assessment, understanding and
  prevention of adverse effects or any other
  drug-related problems WHO
• Signal in pharmacovigilance
• any reported information on a possible causal
  relationship between a drug and an adverse drug
  reaction, the relationship being unknown or
  incompletely documented previously WHO
• WHO-UMC database contains 4M of cases.
• Signal detection
• Statistical detection. (Bate A.)
• Manual review.

3
Hypothesis

• ADR are coded with a standard terminology
  WHO-ART (Word Health Organisation Adverse
  Reaction Terminology)
• Currently semantic information is not taken into
  account in signal detection (data mining)
• Semantic grouping of ADR terms would improve the
  signal detection

4
WHO ART (limits)

• Encoding data
• no complaints until now (35 years of existence
  and 5 new terms added in the last 5 years).
• transmission, unifying resources in a domain
  where sharing data is vital.
• Regulatory activities
• primary conception criteria of the groupings
• surveying drug induced disorders
• public health general problems (foetal disorders,
  neoplasm)
• multiple-inheritance constraints on terms
• Data retrieval for data mining
• recognized limit
• almost flat hierarchy
• impossible to do data mining on other topics than
  the regulatory ones

5
WHO ART (limits)

• Diseases, signs, symptoms and abnormal laboratory
  exams results (1857 PT)
• Three hierarchical levels (PT/IT,HLT,SOC) that
  are not systematically respected over its
  hierarchy.
• 68.7 of PTs linked directly to a SOC (average of
  58 PT/SOC)
• 31.3 under a HLT (average of 4 PT/HLT)
• either very large clusters (sensible but not too
  specific) or very small (specific but not too
  sensible) ones.
• Terms that have different levels of
  generalization may be siblings.
• thrombosis arterial has as siblings
  "thrombosis arterial arm", "thrombosis arterial
  leg", thrombosis cerebral arterial, thrombosis
  carotid, thrombosis retinal artery
• Objective Investigate the use of SNOMED CT to
  provide relevant groupings of WHO ART terms.

6
Snomed CT

• Integrated in UMLS as well as WHO-ART
• a mapping has been already done
• Large coverage of the medical domain and includes
  the medical areas of WHO-ART
• More granular than WHO-ART
• Better organized and multi axial hierarchy
• Extended utilization

7
Improving the hierarchy of WHO-ART
8
Formal definitions OWL
9
Achievements

• A better hierarchy of WHO-ART
• Use of SNOMED CT grouping terms to extract topic
  oriented grouping of WHO-ART terms that are
  overlapping on gold standard manually created
  groupings
• This new model does not capture all the knowledge
  (the associative relations) from SNOMED CT and
  can be improved

10
Limits of IS_A for relevant grouping WHO-ART terms
11
OWL capture model

• Cwho complete
• (is_syn some (Csnmct or Csnmct or .))
• or (has_finding_site some (Csnmct and Csnmct and
  .))
• or (has_associated_morphology some (Csnmct and
  Csnmct and .))
• For example
• Gastritis
• complete
• (is_syn some (SNMCTGastritis))
• or (has_finding_site some (SNMCT
  Stomach_structure))
• or (has_associated_morphology some
  (SNMCTInflammation))
• Gastritis acute
• complete
• (is_syn some (SNMCTAcute_Gastritis))
• or (has_finding_site some (SNMCTStomach_structure
  ))
• or (has_associated_morphology
• some (SNMCTAcute_Inflammation))

12
The model
Cwho complete (is_syn some (Csnmct or Csnmct
or .)) or (has_finding_site some (Csnmct and
Csnmct and .)) or (has_associated_morphology
some (Csnmct and Csnmct and .))

• By putting an union operator between different
  role restrictions we are simplifying the
  reasoning process.
• The synonymous relationships are modeled as
  restrictions on unions of UMLS asserted synonyms
  in order to capture all the possible meanings.
• The associative relationships are modeled as
  intersection of SNOMED CT fillers in order to
  insure inclusion in query classes

13
By putting an union operator between different
role restrictions we are simplifying the
reasoning process.

• Logically an AND (intersection) should stand
  between these restrictions.
• Given the fact that all classes are defined
  according to this pattern, the operator (and/or)
  does not matter in the classification process.
• All classes have a synonymy restriction.
• We are dealing with a large and complex ontology
  (10k classes)
• In the reasoning process comparisons of unions
  are easier to solve computationally than
  comparisons of intersections.
• We are not intending to perform reasoning on
  individuals (pharmacovigilance cases are already
  coded).
• This artefact imposes to define owlthing for
  undocumented associative relations

14
Classification process
GASTRITIS
has_associated morphology some
has_finding_site some
is_syn some
Gastritis
Stomach Structure
Inflammation
15
Query process
GASTRITIS
OWLThing
OWLThing
INFLAMATIONS
Gastritis
Stomach Structure
Inflammation
16
The undocumented relations have to be replaced by
owlthing in order to compensate the artefact
induced by unions.
OWLThing
OWLThing
INFLAMATIONS
OBESITY
Inflammation
Obesity
Obese
OWLNothing
OWLNothing
OWLThing
OWLThing
17
The synonymous relationships are modeled as
restrictions on unions of UMLS asserted synonyms
in order to capture all the possible meanings.

• Hiatus Hernia is synonymous with Hiatus Hernia,
  Hiatus Hernia NOS
• Hiatus Hernia NOS is a subclass of Hiatus Hernia
• Obesity is synonymous with Obesity, Obese
• In this case the two terms are far in the SNOMED
  CT hierarchy.
• This is a ambiguity and needs to be reviewed.
• Cases like this are quite easy to be spotted.

18
The associative relationships are modeled as
intersection of SNOMED CT fillers in order to
insure inclusion in query classes
OWLThing
OWLThing
INFLAMATIONS
DERMATITIS ATOPIC
Inflammation
Skin Structure
Papulovesicular rash
Atopic dermatitis
19
Conclusions

• 14 of WHO-ART not mapped on SNOMED-CT is a limit
  that produce silence in clusters.
• The mapping needs to be manually reviewed.
• Reuse of an already existing resource
• Using concepts from SNOMED-CT to cluster WHO-ART
• A first step towards an evolving terminology
• provides an developing environment
• reasoning insures logical consistency
• reviewers of WHO-ART can consider the SNOMED-CT
  clustering concepts to create finer coarse
  clustering levels
• insuring future EHR compliance
• This model allows more effective groupings of
  WHO-ART terms, that are not possible by IS_A
  relationships.

20

• Thank you !
• We acknowledge the WHO Collaborating Centre for
  International Drug Safety Monitoring for
  providing the WHO-ART sources.