Ontological analysis of SNOMED CT

1
Ontological analysis of SNOMED CT

• Gergely Héja MSc., György Surján MD., Péter Varga
• MSc.
• National Institute for Strategic Health Research,
  Budapest, Hungary

2
Aims

• ESKI needs a reference terminology
• To represent of classification systems (ICD,
  ICPM) and public health indicators
• To provide pre co-ordinated code lists for
  enabling semantic interoperability in Hungary
  (and potentially in the EU)
• SNOMED CT seems to be a first choice candidate

3
Needs for classification systems

• Combinatory representation of categories based on
  a reference ontology to enable
• supporting statistical analysis
• formal consistency checking
• (semi)automatic interconnection of different
  classification systems (e.g. ICD and ICPM)
• A high-quality formal ontology is needed, but it
  needs not to be very detailed

4
Needs for interoperability

• Detailed common terminology covering the whole
  domain of medicine
• Mapping the concepts of the HIS to the concepts
  of the common terminology and vice versa
• A common, consistent, comprehensive and decidable
  ontology is needed

5
Computational issues

• Computability vs. comprehensiveness
• Clean hierarchies
• Less emphasis on coverage (the multitude of
  non-defined leaf categories) and greater emphasis
  on rich and well-organized high-level categories
  is needed
• SNOMED CT contains too much concepts

6
Methods

• Based on DOLCE formal top-level ontology
• Review of high and low-level concepts needed for
  the representation of classification systems and
  public health indicators
• Is the subsumption relation valid?
• Meaning derived from the FSN vs. meaning derived
  from synonyms vs.meaning derived from formal
  definition

7
Error types - 1

• Misplacing concepts in the hierarchy
• smoker (an agent) is subsumed by tobacco smoking
  behaviour finding (a role)
• severe asthma is not a kind of asthma, but a kind
  of asthma finding.
• Mixing the subsumption relation with other roles
  (typically part of)
• haemoglobin subsumes haemin (instead of
  constitutional part)
• exacerbation of asthma attack is subsumed by
  asthma (instead of temporal part)

8
Error types - 2

• Hierarchy violating medical thinking and
  biomedical knowledge
• Disease, observation and finding are subsumed by
  clinical finding
• acute on chronic, which is both subsumed by acute
  and chronic
• polycarbonate is a polymer (instead of synthetic
  polymer)

9
Error types - 3

• Contracting disjoint entities into one concept
• Smoker (an agent) and smoker (finding) (a
  description of a situation)
• additional pathologic finding in tumor specimen
  (observable entity) and additional pathologic
  finding
• Function is classified as an observable entity
• Ontological definition ability of an object to
  play a certain role in a certain kind of activity
• functions (e.g. gene function, adaptation)
• measures (quality) that evaluate the realisation
  of a function (e.g. respiratory rhythm, excretory
  rate)
• Inflammation (morphological abnormality) (a
  physical object) and inflammatory reaction
  (perdurant)

10
Additional problems - 1

• Categories taken form classification systems
• pneumonia in other diseases classified elsewhere
  (marked as ConceptStatus Limited)
• The danger of taking over concepts from other
  conceptual systems the context of the concept is
  lost. What is meant by other diseases classified
  elsewhere?
• relations (such as part of) are represented also
  as concepts
• It prohibits the direct conversion to any
  formalism based on first order logic, thus to any
  DL formalism

11
Additional problems - 2

• Underspecification
• roles are not quantified (existential /
  universal)
• criteria are not specified (necessary /
  sufficient)
• conversion to DL do we have to decide in each
  particular case, or can it be done universally?
• Multiple hierarchy
• alcoholic beverage (through its parent ingestible
  alcohol) is subsumed by central depressant, ethyl
  alcohol and psychoactive substance of abuse
  non-pharmaceutical. Alcoholic drinks contain
  ethyl alcohol, which plays a role of depressant
  and substance of abuse (with respect to human
  beings)
• Is this a general phenomenon in SNOMED?
• Which relations are asserted and which are
  inferred?

12
Discussion - 1

• The intended meaning of the categories is not
  always clear possible translation errors
• Is it reasonable to import categories from
  medical classifications?
• Size
• Artificial concepts
• Consistency errors

13
Discussion - 2

• Real world entities listed heterogeneously
• Mars bar and Kit Kat (chocolate candy would
  suffice)
• UFO is subsumed by transport vehicle
• tendon pulley reconstruction is represented, but
  tendon pulley not

14
Solutions

• Use SNOMED as a plain or loosely structured list
  of terms (with extending the coverage). Not
  appropriate for intelligent services.
• Restructure SNOMED into a high-quality ontology.
• Build a new medical ontology from scratch
  (partial reuse of the existing ones), and to
  restrict the use of SNOMED for interoperability
  by mapping concepts to it.

15
Restructuring SNOMED

• A formal top level ontology (e.g. DOLCE).
• A high level core reference ontology of general
  medical knowledge (e.g. anatomy, physiology,
  pathology, medical procedures).
• Logic-based formalism
• Single hierarchies with formal definitions
• (sub)domain ontologies of specialities
• Compound entities (e.g. tonsillitis)
• Manual assertion (e.g. autism)

16
Questions?