MoST: A System To Semantically Map Clinical Model Data to SNOMED-CT

1
MoST A System To Semantically Map Clinical
Model Data to SNOMED-CT
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Rahil Qamar, Alan Rector
• Medical Informatics Group
• Department of Computer Science
• University of Manchester
• Manchester, U.K.
• qamarr_at_cs.man.ac.uk, rector_at_cs.man.ac.uk

2
Outline
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Background
• Clinical Models Archetype Models
• Clinical Terminologies - SNOMED-CT (need for
  further explanation??)
• Lexical and Semantic Mapping
• The Model Standardisation using Terminology
  (MoST) System
• Results
• Issues

3
BACKGROUND .. (1)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Data standards are basic building blocks for
  achieving data interoperability.
• Data interoperability enables information systems
  to be interoperable.
• Interoperable information systems are vital for
  reducing medical errors and increasing care
  efficiency.

So, we need systems that can produce and work
with standardised data. This means we first need
to map data to some standard terminology!
4
BACKGROUND .. (2)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• DATA SOURCES
• HL7 V3 Messages
• Templates (Data-entry forms)
• Archetypes
• Narratives
• EHRs
• Medical documents
• .

• TERMINOLOGY SOURCES
• SNOMED-CT
• GALEN
• ICD9, ICD10,
• LOINC
• GO
• .

5
BACKGROUND .. (2)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• TERMINOLOGY SOURCES
• SNOMED-CT
• GALEN
• ICD9, ICD10,
• LOINC
• GO
• .

• DATA SOURCES
• HL7 V3 Messages
• Templates (Data-entry forms)
• Archetypes
• Narratives
• EHRs
• Medical documents
• .

Standardised Health Data Repository
Terminology Files
6
BACKGROUND .. (2)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• TERMINOLOGY SOURCES
• SNOMED-CT
• GALEN
• ICD9, ICD10,
• LOINC
• GO
• .

• DATA SOURCES
• HL7 V3 Messages
• Templates (Data-entry forms)
• Archetypes
• Narratives
• EHRs
• Medical documents
• .

Standardised Health Data Repository
Terminology Files
7
Archetype Models
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Computable expressions of a domain content model.
  
• Expressions in the form of structured constraint
  statements inherited from openEHR Reference Model.

openEHR Archetype Models
Taken from the Ocean Informatics website
http//www.oceaninformatics.biz/
8
Example Barthel Index Archetype Model
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
Java Archetype Editor developed at Department of
Biomedical Engineering, Linkoping University,
Sweden
9
Barthel Index Terminology Section
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
From SNOMED CT Jan 2006
10
Autopsy Examination Terminology Section
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
11
Autopsy Examination Terminology Section with
enhanced intelligence - intended meaning
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
SNOMED CT Categories Jan 2006
12
The Model Standardisation using Terminology
(MoST) System
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
METHODOLOGY
Clinical Data Model E.g. Archetypes or HL7 v3
messages
Generalised hierarchy E.g. Archetype in MoST XML
Model transformation
input
PRE-FILTERED RESULTS
POST-FILTERED RESULTS
Non-context Methods (1) EMT-P, UMLS, and Lexical
Lookup
Context Methods (2) UMLS, and Training dataset
Lookup
Filter Methods (Results from 1 and 2 get filtered
for semantic appropriateness)
output
UMLS
GATE/ WordNet
Training Dataset
SNOMED-CT
Intended meaning document (IM Doc)
Candidate Mappings Candidate terms with metadata
13
PRE-FILTERED RESULTS
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• SNOMED-CT results prior to being filtered
• Archetype Terms ? SNOMED-CT Concepts
• (1..1 matches)

Terms Sent (1) Codes retrieved (2) Relevant codes (3) Precision (3/2) Recall (2/1)
475 425 350 82.35 89.47

• 50 archetype terms did not find any match in
  SNOMED resulting in an overall high recall value
  of 89.47
• A manual inspection by the clinical modelers
  gave an estimated precision of 82.35. Good value
  but not reliable at this point!

14
POST-FILTERED RESULTS
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• SNOMED-CT results after being filtered
• Archetype Terms ? SNOMED-CT Concepts
• (1..1 matches)

Terms Sent (1) Codes retrieved (2) Relevant codes (3) Precision (3/2) Recall (2/1)
475 425 385 90.58 89.47

• Of the 425 SNOMED codes only 385 were found to
  be relevant by the MoST system resulting in a
  precision of 90.58.
• The precision improved as most of the irrelevant
  results were eliminated resulting in a smaller
  but better final result set.
• This result set was displayed to the clinical
  modeler as candidate mappings.

15
RESULTS OVERVIEW
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
Processing time 60 sec 30 sec without Spell
Check Processing time 70 sec 40 sec with
Spell Check (GSpell)
16
FILTERING PROCESS .. (1)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
Rule 1 If one input concept subsumes another
then the subsuming concept is selected.
Example from autopsy examination archetype for
term Respiratory System
17
FILTERING PROCESS .. (2)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Rule 2 If the input concepts are disjoint with
  no common subsuming concept then all disjoint
  concepts are selected.
• Example from blood gas assessment archetype for
  archetype term pH .
• Filtering input Hydrogen ion concentration,
  Past history of, ph, pH measurement arterial
• Hydrogen ion concentration (observable entity)
• is_a Fluid observable
• Past history of (context-dependent category)
• is_a Context-dependent categories
• ph (qualifier value)
• is_a skin reaction grades
• pH measurement arterial (procedure)
• is_a pH measurement

From SNOMED-CT Jan 2006

• Output Hydrogen ion concentration, Past
  history of, ph, pH measurement arterial

18
FILTERING PROCESS .. (3)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006
Rule 3 All results are filtered using the SNOMED
categories stated in the IM Doc. Archetype term
descriptions are also considered.
19
ISSUES .. (1)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Difficult to find SNOMED matches for all
  archetype terms
• Archetype modelers do not always have SNOMED in
  mind when modeling
• No clear guidelines for categorising archetypes
  Observation, Act, Evaluation, Instruction
• Cannot be strict about categories in archetypes
  and their inheritance by child nodes. E.g.
  Observation archetype Complete blood picture
  has Haemoglobin with IM finding or
  procedure.
• SNOMED categories are not Gold Standard.

20
ISSUES .. (2)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Ambiguous Intended Meaning
• E.g. (1) Blood film archetype
• Complete blood picture hasElement haemoglobin
  (hasDescription mass concentration of
  haemoglobin)
• Intent Haemoglobin (substance) or haemoglobin
  concentration (procedure)?
• E.g. (2) Autopsy Examination archetype
• Internal examination hasElement System hasElement
  Gastro-intestinal system (hasDescription Findings
  about oesophagus, peritoneum, bowel, liver
  (including gallbladder) and omentum)
• Intent Gastro-intestinal system (body structure)
  or findings in gastro-intestinal system
  (findings)?

21
ISSUES .. (3)
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Post-coordination
• Needs help but can do about half
  unaided -gtDressing/undressing (Barthel Index)
• Independent (on off, dressing, wiping) -gt
  Toilet Use (Barthel Index)
• Grimace and cough/sneeze during airways
  suction -gt Reflex response (Reflexes)
• Occupation risk if drowsy -gt Alert (Alert)
• Reflex possibly present - markedly reduced -gt
  Biceps (Reflexes)
• Is it possible to post coordinated these terms
  with some degree of clinical sense!

22
Discussion
Semantic Mining Conference on SNOMED-CT Oct 1-3
2006

• Methodology is scalable. Theoretically, can be
  applied to other data models and terminologies.
  E.g. HL7 v3 messages.
• It is not sufficient to have lexical lookups to
  determine appropriate mappings. Semantic
  procedures are also needed.
• Reliable external resources should be used
  wherever possible to augment the performance of
  the system.
• Intelligence added to middleware such as IMs
  also help improve performance significantly and
  can alter results.

23

• Thank you. Questions?

Rahil Qamar, Alan Rector Medical Informatics
Group Department of Computer Science University
of Manchester Manchester, U.K. qamarr_at_cs.man.ac.uk
, rector_at_cs.man.ac.uk
24
Semantic Results

• Blood gas assessment
• PvO2 (context Venous) -gt Mixed venous oxygen
  concentration measurement (procedure) (250558000)
• -gt Measurement of venous partial pressure of
  oxygen (procedure) (250547009)
• -gt Measurement of mixed venous partial pressure
  of oxygen (procedure) (167027009)
• Visual acuity
• Perceive light (context Visual acuity) -gt Visual
  acuity perception of light - inaccurate
  projection (finding) (264943005)
• -gt Visual acuity, no light perception (finding)
  (63063006)
• -gt Visual acuity perception of light - accurate
  projection (finding) (264944004)