Formalization of guidelines exploiting medical thesauri

1
Formalization of guidelines exploiting medical
thesauri

• Radu Serban, Annette ten Teije
• Vrije Universiteit Amsterdam
• Protocure project www.protocure.org
• serbanr,annette_at_few.vu.nl
• 8 April 2006

2
Objective Improving quality of Medical
Guidelines by Structured GL Development

• Change management for Evidence-based Guidelines
• systematic review of safe practice, updates due
  to clinical trials
• quality of the guideline document
• Problem Maintaining clinical GLs is expensive
  and ineffective
• typically, small updates concerning specific
  medical knowledge
• we lack computerised objects reflecting changing
  medical knowledge
• Solution use GL formalization to make explicit
  modular organization of medical knowledge and
  support structured GL development
• GL program GL development SW development

3
Context Guideline Formalization

• PROCEDURES Diagnosis Treatment Supervision
  Critical situations
• BACKGROUND
• Terminology, Therapies
• Normal values for parameters
• Decisions, preferences
• CONCLUSIONS
• (A1)In the event of local recurrent breast
  cancer in a previously irradiated area, the
  treatment of choice is low-dose re-radiation with
  hyperthermia.
• RECOMMENDATIONS
• In the event of isolated local recurrence
  following MST, salvage mastectomy is recommended.
• CONSIDERATIONS
• Contraindications for special cases.

• GL Quality
• Consistency
• Exactness
• Completeness

4
Structured GL Formalization
5
Context and Objectives

• Context Formalizing medical guidelines (GLs)
• Exploiting domain knowledge from medical thesauri
• Objective Make GLs modular and maintenable by
  exploiting domain knowledge in medical thesauri
• Knowledge objects obtained by mapping control
  templates and engineered linguistic templates
• Establish categories of knowledge conveyed by
  guidelines
• Extract and organize these categories as
  knowledge templates
• Evaluate usefulness of the templates in guideline
  formalization
• Research questions
• What kind of regularities can be identified?
• what knowledge do they capture?
• how to identify and describe knowledge templates?
• How can knowledge objects be used for GL
  authoring and formalization?
• what knowledge shared by guidelines is used in an
  executable model of a GL?
• is this knowledge covered by existing medical
  thesauri?

6
Pre-requisites for formalizing GL based on
knowledge objects

• Medical terms typically used in GLs are common
  medical terms described in medical vocabularies
• Hypothesis1 Controlled vocabularies cover domain
  knowledge shared by guidelines
• The medical knowledge used to elaborate a GL is
  present in GL text
• Hypothesis 2 Enrichment of medical knowledge,
  semantic tagging of GL content and mapping of
  domain knowledge to linguistic templates is
  feasible
• Control knowledge is abundant in GL text and can
  be recognized using linguistic regularities
• Hypothesis 3 Essential control knowledge is
  present in the guideline text and is well
  structured

7
Assumption 1 Medical terms typically used in GLs
are covered by a controlled vocabulary

• Experiment
• 3 breast-cancer guidelines (CBO,SIGN,RCR)
• Relevant terms (TextToOnto)
• SIGN 174 terms CBO 267 terms RCR 190 terms
• UnionSIGN u RCR u CBO 394 terms
• IntersectionSIGN n RCR n CBO 70 terms
• Extract medical terms in existing thesauri
• MeSH 155000 terms NCI 27700 terms MeSH n
  NCI 5000 terms
• Look-up in existing medical thesauri
• Union n MeSH 202 terms Union n NCI 144 terms
  Union n MeSH n NCI 120 terms
• Intersection n MeSH 60 terms Intersection n
  NCI 45 terms Intersection n MeSH n NCI 42
  terms
• Conclusion 60 of shared vocabullary is
  controlled vocabulary (up to 85 covered by MeSH)

8
Information in existing thesauri

• MeSH
• DescriptorNameRadiotherapy
• AllowableQualifierNameadverse
  effects,contraindications
• ConceptList ConceptName, ConceptUMLSUI
• SemanticTypeList, SemanticTypeNameTherapeutic
  or Preventive procedure
• ConceptRelationList Relation-Concept1-Concept2
• NCI
• owlClass IDRadiation_Therapy
• rdfssubClassOf rdfsresourceCancer_Treatment
• Semantic Type Therapeutic or Preventive
  Procedure
• Preferred Name, Synonym1Radiotherapy, UMLS_CUI
• Merged
• Radiotherapy same_as Radiation_Therapy
• Radiotherapy isa_kind_of Therapeutic_or_Preventive
  _Procedure
• Radiotherapy subclass_of Cancer_Treatment

9
Assumption 2 Mapping domain knowledge to
linguistic templates is feasible
10
Medical Background Knowledge

• Abstracted from semantic nets UMLS, MeSH, NCI
• Semantic types
• medical specific categories disease, medication,
  med_effect, med_action
• lexical operators relational operators
  (temporal, causal) or action operators
  (decomposition, sequence)

11
Ex Generating templates from domain knowledge

• instanceradiotherapy, produces, skin_reactions
  instance_of
• templatemed_action, effect_op, med_effect
  covers
• o_fragment(MedAction produces MedEffect)

12
Assumption 3 Control knowledge is abundant in GL
text and can be recognized using linguistic
regularities
13
Method 1/3

• Building GL components
• reconstructs formalizable procedural knowledge
  which produces lexical regularities in the text
• templates generated using medical ontology
  linguistic regularities
• Observations
• conclusions,recommendations have modular
  structure
• In the event of MedContext, the treatment of
  choice is Treatment.
• In the event of MedContext, Treatment is
  recommended.
• Extraction patterns for procedural knowledge help
  formalization
• Assumptions
• guidelines share a controlled vocabulary
• procedural medical knowledge used for operational
  model corresponds to linguistic regularities and
  can be formalized using knowledge from thesauri

14
Method 2/3

• Guideline
• Requirements(TargetGroup1,TreatmentPath1),
• Definitionsltterminologygt
• Procedural-Narrativeslthow-to knowledgegt
• Conclusions-Referencesltexternal knowledge
  sourcesgt
• RecommendationsltGL commitments wrt Requirementsgt
• Argumentation

15
Method 3/3

• Mapping GL Ontology to UMLS Classes
• TargetGroup ? Age_Group, Patient_or_Disabled_Grou
  p, Population_Group
• Medication ? Clinical_Drug
• BodyPart? Body_Location_or_Region
• MedAction? Diagnostic_Procedure,
  Therapeutic_or_Preventive_Procedure
• Disease? Disease_or_Syndrome
• MedContext? Sign_or_Symptom
• Mapping GL Ontology to UMLS relations
• Medication affects BodyPart ? Clinical_Drug
  affects Body_Location_or_Region
• MedAction1 consists_of MedAction2 ?
  Therapeutic_or_Preventive_Procedure2 part_of
  Therapeutic_or_Preventive_Procedure1

16
Observations about thesauri

• (Deep) hierarchy of terms
• But
• Very simple relations
• Little operational info, relations between
  activities
• Not sufficient for deriving semantic relations
• MedAction treats Disease uses Medication affects
  BodyPart
• MedAction part_of Treatment achieves MedGoal
  produces MedEffect
• Still useful
• for populating GL terminologies
• for enriching GL ontology

17
Results 1/3

• Process pattern
• I1 Radiotherapy following anthracycline-contain
  ing chemotherapy
• CT med_action1 seq_act_op med_action2
• FR action1 SEQ action2
• I1 instantiates CT using mapping
• CT translates_as FR using mapping

18
Results 2/3

• Goal pattern
• I1 Attention should be paid to
  side-effectsmed-effect, such as
  skin-reactionsmed-effect
• CT neg_rec_op med_effect inst_op
  med_effect_val
• FR AVOID med_effect med_effect_val
• Testing targets
• ? Avoid or monitor actions that are known to
  produce as side-effects skin reactions.
• ? IF action.statusconsidered AND
  skin-reactions IN effects(action) THEN
  ask-confirmation(action)

19
Results 3/3

• Role of knowledge patterns
• Modules used in modelling medical processes and
  defining constraints
• Based on thesauri knowledge, a more complete
  component can be built
• Detection of terminological problems
  ambiguities, inconsistencies, incompleteness
• Authoring of GLs

20
Lessons learned

• Knowledge templates
• Extracted from guideline using GL ontology
• semantic annotation knowledge-driven templates
• Advantages
• Correlated with background knowledge, easy to
  extend and to validate
• Exploit shared controlled vocabulary, overlap
  between medical terminologies
• Disadvantages
• Most do not have an operational translation
• Validation only by medical expert
• Benefits of modular guideline components for GL
  maintenance
• map GL text to underlying knowledge, ease
  modelling validation of models, improve GL
  modularization
• only components concerned with changing knowledge
  are updated
• Method to identify knowledge objects
• Domain ontology and linguistic regularities guide
  the generation of knowledge templates and search
  for their instances
• Control structures of the target GL language
  determine which knowledge objects are used in
  formalization