Knowledge Engineering and SemiAutomatic Population of Medical Ontologies Using NLP Methodologies

1
Knowledge Engineering and Semi-Automatic
Population of Medical Ontologies Using NLP
Methodologies

• Munich 11.06.2007
• Pinar Oezden Wennerberg
• pinar.oezden_at_jrc.it

2
Agenda

• Knowledge Engineering and Ontology
• Definitions, methodologies, guidelines
• Medical Terminology and Natural Language
  Processing (NLP)
• The problem of medical terminology
• The context users, tasks, types of information
  in the medical domain
• The role of NLP and knowledge engineering
• Motivation for Semi-Automatic Ontology Population
• The knowledge acquisition bottleneck
• Vast amount of knowledge available in (un- /
  semi-)structured text, WWW, databases etc.
• One example approach
• Ontology population via Supervised Machine
  Learning (ML)
• Challenges

3
Knowledge Engineering and Ontologies

• Some Definitions
• Humans and software agents need knowledge about
  the world in order to reach good decisions
• Such knowledge is typically stored in
  knowledgebases
• Knowledge engineering is the process of building
  a knowledgebase
• A knowledge engineer is someone, who
• investigates a particular domain,
• determines what concepts and relations are
  important in that domain,
• and creates a formal representation of objects
  and relations in that domain.
• (Russel Norvig, 1995)

4
Knowledge Engineering and Ontologies

• An ontology specifies a finite, controlled,
  extensible and machine processable vocabulary for
  a given knowledgebase
• Consists of concepts, properties, relations,
  axioms
• Knowledge engineering guidelines
• Decide what to talk about and on the vocabulary,
• Encode general knowledge and a specific problem
  case
• Execute queries and verify inference
• (Russel Norvig 1995)

5
Medical Terminologies and Natural Language
Processing (NLP)

• Problem statement
• Numerous heterogenious medical terminologies and
  coding schemes exist that need to interoperate
• e.g. Systemized Nomenclature of Medicine (SNOMED)
  for coding paptient notes, ICD (International
  Classification of Diseases), ICD-9-CM for billing
  purposes,RIZIV, IDEWE, ICPC-2, ATC etc.
• Existing efforts UMLS, Galen, MeSH, etc.

6
Medical Terminologies and Natural Language
Processing (NLP)

• Definition of context
• Information types to be collected are about
• Individuals (e.g. medical records)
• Groups (e.g. data about epidemiology, public
  health)
• Institutions (e.g. planning, management in
  hospitals, clinics)
• Domain specific knowledge (e.g. state-of-the-art
  publications, proceedings)
• Domain relevant tasks
• Data entry, query and retrieval about patients
• Information sharing and integration from
  different applications and medical records

7
Medical Terminologies and Natural Language
Processing (NLP)
Information Extraction
Knowledge Representation and Reasoning

• Question
• Answering

Natural Language Processing
Machine Learning
Information Retrieval
Knowledge Discovery, Text Mining
Ontology Engineering
Adapted from Jena University www.julielab.de
8
Motivation for Semi-Automatic Ontology Population

• The knowledge acquisition bottleneck
• Ideally the knowledge engineer interviews the
  knowledge expert to get educated about the domain
  i.e. to acquire knowledge
• ? expensive in time and resources
• ? domain experts not alwaysavailable
• Availability of vast amount knowledge
• In resources such as medical databases, journals,
  publications, conference proceedings, medical
  reports etc.
• World Wide Web

9
Ontology Population via Supervised Machine
Learning

• Problem statement
• Identify and extract relevant knowledge (terms,
  phrases, relations, facts) in text e.g.
• Terms health disorder, malfunction,
  sickness, illness, maladie, Krankheit ?
  Disease
• Smoking causes cancer ? ltSmoking, Cancergt
• Goal
• Assign them to the appropriate concepts of the
  ontology as instance
• Concept Disease
• Relation causes

10
Ontology Population via Supervised Machine
Learning

• Processes
• Annotate (i.e. supervised)
• ltCAUgtSmokingltCAU/gt ltCAU-Rgtcauseslt/CAU-Rgt
  ltDISgtcancerlt/DISgt
• CAU DiseaseCause, CAU-R causalRelation, DIS
  Disease
• Learn and extract from a training set (i.e. ideal
  world)
• Extract from the test set (i.e. unknown world)
• Apply the learned rules on new documents to
  discover and extract new knowledge

11
Ontology Population via Supervised Machine
Learning

• Learn and extract from a training set (i.e. ideal
  world)
• Recognize syntactic constructs such as NPs, VPs,
  PPs
• Generate extraction rules
• Rule for concept Disease
• Disease- ltNP smokinggtltVP causesgtltNP DIS gt
• Rule for concept DiseaseCause
• DiseaseCause- ltNP CAUgtltVP causesgtltNP cancer
  gt
• Rule for relation causalRelation
• causalRelation- ltNP smokinggtltVP CAU-RgtltNP
  cancer gt
• Classify
• Disease cancer
• DiseaseCause smoking
• causalRelation causes

12
Ontology Population via Supervised Machine
Learning

• Possible problems
• More than one value was extracted for a given
  relation
• Entities from different classes were extracted
  (multiple concept assignment i.e. ambiguity)
• Nothing was extracted
• Possible solutions
• Present the user all possible values, let the
  user decide
• To assist user with the decision process by
  assigning confidence scores to possible values
• i.e. how much does the system believe what it
  suggests is relevant/true
• Provide context information via text highlighting
  to justify the systems confidence
• Provide empty data entry slots for users to enter
  their knowledge

13
Challenges

• General challenges
• It is difficult to eliminate the knowledge
  acquisition problem entirely
• Due to the sensitivity of the domain (human
  health) the knowledge experts cannot be
  completely avoided
• Computer scientists need to work together with
  domain experts to a certain extent
• Systems should be able to be used by
  non-technicians
• Multilinguality
• Healthcare workers, patients, administrators
  should be able to have access to information in
  their own language

14
Challenges

• Knowledge/ontology engineering specific
  challenges
• Implicit information (typical for natural
  language) i.e. not machine-processable (not
  explicit)
• Different levels of detail (granularity) is
  required to meet different expectations
• i.e. provide sufficient detail but abstract away
  irrelevencies
• Poly-hierachies to support multiple views
• may lead to ambiguities, contradictions
• Adaptability, extensibility for changing user
  demands and for standards
• Expressibility vs. computational tractibility
• Achieving consensus between practitioners

15
Questions?

• Evaluation
• How do we know if we have a good system?
• Practitioners to evaluate the effficiency and
  reliability of the developed systems?

16

• Thank You!