Hybrid Representation of Procedural Medical Knowledge: The DeGeL Architecture

1
Hybrid Representation of Procedural Medical
KnowledgeThe DeGeL Architecture

• Yuval Shahar, M.D., Ph.D.
• Medical Informatics Research CenterDepartment of
  Information Systems EngineeringBen Gurion
  University
• Beer Sheva
• Israel

2
Medical Decision-Support Systems

• Support multiple different tasks, such as
• Diagnosis
• Therapy
• Clinical and pharmaceutical (drug design)
  research
• Information visualization, summarization, and
  exploration
• Although specific clinical areas can benefit from
  diagnostic support, the current emphasis has
  shifted to therapy, in particular automated
  support at the point of care, based on clinical
  guidelines
• We will focus on automated support to
  specification and use of clinical guidelines

3
Clinical Guidelines

• A standard of care, typically an experts
  consensus
• Usually specify diagnostic and/or therapeutic
  procedures
• Also known as clinical protocols (e.g., in
  oncology) or care plans
• A powerful method to standardize and improve the
  quality of medical care Grimshaw and Russel,
  1993.
• Increasingly widespread use, to spur best
  practices in medical care and to incorporate
  evidence-based medicine
• Focus on chronic patients Diabetes, Asthma,
  Hypertension, Cardiology, Oncology, AIDS,
  Depression
• Computer-based techniques needed to automated the
  support of multiple tasks involved in
  guideline-oriented medical care

4
Guideline-BasedCare Support Tasks

• At design time
• verification of the guideline process
  specification (syntax)
• validation of the guideline versus its goals
  (semantics)
• At execution time
• determination of patient eligibility and
  guideline applicability
• visualization of one or more potentially
  applicable guidelines
• application (execution) of the guideline
• quality assessment of providers actions
• modification of guideline or provider plans
• evaluation of guideline effectiveness

5
Characteristics of Automated Support
toGuideline-Based Care

• Dialog Care provider ? automated support
  system
• Both have relative strengths
• Care provider
• Better access to clinical data of the patient
• Broader medical knowledge
• Automated system
• Better access to guideline specifications
• Faster and more accurate computation and
  detection of complex patterns
• The aim is synergy

6
Automated Support for Clinical Guidelines
Examples of Prescriptive approaches

• DILEMMA, PRESTIGE, Proforma, Prodigy (UK/EU)
• Oncocin, T-Helper, EON (Stanford)
• Arden Syntax/MLMs (Columbia/LDS)
• GLIF (Columbia, Harvard, Stanford)
• ActiveGuidelines (Epic Systems Co., USA)
• The Pavia GUIDE project (Italy)

7
Automated Support for Clinical GuidelinesExample
s of Critiquing Approaches

• VT-Attending (Miller, Yale)
• HyperCritique (Van der Lei and Musen, Rotterdam)
• The Asgaard project (Stanford, Vienna, London,
  BGU) integrates both prescriptive and critiquing
  approaches by representing both the prescribed
  interventions and the underlying process and
  outcome intentions

8
The BGU/Stanford/Vienna/UK Asgaard
Project(Shahar, Miksch, and Johnson, AIM 1998)

• A task-specific framework for the representation,
  application, and quality assessment of
  time-oriented clinical guidelines
• Uses the highly expressive Asbru
  guideline-specification language
• Enables explicit representation of process and
  outcome intentions
• The quality-assessment algorithms try to explain
  care-provider intentions given their actions, the
  intentions of the guideline they are applying,
  and a domain-specific knowledge base
• Includes the DeGeL hybrid guideline server at
  BGU, which allows access to and manipulation of
  guidelines represented as free text,
  semi-structured text, and formal Asbru

9
Intention-Based Quality Assessment

• Guideline prescribed actions ltgt applied
  (observed) actions
• Guideline intended plan ltgt pattern of
  applied actions
• State intentions of guideline ltgt state
  intentions of provider
• State intentions of guideline ltgt abstracted
  patient state
• State intentions of provider ltgt abstracted
  patient state
• gt A comparison (critiquing) vector is a behavior

10
Example Behaviors

• Everything goes according to plan
• - - The guideline does not work
• - Process intention achieved by other
  actions
• - - Outcome intention achieved by
  another plan
• - - - - - Process and outcome intentions
  not achieved

11
A Plan-Recognition and Critiquing Example (I)
Patient state 2 weeks of severe
anemia Protocol outcome intention
Avoid more than 2 weeks of severe
anemia Protocol process recommendation
Decrease dose of drug toxic to bone
marrow Physician's action Transfusion
of two units of packed red blood cells  
12
A Plan-Recognition and Critiquing Example (II)

• Analysis
• Both methods compatible with elevating
  hemoglobin level
• 1. Decrease a plan argument inversely
  related to it
• 2. Initialize a plan that increases it
  directly
• Conclusions
• a. The protocols outcome intention is
  achieved by another plan
• b. Next therapy session, monitor transfusion
  complications

13
Asbru Guideline Conditions

• All conditions are temporal patterns for
  example
• Filter preconditions need to be true at start of
  a guideline (e.g., patient is pregnant and has
  gestational diabetes)
• Setup preconditions need to be achieved to enable
  a guideline (e.g., patient has undergone a
  glucose tolerance test)
• Completion conditions specify when the guideline
  application is completed (e.g., patient has
  undergone a baby delivery)
• Abort conditions specify when the guideline
  application needs to stop immediately (e.g., an
  inadequate response to a dietary regimen)

14
Asbru Guideline Intentions

• Intentions temporal-constraint patterns
• Each pattern is meant to be maintained, achieved,
  or avoided
• Two main categories of intentions might annotate
  a guideline
• Outcome intentions (refer to patient states)
• E.g., achieve diastolic blood pressure of 90 or
  less within 1W
• Process intentions (refer to provider actions)
• E.g., use ACE inhibitors, or Avoid use of
  thiazides

15
The Digital Electronic Guideline Library (DeGeL)

• Uses hybrid (multiple-format) electronic
  representation of clinical guidelines to cater
  for different specification skills of the editors
  and different needs of each guideline-based-care
  task
• Allows distributed Web-based access for knowledge
  acquisition, maintenance, retrieval, and
  application
• Includes tools for distributed performance of
• Upload of source documents semantic indexing of
  a guideline (IndexiGuide)
• semantic markup/specification using multiple GL
  ontologies (Uruz, Gesher)
• Context-sensitive search and retrieval (Vayduria)
• Visualization (VisiGuide)
• Eligibility Applicability determination
  (Dipole)
• Runtime application (Spock)
• Quality assessment (QualiGuide)
• Supports a multiple-group permission model for
  knowledge access

16
Hybrid Representations

• A hybrid representation supports gradual
  conversion of free-text guidelines into a formal
  executable representation
• A hybrid framework represents guidelines using
  three formats
• Free text
• Semi-structured representation, using the
  semantic roles of the formal language
• Semi-formal representation, which includes
  control knowledge
• Fully structured, formal, executable language
  (e.g., Asbru or GLIF)
• In DeGeL, The guideline is first converted to a
  semantically semi-structured text by a medical
  expert using a markup editor, then to a
  semi-formal representation with the assistance of
  a knowledge engineer, and finally to a formal
  representation by a knowledge engineer (the
  default target language being Asbru), thus best
  using each expert type
• Each of the application tools (e.g., search,
  application) is designed to handle all three
  representations

17
Incremental Hybrid Guideline Specification
18
The DeGeL Architecture
19
(No Transcript)
20
(No Transcript)
21
Uploading a New Guideline
22
(No Transcript)
23
The IndexiGuide Classification Tool

• Enables medical experts performance of
• Classification of a new guideline along a
  multiple hierarchy of several semantic axes
  (e.g., diagnosis, therapy), and definition of
  their type (e.g., screening)
• Operates dynamically, by reading semantic-axes
  representation tree, thus supporting indexing and
  markup using any external semantic-axes hierarchy
• Operates at the guideline meta-representation
  level, that is, independent of any particular
  guideline format
• Might use in the future an automated tool to
  suggest multiple hierarchical classifications,
  based on our machine-learning work

24
Classifying a Guideline
25
The Uruz Semantic-Markup Tool

• Enables medical experts performance of
• Creation and semi-structuring of new guidelines
  de novo
• Semantic mark up of existing free-text guidelines
  into structured text, by allowing the experts to
  label (and then modify) parts of the text using
  the semantic tags of the selected target
  guideline-representation language (e.g., Asbru,
  GEM)
• Operates dynamically, by reading the structure
  tree, thus supporting indexing and markup using
  any format
• A client-based graphical version, Gesher, is
  being developed as well

26
Semantic Markup Filling a Semantic Role
27
The Plan-Body Wizard A Sequential-Plan Form
28
The Vayduria Search, Retrieval, Visualization
Tool

• Enables the user to search within
• The structure of the semantic-classification
  axes one or more concepts can be selected in any
  of the indexing axes (e.g., Dx, Tx)
• the structure of the chosen semi-structured
  representation format (e.g., Asbru, GEM)), which
  enables querying a marked guideline for the
  existence of one or more terms in a particular
  context (e.g., eligibility conditions)
• Enables visualization, browsing, and exploration
  of the search results, exploiting the original
  semantic classifications, and the internal
  guideline structure, to determine which guideline
  is most relevant, by using the VisiGuide tool
• Operates dynamically, thus supporting search
  using any semantic-axes hierarchy, or any
  guideline ontology

29
Searching Using Only The Source
30
Search Results Using Only the Source Text
31
Search Using The Semantic Axes
32
Search Results Using The Semantic Axes
33
Adding The Context-Sensitive Search
34
Results of Adding Context-Sensitive Search
35
Browsing The Semantic Indices of a Result
36
Quick-Viewing of The Search ResultsBrowsing
the Source
37
Quick-Viewing of The Search ResultsBrowsing a
Semantic Role
38
The VisiGuide Guideline-Browsing Tool
39
The Need for Standard Schemas and Terms

• Guidelines cannot be (re)used at multiple local
  sites without well-defined standards for
  representing the knowledge
• Problems in (local) knowledge reuse
• The local database schema is unknown.What are
  the names of the tables and what are the columns?
  Where is each clinical concept?
• The local vocabulary is unknown.How are various
  types of hemoglobin measures referred to? (e.g.,
  "hgb", or any other arbitrary string)
• The local measurement units might be unknown, and
  might even be non-standard. (e.g. mg/dl instead
  of gr/l)
• Our Solution The Medical Database Adaptor
  (MEIDA) system includes tools and vocabulary
  servers for mapping all three aspects of the
  local database to international standards such as
  LOINC, ICD, CPT, and HL7

40
The LOINC Server Search Engine
41
LOINC Search Results
42
Summary

• Automated support to the use of clinical
  guidelines is a major informatics focus
• major issues
• Grounding of guidelines in the terms of shared
  vocabularies
• Facilitation of authoring, search, application,
  and quality assessment
• Sufficient expressiveness to capture process and
  outcome intentions
• Integration at the point of care
• The BGU DeGeL framework and software architecture
  uses a hybrid representation (free text,
  semi-structured, semi-formal, and formal formats)
  to facilitate markup classification of
  guidelines by physicians, formal specification by
  knowledge engineers, and retrieval use of the
  guidelines by clinical end users
• Most of the tools are ontology independent
• Encouraging preliminary assessments of all tools
  formal evaluations in process