Transcript Royal College of Surgeons in Ireland Coláiste Ríoga na Máinleá in

Royal College of Surgeons in Ireland
Coláiste Ríoga na Máinleá in Éirinn
Knowledge representation in TRANSFoRm
AMIA CDSS workshop, 24th October 2011
Derek Corrigan, Borislav Dimitrov, Tom Fahey
PHS / Department of General Practice
Overview
•
Aim to provide overview of TRANSFORM approach to knowledge
representation – provide discussion points
•
Distinguish between clinical knowledge vs. patient data
•
Description of development of ontology to support clinical evidence
•
Examples of how the ontology can be used to support querying data
•
Discuss the benefits of this approach
•
Discuss the challenges and issues encountered using this approach
PHS / Department of General Practice
Clinical knowledge – what do we mean?
•
Patient Data – traditional model focus
– Documentation to support record of patient encounter
– Tends to be historic and static to a point in time in nature
– Data or document presentation focussed
– Existing clinical models traditionally have been EHR focussed
•
Clinical Knowledge
– Clinical facts derived from research data that stands alone and
separate from a patient context
– Dynamically changing as research evolves and develops
– Rule based to implement forms of reasoning PHS / Department of General Practice
The TRANSFoRm Project
PHS / Department of General Practice
TRANSFoRm
Services
1 CPR Repository
Clinical Prediction
Rules Service
2 Distributed
GP EHRs
With CDSS
3 Research Study Designer
Study Criteria
Design
CP Rules
Manager
CP Classifier
5 CPR Data Mining
and Analysis
CPR Analysis &
Extraction Tool
Find Eligible Patient
4 Research Study Management
Recruit Eligible Patient
Study Data Management
PHS / Department of General Practice
TRANSFoRm approach
•
Clinical Prediction Rule – core model structure
– Well defined and has underlying statistical model in the form of
logistic regression models to support electronic derivation from
research data
– TRANSFoRm has potential to address limitations of traditional CPR
development – large populations for derivation, validation
infrastructure, dissemination of CPRs as guidelines
•
Ontology of clinical evidence
– Using Protégé to define an ontology of clinical evidence that
implements CPRs as an evidence interpretation mechanism
PHS / Department of General Practice
Ontology Development Tools
•
Protégé – Ontology Development
•
Sesame Triple Store – provides persistent representation
•
Sesame API – provides for programmatic update/manipulation
•
Ontology will provide a service oriented semantic contract for the
representation of clinical evidence knowledge for other TRANSFoRm
services and software artifacts e.g. provenance, data mining, CDSS
interface
PHS / Department of General Practice
Ontology Data Representation
•
Generic model representation constructs and rule formulation
–
–
–
–
•
Data instance representation
–
–
–
–
•
RDFS (Schema language) and Web ontology language (OWL)
E.g. “EvidenceSymptom” – “isSymptomOf” – “EvidenceDiagnosis”
SWRL (Semantic Web Rule Language) allows definition of complex chained rules
Person (?x1) ^ hasSibling(?x1,?x2) ^ Man(?x2) → hasBrother(?x1,?x2)
Resource Description Format triples (RDF) – “Subject – Predicate – Object”
E.g. “Dysuria” – “isSymptomOf” – “UrinaryTractInfection”
Predicates/relationships are directional in nature
E.g. “UrinaryTractInfection” – “hasSymptom” – “Dysuria”
Distribution format – supports concept composition
–
–
Tagged text file in XML like syntax for easy distribution
Import and reuse other ontologies as building blocks
PHS / Department of General Practice
Example Question: Provide all differential
diagnoses relating to a reason for
encounter ICPC2 code “D01” (abdominal
pain /cramps general)
SELECT ?anyDifferentialDiagnosis
WHERE
{?anyRFE
?anyDifferentialDiagnosis
hasICPC2Code
isDifferentialDiagnosisOf
"D01"^^xsd:string .
?anyRFE .}
EctopicPregnancy
Pyelonephritis UrinaryTractInfection
ChronsDisease
Appendicitis
BowelCancer
IrritableBowelSyndrome
BacterialEnteritis
PHS / Department of General Practice
Example Question: Give me all rule criteria
and cues for all elements of the Little
Symptom Rule for UTI
SELECT ?anyCriteriaElement
WHERE
{?anyRuleElement
?anyCriteriaElement
?anyCueElement
?anyCriteriaElement
?anyProperty
UTI1Crit1
UrineCloudiness
isPresent
1 (True)
?anyCueElement
isRuleElementOf
isCriteriaOf ?
isCueElementOf
?anyProperty
rdf:type
UTI1Crit2
UrineSmell
isPresent
1 (True)
?anyProperty
?anyValue
LittleSymptomRule .
?anyRuleElement.
?anyRuleElement.
?anyValue.
owl:DatatypeProperty. }
UTI1Crit3
Dysuria
isPresent
1 (True)
UTICrit4
Nocturia
isPresent
1 (True)
PHS / Department of General Practice
WP4 – Technical
Architecture
Diagram
EHR Client
Evidence Update via
Research Tools
WP5 Dynamic Interface
Linked to EHR
WT 4.5 Data Mining Process
Clinical Evidence
Web Client
Cached
Client Data
Repository
Management Tools
HTTP
HTTP
WP4 Web Services
WP4 Client Interfaces
Clinical Evidence Web
Service (SOAP/WSDL)
SPRING MVC
Framework
TRANSFORM
Provenance
Service
HTTP
Evidence Update
Service (SOAP/WSDL)
HTTP
HTTP
JAVA Business
Objects
Sesame RDF API
Object / RDF Interface
TRANSFORM
Security
Framework
HTTP
HIBERNATE
Relational to
Object Mapping
TRANSFORM
Vocabulary
Service
Application Layer Objects
TRANSFORM
Shared
Services
SQL Queries
Clinical
Evidence
Repository
My SQL Database
SPARQL Queries
Sesame RDF
Repository
My SQL Database
RDF/XML
OWL/XML
Protégé Ontology
Development
Observations on the ontological approach
•
RDF provides an alternative model approach by reducing data representation to
a very simple form without use of complex reference models – reduced
complexity paradoxically increases power!
•
The addition of RDFS and OWL add a semantic interpretation layer on top of
the data representation that supports composition and merging of diverse data
sources and subsequent inference to generate new facts
•
SPARQL allows for very complex querying using compact data representation
that can be easily be done in ‘two directions’ to support ‘top-down’ analysis or
‘bottom-up’ analysis – works well for diagnostic view of data
PHS / Department of General Practice
Challenges of ontological approach
•
Ontology validation – who arbitrates on the clinical accuracy and completeness
of models? Knowledge governance Vs. Standards governance
•
An ontology is not a working application – development tools are not
application focussed and needs ontological to relational mapping to support
integration with relational data to provide the ‘working’ application context –
duplication of effort?
•
Ontology maintenance is intensive – tools still immature/poorly integrated in
development environments
•
Integration /interoperability with EHR using standards and clinical vocabularies
– granularity/mapping issues e.g. ICPC2
PHS / Department of General Practice
Thank You
Discuss!
PHS / Department of General Practice