1. Simplifying Semantic Solutions
for Biomedical Applications:
San Diego Semantic Web Meet-up
Eric Little, PhD
Chief Knowledge Engineer
Eric.Little@CTG.com
2. Current Biomedical Ontologies Exist in 3
Modes
Metaphysically-based category
systems that can properly classify
biomedical entities & relations.
Improperly structured coding
systems (ICD-9/10, LOINC,
EHRs, etc) where little care has
been taken in providing the proper
formal standardization.
Ontology development tools used
for implementing models and their
accompanying GUIs.
4. Promoting Fundamental Change
Decisions based upon
Medical Informatics
Decisions based upon
Claims and Codes
ProactiveReactive
Rich Internet Experience &
Self Service
Static or limited delivery
Current Technology ApproachesCurrent Technology Approaches New Technology ApproachesNew Technology Approaches
Example: Current systems rely on things like ICD-9 and CPT codes
- so patients already possess the chronic condition (past the threshold for a disease)
- it is hard to objectively judge treatment efficacy
The need is to proactively recognize the complexities that lead to chronic disease and
use objective measures to judge appropriate courses of action.
5. Ontology and Advanced Business
Intelligence
• CTG provides a combination of innovative technologies that deliver
unique medical insights into:
– New Research Areas
• Reuse & repurposing of knowledge
– Disease Registries
– Interventions
– Physician Feedback
– Outcomes Analysis
– Treatment Efficacy
– Provider Efficacy
– Cost Effectiveness
– Waste Reduction
OntologyReasoning
Engines
Semantic
Web
7. Ontologies should provide “actionable intelligence” across
an entire organization
• Knowledge Specification – Provide formal definitions of all items/relations
in a domain.
– E.g., A substance such as Alpha-fetoprotein exists in high levels in
healthy fetuses and infants and plays an important role in their health
and development. High levels in adults is an indicator of liver cancer.
Thus the same protein can be positive or negative based on age.
• Knowledge Elicitation – Provide search/query capabilities to better find
and utilize pertinent information.
– E.g., What are all the known compounds that can act as an SSRI drug?
Which treatments have proven the most effective for a particular
disease type?
• Knowledge Transfer – Provide improved communication between 1)
human-to-machine, 2) human-to-human, 3) machine-to-machine.
– E.g., Improves understanding of complex things like disease states and
provides added insights – requires linking disparate knowledge from
different experts.
Traditional technologies can prove difficult to provide these kinds of capabilities.
Ontology-based systems can help to integrate people through the use of a
common semantic framework.
8. Capturing Medical Knowledge in the Ontology
(Knowledge of Ranges Rules)
• The normal ranges for hemoglobin depend on the age and, beginning in adolescence, the gender
of the person. The normal ranges are:
• Newborns: 17-22 gm/dl
• One (1) week of age: 15-20 gm/dl
• One (1) month of age: 11-15gm/dl
• Children: 11-13 gm/dl
• Adult males: 14-18 gm/dl
• Adult women: 12-16 gm/dl
• Men after middle age: 12.4-14.9 gm/dl
• Women after middle age: 11.7-13.8 gm/dl
9. Merging Independent Domain Ontologies
• Ontologies provide:
– Relations between
blood chemistry,
medical conditions and
symptoms
– Relations between
diseases, diagnoses
and treatments
– Knowledge integration
across a broad
spectrum of
applications
– A common medical
semantics – gives
meaning to codes and
integrates the
information contained in
those codes
Integrate
Integrate
Integrate
Diagnosis
Ontology
Disease
Ontology Allows
inference
across
various
domains
Cell
Ontology
LOINC
Ontology
10. Diagnosis
SNOMED-CT ICD9
Ontology Storyboard for Medical Informatics
Physical Test Results
LOINC
Medical Conditions
SNOMED-CT
Medical Claims
ICD9 - CPT
EHR
Disease Risk Registry
SNOMED-CT
Disease Registry
SNOMED-CT
Treatment
SNOMED-CT ,CPT
Co-morbidities
Disease
Model
Example of Utilizing MedMaP for
Liver Cancer Screening
LOINC CODE
1834-1
AFP Ser-mCnc
ICD-9
155
Elevated AFP
Age Ontology:
Adult Male
Disease Risk Ontology:
Hepatocellular
Carcinoma (HCC)
ICD-9
155
Foundational Model
Human Anatomy
Ontology:
LIVER – Yolk Sac
Cell Ontology:
Liver Cells Dendrite Cell
Ontology:
Immune Cells
Patient Cohorting
Gene Ontology:
Chromosome 4
Gene ID - 174
CPT 47135:
Liver Transplant
Pharmaceutical
Treatments
Drug Bank Ontology:
Sorafenib
Hepatitis B
Hepatitis C
Patient Age, Sex, etc
11. Objectives of Ontology-based Biomedical Systems
• Find new insights
• Provide new knowledge and insights by running inference engines (the system itself should multiply
the knowledge)
• Establish and apply business rules which entail a higher level of reasoning
• Capture SME expertise and deliver globally
• Establish a model that captures SME approach and science (Establish a collaborative community)
• Insulate the common user from the complexity unless it is requested
• Access to knowledge should not require IT intervention
• Allow user to customize their experience to meet their specific needs and objectives
• Portal to allow for customized experience
• Multi-perspectivealism should allow multiple views or renderings of the same data
• Also support static views to dictate or direct useable experience
• Deliver intelligence in an intuitive easily consumable format
• Ease of Use = High Adoption
• Interactive and tactile feel tp allow the user to experience and interact with the ABI system
• Rendered / skinned in format familiar to users surroundings
• Utilize Common Semantics that are relevant to the user
• Health Care Provider – Hypertensive (140 systolic / 90 diastolic) is more common to Patient as High
Blood Pressure
• Model must support synonyms
• Educate
• Allow users to gain insights to SME expertise, and reasoning results
• News feeds, research, sparql end points, lexicons of relevant terms, video
• Provide Evidence to support findings
• Ability to drill into reasoning and scoring and expose business rules, raw data, formulas, and
foundations of knowledge
• Make intelligence “actionable”
• Allow people to take action with the knowledge gained
12. Technology Stack
RDF RDF SchemaEstablish Fact Base
XML XML Schema XSLTStandard Structure&Syntax
OWL–DL Ontology
Model Concepts into
Knowledge Base
REASONER
Provide Reasoning and
Inference
Rule
Execution
Take Action based on
Knowledge
IncreaseIntelligence
13. Layered Inference based upon
Medical Informatics
Medical Concepts & Facts
Medical Conditions
& Disease Risk Models
Disease Registries
& Recommend treatments
Interventions &
Outcomes & Efficacy
IncreasesCapability
Claims & Labs
Establish Fact Base
Standard Structure&Syntax
Model Concepts into
Knowledge Base
Provide Reasoning and
Inference
Take Action based on
Knowledge
15. Knowledge Management Technologies for
Drugs/Immunizations
• Utilizes
federated sub-
ontologies to:
– Classify types of
diseases treated
– Classify active
and inactive
ingredients
– Classify potential
side effects
-e.g.,
contraindications
– Classify dosages
– Classify
chemical
composition/for
mula
– Classify
pathways
Class Hierarchy
Product of Interest
(Pentacel Immunization)
Graph of Items in
The Ontology
Web Search Capabilities
On Any Item in Ontology
(e.g., Pentacel)
16. Advanced Research Capabilities
• Information can be manually searched or automatically linked (via
RSS feeds)
• Allows researchers to find the information they need very rapidly.
• Information gathered through these inputs can be quickly ingested
into the ontology
– improves the knowledge base over time
Specific Research
Article
Details of the
Article
On-line Resource
17. Graphical Visualizations of Important
Relationships in the Data
• Different data sets can be graphically related to one another.
• Allows for rapid insights into new kinds of relationships.
• Can be used to represent individual records (e.g., indiv. patient data)
versus group data (e.g., groups of patients classified into categories –
morbidly obese, clinically obese, etc.
Shows Overlap of
Different Categories
18. Building a Disease Registry
• Proactive Identification of a Disease Risk Constituency allows a Payer to take action
(intervene or educate) and perform preventive maintenance.
• Abilities for Predictive Modeling of Disease Risk, etc.
19. Inferring Patient Complexity, Conditions, Disease & Co
Morbidity
• The Medical Management ontology utilizes an inference engine auto classifies based upon
industry accepted reference ranges and intelligent rules to determine complexity factors such
as conditions, disease risk, etc.
• Blue Lines in the diagram below identify knowledge learned (inferred) from the inference
engine.
20. Rapid Researching of Disease Management Related Issues
• Allows one to quickly identify and apply
the most available research on a given
topic.
• Information can be manually searched
• Information can be automatically linked to
the ontology (RSS feeds).
• Allows the knowledge base of diseases
(and related symptoms/blood chemistry
readings) to continuously grow over time.
• Knowledge of disease becomes an asset
over time.
21. SELECT ?lotno ?plrcount ?scocount ?sumcount ?factor
WHERE {
{SELECT ?lotno (COUNT(*) AS ?plrcount)
WHERE {
?plr a PLR:Pallet_Reading .
?plr PLR:lot_no ?lotno .
} GROUP BY ?lotno
}
OPTIONAL {
{SELECT ?lotno ?scocount
WHERE {
?sco a SCO:Lot_Score .
?sco PLR:lot_no ?lotno .
?sco SCO:hasPalletCount ?scocount
}
}
}
OPTIONAL {
{SELECT ?lotno (SUM (?count) AS ?sumcount)
WHERE {
?s a SUM:Lot_Summary .
?s PLR:lot_no ?lotno .
?s SUM:hasSummaryCount ?count
} GROUP BY ?lotno
}
}
LET (?factor := ?sumcount / ?plrcount)
} ORDER BY ?lotno
CONSTRUCT {
?uri a SUM:Hourly_Summary .
?uri PLR:lot_no ?lotno .
?uri PLR:captureevtdate ?date .
?uri PLR:hasprocessinghour ?hr .
?uri CLS:hasClassifyGroup ?classifyGroup .
?uri CLS:hasClassifyProperty ?classifyProp .
?uri CLS:hasClassifyCategory ?classifyCategory .
?uri SUM:hasSummaryCount ?LotHrCount .
}
WHERE {
{
SELECT ?lotno ?date ?hr ?classifyGroup ?classifyProp ?
classifyCategory (SUM(?Count) AS ?LotHrCount)
WHERE {
?sum a SUM:Hourly_Summary .
?sum PLR:lot_no ?lotno .
?sum PLR:captureevtdate ?date .
?sum PLR:hasprocessinghour ?hr .
?sum CLS:hasClassifyGroup ?classifyGroup .
?sum CLS:hasClassifyProperty ?classifyProp .
?sum CLS:hasClassifyCategory ?classifyCategory .
?sum SUM:hasSummaryCount ?Count .
} GROUP BY ?lotno ?date ?hr ?classifyGroup ?
classifyProp ?classifyCategory
} .
LET (?uuid := smf:generateUUID()) .
LET (?uri := smf:buildURI("<http://www.ctg.com/SUM#{?lotno}_{?
date}_{?hr}_{?uuid}>")) .
}
Examples of Complex SPARQL Queries
25. How Knowledge is Delivered – Medical
Management Portal (MedMaP)
• Support access to multiple analytical tools (portlets)
• Users customize their experience to meet their specific objectives
• Support multiple perspectives (Different views for different people)
• Allow users to self help – self serve
• Educate
• Allow users to gain insights to SME expertise, and
reasoning results
• News feeds, research, sparql end points, lexicons of
relevant terms, video
• Support static views to dictate or direct the most efficient
experience
26. Graphical Visualizations of Important Relationships in the
Data
• Different data sets can be
graphically related to one
another.
• Allows for rapid insights into
new kinds of relationships.
• Can be used to represent
individual records (e.g.,
indiv. patient data) versus
group data (e.g., groups of
patients classified into
categories – morbidly
obese, clinically obese, etc.
Yellow Balls Represent
Individual Patients
Purple Balls Represent
Individual Patients
With Added Attribute of
HYPERGLYCEMIA
27. Cohort Populations: Identifying Critical Relationships
within the Ontology
• Information from disparate
sources can be rapidly
compared using industry
standard classifications.
– Can relate information
across different medical
areas.
• The system can be used to link
blood chemistry values to one
another (e.g., glucose readings
& BUN scores).
• It can link those values to other
items (e.g., patient attributes –
gender, Body Mass Index (BMI),
etc.)
29. Therapeutic Optimization for Disease Management
• Example: Effective Treatment
Management
– Prescribing large doses of epoetin
alpha for low levels of hemoglobin
(Anemia) in CKD patients.
– Outcomes analysis can be used to
determine whether this is an
appropriate course of action.
30. Therapeutic Optimization for Disease Management
• Example: Potentially Ineffective
Treatment Management
– Prescribing Large Doses of Epoetin
for normal Hemoglobin.
– Can be used to identify when a
treatment is not warranted based on
empirical evidence (outcomes).
– Resulting Question: Why are large
and costly doses of epoetin alpha
being given to patients with normal
hemoglobin levels?
Editor's Notes
If you cannot take action with the intelligence what is the point – this is usually not science for the sake of science
If you cannot take action with the intelligence what is the point – this is usually not science for the sake of science