Vipul Kashyap vkashyap1@partners Senior Medical Informatician

1. From the Bench to the Bedside:The role of Semantics in enabling the vision of Translational Medicine Vipul Kashyap vkashyap1@partners.org Senior Medical Informatician 1Clinical Informatics R&D, Partners Healthcare System Semantic Web and Databases Workshop, ICDE 2006 April 8th , 2006 * Thanks to Dr. Tonya Hongsermeier and Alfredo Morales

2. Outline • What is Translational Medicine? • Current State • HCLS: A Knowledge Driven Endeavor • Use Case • Functional Requirements • Data Integration • Clinical Decision Support • Knowledge Change and Provenance • Conclusions

3. What is Translational Medicine? • Improve communication between basic and clinical science so that more therapeutic insights may be derived from new scientific ideas - and vice versa. • Testing of theories emerging from preclinical experimentation are tested on disease-affected human subjects. • Information obtained from preliminary human experimentation can be used to refine our understanding of the biological principles underpinning the heterogeneity of human disease and polymorphism(s). • http://www.translational-medicine.com/info/about

4. Current State • 17 year innovation adoption curve from Discovery  Practice • Even if a standard is accepted: • 50% chance of receiving inappropriate care • 5-10% chance of preventable, anticipatable adverse event pati • Healthcare inflation, increasing resistance for reimbursement of, new diagnostics and therapeutics

5. Current state: Knowledge Processing Requirements • Medical literature doubling every 19 years • Doubles every 22 months for AIDS care • 2 Million facts needed to practice • Genomics, Personalized Medicine will increase the problem exponentially • Typical drug order today with decision support accounts for, at best, Age, Weight, Height, Labs, Other Active Meds, Allergies, Diagnoses • Today, there are 3000+ molecular diagnostic tests on the market, typical HIT systems cannot support complex, multi-hierarchical chaining clinical decision support Covell DG, Uman GC, Manning PR. Ann Intern Med. 1985 Oct;103(4):596-9

6. Current State: Healthcare IT Vendors • Knowledge “hardwired” into applications • Little or no standardization on terminologies or information models • Knowledge engineering tools update content in transactional environments, no support for versioning, provenance, change propagation. • Clinical systems implementations have inadequate knowledge to meet current workflow and quality needs • Labor of converting knowledge into Clinical Decision Support is vastly underestimated

7. HealthCare and Life Sciences: A Knowledge Driven Endeavor E.g., Application of Clinical/Genomic Decision Support Rules Knowledge Application E.g., Analysis of clinical Care transactions for Rules, Patient Groups, Potential Biomarkers Knowledge Asset Management Knowledge Discovery E.g., Creation and Maintenance of Clinical Decision Support Rules

8. Use Case:Dr. Genomus Meets Basketball Player who fainted at Practice • Clinical exam reveals abnormal heart sounds • Family History: Father with sudden death at 40, • 2 younger brothers apparently normal • Ultrasound ordered based on clinical exam reveals cardiomyopathy Structured Physical Exam Structured Family History Structured Imaging Study Reports

9. Actionable Decision Support Echo triggers guidance to screen for possible mutations: - MYH7, MYBPC3, TNN2, TNNI3, TPM1, ACTC, MYL2, MYL3

10. statistics statistics application application server server population ownership registry manager database database encryption Knowledge-based Decision Support Connecting Dx, Rx, Outcomes and Prognosis Data to Genotypic Data for Cardiomyopathy Gene expression in HCM Test Results person raw value concept date Z5937X 3/4 Syncope Outcomes calculated every week microarray (encrypted) Myectomy ER visit Z5937X 3/4 Atrial Arrhythymi Palpitations Z5937X 3/4 ER visits Gene-Chips Z5937X 3/4 Clinic visits Ventricular Arrhy Echocardio Z5937X 4/6 ICD Gene-Chips Z5956X 5/2 Cong. Heart Failure microarray (encrypted) Cardiomyop Z5956X 5/2 Atrial Fib. Z5956X 5/2 Echocardio Z5956X 5/2 EKG Z5956X 3/9 Cardiac Arr Z5956X 3/9 ER Visit Z5956X 3/9 Thalamus Z5956X 3/9

11. Functional Requirements • Data Integration • Advantages of RDF • Incremental, cost effective approach • Clinical Decision Support • Classification v/s Actionable Decision Support • Advantages of ontology-based inferences • Knowledge Maintenance and Change Propagation • Advantages of using ontology-based inferences

12. A Strawman Ontology OWL ontologies that blend knowledge from the Clinical and Genomic Domains Clinical Knowledge Figure reprinted with permission from Cerebra, Inc. Genomic Knowledge

13. Domain Ontologies for Translational Medicine Research Instantiation Merged RDF Graph • Use of RDF graphs that instantiate • these ontologies: • - Rules/semantics-based integration • independent of location, method of access or underlying data structures! • Highly configurable, minimize • software coding RDF Graph 1 RDF Graph 2 RDF Wrapper RDF Wrapper GIGPAD Study RPDR Data Integration

14. Patient (id = URI1) “Mr. X” name has_structured_test_result related_to Patient (id = URI1) Person (id = URI2) MolecularDiagnosticTestResult (id = URI4) associated_relative has_family_history identifies_mutation indicates_disease problem MYH7 missense Ser532Pro (id = URI5) FamilyHistory (id = URI3) “Sudden Death” Dialated Cardiomyopathy (id = URI6) EMR Data LIMS Data evidence2 95% Bridging Clinical and Genomic Information “Paternal” 1 90% degree type evidence1 • Rule/Semantics-based Integration: • Match Nodes with same Ids • Create new links: IF a patient’s structured test result indicates a disease • THEN add a “suffers from link” to that disease

15. Bridging Clinical and Genomic Information 90%, 95% evidence Dialated Cardiomyopathy (id = URI6) “Paternal” suffers_from 1 “Mr. X” type degree name indicates_disease has_structured_test_result related_to Patient (id = URI1) Person (id = URI2) StructuredTestResult (id = URI4) identifies_mutation associated_relative has_family_history has_gene MYH7 missense Ser532Pro (id = URI5) problem FamilyHistory (id = URI3) “Sudden Death” RDF Graphs provide a semantics-rich substrate for decision support. Can be exploited by SWRL Rules

16. Advantages • RDF: Graph based data model • More expressive than the tree based XML Schema Model • RDF: Reification • Same piece of information can be given different values of belief by different clinical genomic researchers • Potential for “Schema-less” Data Integration • Hypothesis driven approach to defining mapping rules • Can define mapping rules on the fly • Incremental approach for Data Integration • Ability to introduce new data sources into the mix incrementally at low cost • Use of Ontology to disallow meaningless mapping rules? • For e.g., mapping a gene to a protein…

17. Clinical and Genomic Decision Support IF the patient’s LDL test result is greater than 120 AND the patient has a contraindication to Fibric Acid THEN Prescribe Zetia Lipid Management Protocol Contraindication to Fibric Acid: Clinical Definition (Old) The patient is contraindicated for Fibric Acid if he has an allergy to Fibric Acid or has elevated Liver Panel Contraindication to Fibric Acid: Clinical+Genomic Definition (New) The patient is contraindicated for Fibric Acid if he has an allergy to Fibric Acid or has elevated Liver Panel or has a genetic mutation Missense: XYZ3:Ser@$#Pro Please note: Hypothetical – assume a genetic variant is a biomarker for patients contraindicated to Fibric Acid.

18. Clinical Decision Support: A Rules-based Implementation Business Object Model Design Class Patient: Person method get_name(): string; method has_genetic_test_result(): StructuredTestResult; method has_liver_panel_result(): LiverPanelResult; method has_ldl_result(): real; method has_contraindication(): set of string; method has_mutation(): string; method has_therapy(): set of string; method set_therapy(string): void; method has_allergy(): set of string; Class StructuredTestResult method get_patient(): Patient; method indicates_disease(): Disease; method identifies_mutation(): set of string; method evidence_of_mutation(string): real; Class LiverPanelResult method get_patient(): Patient; method get_ALP(): real; method get_ALT(): real; method get_AST(): real; method get_Total_Bilirubin(): real; method get_Creatinine(): real;

19. Definition of “Fibric Acid Contraindication” Clinical Decision Support: A Rules-based Implementation Rule base Design IF the_patient.has_ldl_result() > 120 AND ((the_patient.has_liver_panel_result().get_ALP()  <NormalRange> AND the_patient.has_liver_panel_result().get_ALT()  <NormalRange> AND the_patient.has_liver_panel_result().get_AST()  <NormalRange> AND the_patient.has_liver_panel_result().get_Total_Bilirubin()  <NormalRange> AND the_patient.has_liver_panel_result().get_Creatinine()  <NormalRange>) OR “Fibric Acid Allergy”  the_patient.has_allergy() OR “Missense: XYZ3:Ser@$#Pro”  the_patient.has_mutation()) THEN the_patient.set_therapy(“Zetia Lipid Management Protocol”)

20. Clinical Decision Support:Definitions vs Decisions Commonly occurring design pattern: • The definition of a “Fibric Acid Contraindication” is represented using rules. • The decision related to therapeutic intervention is also represented using rules. Currently, both these inferences are performed by the rules engine.

21. Clinical Decision Support:Decoupling definitions vs decisions • Evaluation of classification based inferences (does patient have a fibric acid contraindication?) can be evaluated by an ontology engine. • Reduces overhead on Rule Engine • Opens up the possibility of plugging-in other specialized inference engines (e.g., spatio-temporal conditions) • Makes knowledge maintenance easier • Each definition may be referred to in 100s of rules..

22. Knowledge Provenance and Maintenance • There is a close interrelationship between knowledge change and provenance • What has changed? – Change • Why did it change? – Provenance • Did someone change it? – Provenance • Did its components change? – Change • Who changed it? – Provenance • Significance: • Rapid Knowledge Discovery and Evolution in Healthcare and Life Sciences

23. When Knowledge Changes… How quickly can you change the content of your rules, order sets, templates, and reports?

24. The “Genetic Revolution” Begins Leading the News: Roche Test Promises to Tailor Drugs to Patients --- Precise Genetic Approach Could Mean Major Changes In Development, Treatment June 25, 2003 Roche Holding AG is launching the first gene test able to predict how a person will react to a large range of commonly prescribed medicines, one of the biggest forays yet into tailoring drugs to a patient's genetic makeup. The test is part of an emerging approach to treatment that health experts expect could lead to big changes in the way drugs are developed, marketed and prescribed. For all of the advances in medicine, doctors today determine the best medicine and dose for an ailing patient largely by trial and error. The fast-growing field of "personalized" medicine hopes to remove such risks and alter the pharmaceutical industry's more one-size-fits-all approach in making and selling drugs.

25. Knowledge Changes • Introduction of a new molecular diagnostic test that identifies genetic variants that have been established as biomarkers for patients contraindicated for fibric acid. Clinical (Old) definition of Fibric Acid Contraindication  Clinical/Genomic (New) definition of Fibric Acid Contraindication • Change in definitions of clinical normality, e.g., change in the definition of normal value ranges of AST Results say from 0  AST  4020  AST  40

26. implemented in an ontology engine Knowledge Maintenance and Provenance IF the_patient.has_ldl_result() > 120 AND “Fibric Acid Contraindication”  the_patient.has_contraindication() THEN set the_patient.has_therapy(“Zetia Lipid Management Protocol”)

27. Patient_with_Biomarker has_mutation: “Missense: XYZ3:Ser@$#Pro” Domain Ontology

28. Bridge – Composition Ontology Rule base

29. Knowledge Change and Provenance • At each stage, Knowledge Engineer gets notified of: • What has changed? • The definition of Fibric Acid Contraindication • Why did it change? • Fibric Acid Contraindication  Patient with Abnormal Liver Panel  Abnormal Liver Panel  Abnormal AST  Change in AST Values • Fibric Acid Contraindication  Patient with Biomarker  Patient with a particular genetic variant • Who/What was responsible for the change? • Knowledge Engineer who entered the changed AST values? • Change in a Clinical Guideline?

30. Semantics of Knowledge Maintenance • Managing change and provenance is a very difficult problem • Semantics can play a crucial role in it: • A reasoner can navigate a semantic model of knowledge and propagate change • One can declaratively change the model at any time • The reasoner will compute the new changes! • Configuration v/s coding. Could read to a huge ROI!

31. Conclusions • Healthcare and Life Sciences is a knowledge intensive field. The ability to capture semantics of this knowledge is crucial for implementation. • Incremental and cost-effective approaches to support “as needed” data integration need to be supported. • Scalable and modular approaches for decision support need to be designed and implemented. • The rate of Knowledge Updates will change drastically as Genomic knowledge explodes. Automated Semantics-based Knowledge Update and Propagation will be key in keeping the knowledge updated and current • Personalized/Translational Medicine cannot be implemented in a scalable, efficient and extensible manner without Semantic Web technologies

32. Shameless Marketing Plug Tutorial Presentation at WWW 2006, Edinburgh, UK Semantics for the HealthCare and Life Sciences - Vipul Kashyap, Eric Neumann and Tonya Hongsermeier