Charlie Mead MD, MSc Chief Technology Officer, National Cancer Institute (NCI)

1. Health Level 7 (HL7):A Brief Overview of a23-Year TrajectoryW3C Semantic WebHealth Care andLife Sciences SIG Charlie Mead MD, MSc Chief Technology Officer, National Cancer Institute (NCI) Center for Biomedical Informatics and Information Technology (CBIIT) Chair, HL7 Architecture Board (ArB)

2. Overview Time flies like an arrow. Fruit flies like a banana. • HL7 Origins: Mission and Version 1.0 • HL7 Adoption: Version 2.x • HL7 Maturation and Expansion: Version 3.0 • HL7 Evolution: Reorganization & Collaboration, Domain Analysis Models, Service-Awareness, and Enterprise Architecture • HL7 Interoperability Contexts: the Translational Continuum • Pharma’s “next” business model: intersection with healthcare • National Cancer Institute: from caBIG™ to BIG-Health™

3. HL7 Origins: Version 1.0

4. 1985-89: A (relatively) Simple Problem • Collection of seven enterprise-related hospitals with ~30 ADT systems needed to exchange data • RS-232-like byte-stream approach • Defined delimiters in message headers , fields/sub-fields, etc. • Syntactic interoperability with agreed-upon semantics in a restricted/closed domain-of-application • An engineering solution that worked! • Message paradigm (ACK/NACK) • “the only game in town” • Version 1.0 (< 10 messages) released circa 1987, 1.1 circa 1989 • Organization name chosen (Health Level 7) based on OSI stack • Solid adoption trajectory in US hospital community  interest in expansion of message repertoire beyond ADT

5. HL7 Adoption: Version 2.x

6. 1989-95: The Problem Gets Bigger • Increasing adoption drives interest in non-ADT domains • Financial/Patient Accounting • Orders management (e.g. labs, etc.) • (gradually) Clinical Care (e.g. order sets, care plans, etc.) • Organizational structure of HL7 established • Technical Committees/SIGs (domain-specific) • Bottom-up message development based on “business triggers driving information exchange” • Optionality allowed in all messages • Complex relationships were hand-tooled on a per-message basis • Z-segments (the equivalent of free text in an RDBMS) allowed • Cross-TC sharing of semantics based on “good citizenship/awareness” • Content expands • Version 2.0, 2.1, and 2.2 released between 1989-92 • Version 2.3, 2.4, and 2.5 relesed between 1992-95

7. 1989-95: The Problem Gets Bigger • Adoption increases • 75% (1992)/95% (1995) of US hospitals utilized HL7 2.x in two or more systems • Initial interest from non-US entities • Canada • Australia • Europe • UK NHS • HL7 becomes an ANSI SDO to facilitate interaction with ISO • Slowly but surely, HL7 was becoming a victim of its own success • “If you’ve seen one HL7 implementation, you’ve seen one HL7 implementation.” • “HL7 isn’t a standard, it’s a style guide.”

8. HL7 Maturation and Expansion: Version 3.0

9. 1995-2006: Success Drives New Approaches • HL7 BoD decided to embark on a new message-development strategy • Adoption of emerging UML as a standard modeling language • Adoption of a more formal abstract data type specification as underpinning for computable semantic interoperability • Decision to “stay out of the terminology business” and concentrate on the structures that bind terminologies, i.e…. • Development of a common Reference Information Model that would provide the “universe of semantics” for all HL7 domains-of-interest and could be used to develop all HL7 message structures • Emergence of a commitment to “more than just messaging in the HL7 2.x sense” • Computable semantic interoperability • Other related standards • Arden Syntax • CCOW • Clinical Document Architecture (CDA)

10. Health Level Seven (HL7) • “HL7 develops specifications that enable the semanticallyinteroperable exchange of healthcare data. ‘Data’ refers to any subject, patient, or population data required to facilitate the management or integration of any aspect healthcare including the management, delivery, evaluation of and reimbursement for healthcare services, as well as data necessary to conduct or support healthcare-related research. HL7 Specifications are created to enable the semanticallyinteroperable interchange of data between healthcare information systems across the entire healthcare continuum.” -- (Mead paraphrase of HL7 Mission Statement) • Conceptually congruent with W3C Semantic Web HCLS SIG Mission Statement

11. The Four Pillars ofComputable Semantic InteroperabilityNecessary but not Sufficient • #1 - Common model (or harmonized sibling models) across all domains-of-interest • Information model vs Data model • The semanticsof common structures – Domain Analysis Model • Discovered (in part) through analysis of business processes • #2- Model bound to robust data type specification • HL7 V3 Abstract Data Type Specification (R2) • ISO DT Specification

12. The Four Pillars ofComputable Semantic InteroperabilityNecessary but not Sufficient • #3 - Methodology for binding terms from concept-based terminologies • Domain-specific semantics • #4 - A formally defined process for defining specific structures to be exchanged between machines, i.e. a “data exchange standard” • Static structures (as defined via Pillars 1-3) bound to explicit data/information exchange constructs • As of Version 3.0, these constructs were still defined as “messages” in the traditional HL7 sense • Documents (content RIM-derived) could also be defined by HL7 TCs and be exchanged within or without accompanying message constructs A single CSI statement is made by binding common, cross-domain structures to domain-specific terminologies (semantics).

13. HL7 V3 Reference Information Model (RIM) “An instance of an Entity may play zero or more Roles. Each instance of a Role may, in turn, play zero or more instances of a Participation in the context of an instance of an Act. Each instance of a Participation participates in a one and only one Act for the ‘duration’ of that Act. Acts may be related to each other through instances of Act Relationship.” ActRelationship • Has component • Is supported by 0..* 0..* 0..* 1 0..* 1 1 Entity Role Participation Act 1 1 1..* • Organization • Place • Person • Living Subject • Material • Patient • Member • Healthcare facility • Practitioner • Practitioner assignment • Specimen • Location • Referral • Transportation • Supply • Procedure • Consent • Observation • Medication • Administrative act • Financial act • Author • Reviewer • Verifier • Subject • Target • Tracker

14. Collection, Context, and AttributionBuilding Complex RIM-based structures • A diagnosis of pneumonia (observation Act) related to three other observations Acts. Each Act is fully attributed with its own context of Entity-Role-Participation values. AR: “is supported by” OBS: Temp 101F Attribution is source for has target OBS: Dx Pneumonia AR: “is supported by” OBS: Abnormal CXR Attribution has target is source for AR: “is supported by” is source for Attribution OBS: Elevated WBC has target PARTICIPAT: Subject ENTITY: Person ROLE: Clinician Attribution PARTICIPAT: Author ROLE: Patient

15. Shakespeare in RIM-speak(courtesy of David Markwell) subject subject responsible party direct target All the world’s a stageAnd all men and women are merely playersOne man, in his time, has many parts,First the infant, Mewling and puking In the nurse’s arms.

16. Information vs Terminology Models:Intersecting and interleaving semantic structures Information Model Terminology Model Common StructuresforShared Semantics Domain-Specific Terms specifying Domain-Specific Semantics Binding/Interface Information Model Terminology Model Common Structures bound toDomain-Specific Structures specifyingDomain-Specific Semantics Domain-Specific Terms specifying Domain-Specific Semantics Example: Appropriately constructed semantic web structures should be able to distinguish between “Grade IV allergic rxn to Penicillin” represented in several ways using various combinations of RIM and SNOMED-CT codes.

17. HL7’s Clinical Document Architecture (CDA) • Emerged coincident with development of XML • Driven by “document-centricity” of much of healthcare practice • Emphasized the importance of transmitting both text and structured data • Identified “fundamental document characteristics” • Persistence • Authentication • Stewardship • Wholeness • Global/Local Context • Human Readability • Nicholas Negroponte’s “bits and atoms” paradigm is particularly relevent • Release 1 (circa 2001) was only partially RIM-derived • Rapid uptake/adoption in European community (less in US) • Release 2 (circa 2006) entirely RIM-derived • Adopted by HITSP (Coordination of Care Document (CCD)

18. Incremental ComputableSemantic Interoperability 1001 0100 0100 1001 0100 0100 1011 1110 0101 1011 1110 0101 Highly “Informational” Systems * 1001 0100 0100 1011 1110 0101 * Less “Informational” Systems 1001 0100 0100 1011 1110 0101 *HL7 Clinical Document Architecture: Single standard forcomputer processable and computer manageable data (Wes Rishel, Gartner Group)

19. HL7 TCs and the Life Sciences • Focus up until circa 1997 was “clinical care” • However, international interest in HL7 led to new domains-of-interest and new TCs • Regulated Clinical Research Information Management (RCRIM) • Pharma (CDISC), FDA • Clinical Genomics • Imaging • Medical Devices • Security • Etc. • Of particular interest to W3C SW HCLS SIG is the RCRIM TC

20. RCRIM TC • The RCRIM TC … defines standards to improve or enhance information management during research and regulatory evaluation of the safety and efficacy of therapeutic products or procedures worldwide. The committee defines messages, document structures, and terminologies to support the interoperability of systems and processes used in the collection, storage, distribution, integration and analysis of ‘clinical trial’ information. Specific areas of interest include: • Structured Protocols • Product Stability and Labeling • Clinical Trial Reporting • AEs (with CDC) • SDTM (with CDISC)

21. A caBIG™ Example(from Covitz et al, Bioinformatics, V19, N18, P2404) • Patient has headache, focal weakness, history of seizures • Workup reveals glioblastoma multiforma, subtype astrocytoma • Is this tumor histology associated with gene expression abnormalities? • Yes, in the p53 signaling pathway including BCL2, TIMP3, GADD45A, CCND1 • Is there documented evidence of aberrant expression of (e.g. CCND1)? • Yes, SAGE tags for cyclin D1 appear with 3x greater frequency in cancerous vs normal brain tissue • Are any gene products of the p53 signaling pathway known targets for therapeutic agents? • Yes, TP53, RB1, BCL2, CDK4, MDM2, CCNE1 • Are any of the agents known to target these genes being specifically tested in glioblastoma patients? • Yes, trials xxx and yyy are currently underway

22. HL7 Evolution: Reorganization & Collaboration, Domain Analysis Models, Service-Awareness, and Enterprise Architecture

23. Reorganization • As HL7 has grew to have ~3000 members, 40+ TCs and SIGs, ~30 International Affiliates, and as several countries implementing (or attempting to implement) its standards, it became increasingly obvious that -- like a growing company -- it needed to reassess its organizational structure and processes. • Multi-dimensional effort began in 2005 and continuuing to present • Reconstituted BoD • CEO • CTO • Technical Steering Committee • Architecture Board • Decreased number of TCs/SIGs • Emphasis on project management and common development methodology • Use of ANSI DSTU to encourage testing before final ballot

24. Collaboration • HL7 actively seeks collaboration with other organizations developing standards for healthcare, life sciences, clinical research internationally • Goal is to avoid redundant efforts • Examples include • ISO • CEN • OMG • CDISC • CDISC (Clinical Data Interchange Standards Consortium) • Established circa 2002 • Virtually entire pharma industry is represented • Prototype collaboration relationship for HL7 via RCRIM TC • Leading developer of a DAM for domain of “Protocol-driven research and associated regulatory artifacts”  the BRIDG Model

25. Domain Analysis Models:The Communication Pyramid Standardized Models (UML) -- DAM ` Non-standard Graphics ad hoc Drawings Problem Space Solution Space Implementation-Independent Abstraction Implementation-Specific Structured Documents Free-text Documents Discussions Communication

26. BRIDG circa 2004:Separating Analysis from Design/Implementation Problem-Space Model (a la HL7 Development Framework) Level of Abstraction RIM / DMIM RMIM / HMD / XSD ODM

27. Lessons Learned:Using a DAM • DAMs need to be applied in the context of a larger development (message, service, application) management process • DAMs should be both domain-friendly and semantically robust (technology useful) • In order to be truly effective, standards development needs to become less like the Waterfall and more ‘Agile,’ i.e. embedded in an interactive, iterative, incremental process. • Exemplar process has been successfully piloted at NCI and is now ready for application to all projects • A DAM that is ultimately used in message, application, or service development needs to address • Data Type bindings • Terminology bindings for coded data types

28. Lessons Learned:Working with BRIDG (1) • BRIDG only makes sense ‘in context’ • e.g. message development, application development, service specification, etc. • Analysis Paralysis occurs otherwise • Most effective use is in the context of an iterative/incremental development process (e.g. RUP, SCRUM, Agile, ect.) • NCI has integrated use of the BRIDG Model (and the use of analysis models in general) into its development practices • HL7 RCRIM appears to be ready to do the same • The BRIDG domain-of-interest is stable after 4+ years of use • Protocol-driven research involving human, animal, or device subjects and associated regulatory artifacts • Recently, questions have been raised as to whether the BRIDG domain-of-interest should include post-marketing safety/adverse events • Initial indications are that the answer is ‘Yes’ and that the effect on the model’s structure will be minimal

29. Lessons Learned:Working with BRIDG (2) • Teams need to start with the existing BRIDG Model • Subset as needed based on project focus • Add new semantics (e.g. classes, attributes, relationships, business rules, etc.) as needed • All new editions must be rigorously defined • Identify existing elements in the BRIDG model which are incorrect, unclear, too restrictive, etc.

30. BRIDG circa 2008:Separating Analysis from Design/Implementation Requirements Analysis Messages Messages Messages Services Services Services Applications Applications Applications Implementation-dependencies Design Technology/platform bindings Implementation From Wish to Reality

31. The Current BRIDG Model Understandable to Domain Experts Unambiguously mappable to HL7 RIM Varying levels of abstraction, explicitness, and ‘RIM-compliance’

32. The Revised, 2-layered (2-views) BRIDG Model Consistent levels of abstraction and explicitness in multiple sub-domain ‘Requirements Models’ Understandable to Domain Experts (DaM) Sub-Domain 1 Sub-Domain 2 Sub-Domain 3 Sub-Domain 4 Sub-Domain 5 Consistent levels of RIM-compliance and explicitness in a single ‘Analysis Model’ Unambiguously mappable to HL7 RIM (DAM) NOTE: Sub-domains may or may not intersect semantically

33. DAMs and Ontologies (1) An OWL-DL definition is worth at least several UML classes Domain-of-Interest described by Ontologic Representation (OWL-DL) visualized by Visual Conceptualization (UML DAM) A UML picture is wortha thousand Requirements Documents words

34. DAMs and Ontologies (2) Is described by /facilitates computational in Domain-of-Interest Ontologic Representation (OWL-DL) An OWL-DL definition is worth at least several UML classes A UML picture is wortha thousand Requirements Documents words provides input for /is vetted by ODM formally specifies / informsthe representation of Visual Conceptualization (UML DAM)

35. Service Awareness within HL7 • Initial work began in 2006 with the development of the Health Services Specification Project (HSSP), a joint effort with OMG • By CTO directive, has evolved to a directive to the newly-established ArB to develop a “Services-Aware Enterprise Architecture Framework” (SAEAF) for HL7 • Requirements include • Maximum utilization of existing static artifacts • Development of computationally robust behavioral/interaction model • Development of a scalable Conformance/Compliance framework

36. Enterprise Integration Strategies:Objects vs Messages vs Services • Objects • Finely-granulated • Difficult to trace to business functionality • Encapsulation not a positive when crossing enterprise boundaries • Messages • Payloads based on standards support semantic interoperability • Embedding dynamic/behavioral semantics within message causes run-time context ambiguity or non-enforceable contract semantics • Application Roles • Receiver Responsibilities • Services • Traceable to business-level requirements • Separation of static semantics (message payload) from dynamic semantics (“integration points,” contracts)

37. Service Awareness within HL7 • Historically, HL7 as conceptualized the world as “communicating clouds” but has not formally specified the semantics of the interactions that occur • HSSP began the specification process with its Service Functional Model (basically a services “requirements document”) • SAEAF extends the definitional space

38. HL7, MDA, CSI, SOA, and Distributed Systems Architecture • The intersection of HL7, MDA, Distributed Systems Architecture, SOA, and CSI provide a goal, the artifacts, portions of a methodology, and the framework for defining robust, durable business-oriented constructs that provide extensibility, reuse, and governance. Health Level 7 Service Oriented Architecture Computable Semantic Interoperability Reference Model For Open Distributed Processing Model Driven Architecture You are here (Vousêtesici)

39. Choreography: an Analysis Perspective NCI is using CDL as an analysis tool (via pi4soa open-source tool)

40. SAEAF Behavioral Framework The HL7 Specification Stack - Overview

41. SAEAF Specification Pattern

42. An Exemplar Service:Clinical Research Filtered Query (CRFQ) List Qualified Protocol Interface I/E criteria I/E criteria Qualified protocols CRFQ client (clinician, caregiver, patient C R F Q P2 P1 P4 P4 P2 P1 P3 I/E criteria Clinical data set I/E criteria Count Qualified Patients Interface Pt data Pt data C R F Q P3 Qualified patients CRFQ client (trial sponsor, CRO, Pharma) Pt data Protocol I/E criteria/ Safety criteria Pt data

43. HL7 Interoperability Contexts: • The Translational Continuum • -- Pharma’s “next” business model: intersection with healthcare • -- National Cancer Institute: from caBIG™ to BIG-Health™

44. Pharma’s essential challenge:Increased R&D expenditures, decreased NCEs to market Genomics Proteomics Combichem UHTS Phase I rejection Phase II rejection Phase III rejection Approved NCE Today’s R&D Infrastructure Source: PhRMA Ian Ferrier, PhD

45. The Transformation:Better early decisions, fewer late stage failures, decreased time-to-market Increased Early Drug Development Capabilities Genomics Proteomics Combichem UHTS Decreased Time in Pipeline Decreased Cost Phase I rejection Phase II rejection Phase III rejection Fewer late stage failures Ian Ferrier, PhD Increased Approved NCEs

46. The Vision is simple, and well understood …it is based on individualized data…and appropriate tools… Clinical data Collection Pharma-supplied Queries Sophisticated Knowledge Creation Tools Genomics, Proteomics, Chemistry, etc. Clinical Data Repository Ian Ferrier, PhD … ‘Knowledge-creation’  CSI platform

47. caBIG™ and BIG-Health™:Addressing the Infrastructure of the Current World of Biomedicine • Isolated information “islands” • Information dissemination uses models recognizable to Gutenberg • Pioneered by British Royal Academy of Science in the 17th century • Write manuscripts • “Publish” • Exchange information at meetings Need to convert islands into an integrated system

48. The caBIG™ Initiative caBIG™ GoalA virtual network of interconnected data, individuals, and organizations that whose goal is to redefine how research is conducted, care is provided, and patients/participants interact with the biomedical research enterprise. caBIG™ Vision • Connect the cancer research community through a shareable, interoperable electronic infrastructure • Deploy and extend standard rules and a common language to more easily share information • Build or adapt tools for collecting, analyzing, integrating and disseminating information associated with cancer research and care

49. caBIG™ Strategy • Connect the cancer research community through a shareable, interoperable infrastructure • Deploy and extend standard rules and a common language to more easily share information • Build or adapt tools for collecting, analyzing, integrating and disseminating information associated with cancer research and care

50. caBIG™ is utilizing information technology to join islands into a community Alabama Birmingham: UAB Comprehensive Cancer Center Arizona Phoenix: Translational Genomics Research Institute Tucson: University of Arizona California Berkeley: University of California Lawrence Berkeley National Laboratory University of California at Berkeley Los Angeles: AECOM California Institute of Technology University of Southern California Information Sciences Institute University of California at Irvine The Chao Family Comprehensive Cancer Center La Jolla: The Burnham Institute Sacramento: University of California Davis Cancer Center San Diego: SAIC San Francisco: University of California San Francisco Comprehensive Cancer Center Colorado Aurora: University of Colorado Cancer Center District of Columbia Department of Veterans Affairs Lombardi Cancer Research Center - Georgetown University Medical Center Florida Tampa: H. Lee Moffitt Cancer Center at the University of South Florida Hawaii Manoa: Cancer Research Center of Hawaii Illinois Argonne: Argonne National Laboratory Chicago: Robert H. Lurie Comprehensive Cancer Center of Northwestern University University of Chicago Cancer Research Center Urbana-Champaign: University of Illinois at Urbana-Champaign Indiana Indianapolis: Indiana University Cancer Center Regenstrief Institute, Inc. IowaIowa City: Holden Comprehensive Canter Center at the University of Iowa Louisiana New Orleans: Tulane University School of Medicine Maine Bar Harbor: The Jackson Laboratory Maryland Baltimore: The Sidney Kimmel Comprehensive Cancer Center at Johns Hopkins University Bethesda: Consumer Advocates in Research and Related Activities (CARRA) NCI Cancer Therapy Evaluation Program NCI Center for Bioinformatics NCI Center for Cancer Research NCI Center for Strategic Dissemination NCI Division of Cancer Control and Population Sciences NCI Division of Cancer Epidemiology and Genetics NCI Division of Cancer Prevention NCI Division of Cancer Treatment and Diagnosis Terrapin Systems Rockville: Capital Technology Information Services Emmes Corporation Information Management Services, Inc. Massachusetts Cambridge: Akaza Research Massachusetts Institute of Technology Somerville: Panther Informatics Michigan Ann Arbor: Internet2 University of Michigan Comprehensive Cancer Center Detroit: Meyer L. Prentis/Karmanos Comprehensive Cancer Center Minnesota Minneapolis: University of Minnesota Cancer Center Rochester: Mayo Clinic Cancer Center Nebraska Omaha: University of Nebraska Medical Center/Eppley Cancer Center New Hampshire Lebanon: Dartmouth College Dartmouth-Hitchcock Medical Center New York Buffalo: Roswell Park Cancer Institute Bronx: Albert Einstein Cancer Center Cold Spring Harbor: Cold Spring Harbor Laboratory New York: Herbert Irving Comprehensive Cancer Center Columbia University Memorial Sloan-Kettering Cancer Center New York University Medical Center White Plains: IBM North Carolina Chapel Hill: University of North Carolina Lineberger Comprehensive Cancer Center Raleigh-Durham: Alpha-Gamma Technologies, Inc. Constella Health Sciences Duke Comprehensive Cancer Center Ohio Cleveland: Case Comprehensive Cancer Center Columbus: Ohio State University Comprehensive Cancer Center Oregon Portland: Oregon Health & Science University Pennsylvania Philadelphia: Drexel University Fox Chase Cancer Center Kimmel Cancer Center at Thomas Jefferson University Abramson Cancer Center of the University of Pennsylvania Pittsburgh: University of Pittsburgh Cancer Institute Tennessee Memphis: St. Jude’s Children’s Research Hospital Texas Austin: 9 Star Research Houston: M.D. Anderson Cancer Center Virginia Fairfax: SRA International Reston: Scenpro Washington Seattle: DataWorks Development, Inc. Fred Hutchinson Cancer Research Center International Paris, France: Sanofi Aventis