Clinical Bioinformatics: New Challenges for Human-Centered Decision-Support Systems

Speakers: Riccardo Bellazzi, University of Pavia Patrice Degoulet, Hospital Georges Pompidou (Paris) Amnon Shabo, IBM Research (Haifa) Vimla Patel, New York Academy of Medicine Moderator: Edward H. Shortliffe, Arizona State University

www.mie2012.it! 1990s: Relationship of Medical Informatics and Bioinformatics! Biological and Clinical Applications of Interrelated Techniques and Medical Bioinformatics Solution:Methods ! Informatics “Biomedical” Informatics! Anticipation of their Future Clinical Interdependencies Biomedical Informatics in Perspective

BiomedicalClinical Informatics (BMI) Basic! Education and Research Research ! Bioinformatics!

Methods, Techniques, Theories!

Bioinformatics and Health Informatics: Structural Clinical Informatics Applied (Imaging) and Public Health Research Informatics Informatics and Practice! Informatics in Translational Science: Translational Bioinformatics (TBI) and Clinical Research Informatics (CRI)

Molecules, Cells, Tissues, Patients, Individuals, Populations , Organs! Societies! Four Presentations

• Clinical Bioinformatics – decision support through knowledge and data integration (Riccardo Bellazzi) • Bioclininical data warehouses and translational research (Patrice Degoulet) • The Data Representation Challenge: Encapsulate and Bubble-up (Amnon Shabo) • Cognitive Foundations for Decision Support in Clinical Bioinformatics (Vimla Patel)

Billedet kan ikke vises. Computeren har muligvis ikke hukommelse nok til at åbne billedet, eller billedet er muligvis blevet beskadiget. Genstart computeren, og åbn derefter ﬁlen igen. Hvis det røde x stadig vises, skal du muligvis slette billedet og indsætte det igen. www.mie2012.it! Discussion

Please use microphone and identify yourself and your afﬁliation

Bioclinical Data Warehouses & Translational Research : the HEGP Case

Patrice Degoulet & Paul Avillach

Faculté de Médecine René Descartes, Université Paris 5 Hôpital Européen Georges Pompidou (HEGP)

INSERM UMR_S 872 eq22 August 2012 Why should we develop Bioclinical Data Warehouses

1. EHR driven research – Patient selection for CR studies (e.g., EHR4CR) – Biomarker discovery (phenotype-genotype) – Personalized medicine

2. Phenotypic augmentation for clinical research studies – Additional source of bioclinical data for cohort studies – Reuse of EHR data to feed a CR study

• Prokosch HU, Ganslandt T. Perspectives for medical informatics. Reusing the electronic medical record for clinical research. Methods Inf Med. 2009; 48(1): 38-44. • Kohane IS. Using electronic health records to drive discovery in disease genomics. Nat Rev Genet. 2011; 12(6): 417-28. • Jensen PB, Jensen LJ, Brunak S. Mining electronic health records: towards better research applications and clinical care. Nat Rev Genet. 2012; 13(6): 395-405. HEGP Background

• HEGP is the most recent acute care hospital within the 37 AP-HP hospitals • HEGP meets the needs of the 600, 000 inhabitants of the Paris south-west HEGP background Shared Biobank (2008-) HEGP BDW EHR/BDW integration

Production environment Evaluation/Research environment

ETL suite (Talend Open Studio)

EHR : EHR : Biomedical Data External Operational Mirrored Warehouse Databases Database Database (BDW) (ODS)

R Real time requests i2b2/tranSMART tools Business Object Data Analysis IBM Ilog Rules Data Mining HEGP BDW I2b2 + tranSMART data sources

Health care Clinical Research Health Information System

Browser and Analysis DRG tools available for MD & Pharm EHR forms BDW EHR reports i2b2 ETL Structured data from research Biology CDW studies

Imaging ‘omics ETL once aweek ETL +’omics ETL Pathology ‘omics data

Rx BDW content i2b2 CDW content (July 2012)

Dimensions Categories Numbers Patient 432 033

Concept ICD10 classification 21 356 dimension Laboratory results classification 8 272 Drug classification (ATC) 33 612 EHR forms concepts 5 950

Observation facts ICD10 Diagnosis 2 537 633 Laboratory results 85 598 854 Drug prescriptions 2 474 985 EHR forms items 28 641 547 Text reports 902 747 BDW use i2b2 CDW queries (Jan. 2011-Aug. 2012)

• 158 MD + Pharm trained • 1386 requests - 15 cohorts created (i2b2 datamarts) CIS use In vivo evaluation of decision rules (EHR)

Rate of inappropriate prescriptions 6 alternating 2-month phases: control vs. intervention (Aug. 2006- Aug. 2007)

Physicians Alerting off Alerting on p Junior 21.5% 16.3% Senior 20.9% 29.3% p p=0.88 p=0.01

Total 21.3% 19.9% 0.63 (NS)

Sellier E et al. Effect of alerts for drug dosage adjustments in inpatients with renal insufficiency. JAMIA 2009; 16:203-10. BDW use In silico evaluation of decision rules (BDW)

Development and optimization of clinical decision rules for dosage adjustment of « renal function dependent » drugs

Boussadi A, Caruba T, Zapletal E, Sabatier B, Durieux P, Degoulet P. A clinical data warehouse– based process for refining medication orders alerts. J Am Med Inform Assoc 2012; 19(5): 782-5 BDW use In silico evaluation of decision rules (BDW)

Initial set Final set

(CPOE vn) Suppressed : 45 (16.1%) (CPOE vn+1)

280 rules 371 rules Modified : 105 (37.5%) (10 drugs) (10 drugs)

Added : 136 (48.2%)

http://www.transmartproject.org ! The Transmart integration process

http://www.transmartproject.org BDW content Pilot Study

• Laurent-Puig P, Cayre A, Manceau G, Buc E, Bachet J-B, Lecomte T, et al. Analysis of PTEN, BRAF, and EGFR status in determining benefit from cetuximab therapy in wild-type KRAS metastatic colon cancer. J. Clin. Oncol. 2009 Dec 10;27(35):5924–30. TranSMART-based Ontology definition Proof of concept

• R module in tranSMART

• Published figure in JCO The HEGP BDW Conclusion (1)

Achievements • A methodology to export concepts and data from an integrated CIS (2009-) • An operational CDW directly used by MD and Pharm (2011- ) • Installation & evaluation of a tranSMART platform to augment clinical data with omic information (2012-)

Benefits of the approach

• Availability of clinical data has generated a virtuous cycle at the HEGP end-user level (e.g., improved standardized questionnaires) • tranSMART is based on the i2b2 model for clinical data

The HEGP BDW Conclusion (2)

Limits of the approach • Lack of semantic integration tools to merge i2b2-tranSMART concepts • i2b2-tranSMART data model heterogeneity

Perspectives

• Semantic integration within the i2b2-tranSMART platform to enable: 1) meta-analysis from multiple cohorts 2) phenotypic augmentation in genomic driven research

Acknowledgments

www.i2b2.org Informatics & Public Health Dept. www.transmartproject.org Pierre Durieux, MD Eric Zapletal, PhD www.recomdata.com Kevin Zarca, MD Vincent Canuel, Resident

Contacts : [email protected] [email protected] MIE 2012 – Panel on Clinical Bioinformatics August 27, 2012 New challenges for human-centered decision- support systems

The Data Representation Challenge

Presented by Amnon Shabo (Shvo) • Chair, Medical Informatics Community of IBM Research • Head of IBM HCLS* Standards Program • Co-chair & facilitator of HL7 Clinical Genomics • Co-editor of CDA R2, CCD and Pedigree IBM Haifa Research Lab (HRL)

1. Flat representations are flat tires for CDS engines…!

Health data semantics and context cannot be faithfully represented using flat structures (e.g., a list of entries), rather it requires a compositional language that associates various data entries to a meaningful statement 2 Haifa Research Lab The New Generation of Standards is Object-oriented!

Observation Procedure Object Object Code Insert into basic Participant Medication health objects Code Object Grammar* Object Code

SNOMED, LOINC, ICD, etc.

Observation O1 (consisting of Observations O11 and O12 and related to Subject S1), is Clinical Statement the reason for Procedure P1 (performed by Clinician C1) which is the cause of Observation O2… Lab Pharma Docs Others

* The grammar should be part of a health-dedicated reference information model, e.g., the HL7 RIM or the openEHR RIM 3 Haifa Research Lab Clinical Omics Statement ! § An effort within the HL7 Clinical Genomics Work Group § Latest effort: CDA Implementation Guide for Genetic Testing Reports (GTR)

§ An abstract Clinical Omic Statement (COS) template that § Has at its core a genomic observation (e.g., a DNA sequence variation) § If it’s a reportable finding, then it should be associated with indications and interpretations, specimen and genomic source class § The major finding can be associated with associated observation (e.g., amino acid change) § Optionally, performers may be specified (overriding header performers)

§ The COS abstract template is instantiated by specialized COS’s, e.g., for genetic variations or cytogenetics

Associated Observations

Indications Omics Interpretations Performers Observation Clinical Genomic Statement Genomic Clinical Genomic Specimen Source

* Taken from the HL7 CDA Implementation Guide for Genetic Testing Reports. 4 Haifa Research Lab Main Concepts 2. ‘Encapsulate & Bubble-up’…

Personal biological key data should be encapsulated in its native format into clinical data structures, where 'bubbled-up' items could be associated with phenotypic data, as part of clinical data standards 5 Haifa Research Lab Clinical Genomics Data Standards: Encapsulate & Bubble-up

Genomic Data the challenge… Clinical Practices Sources

Knowledge (KBs, Ontologies, registries, reference DBs, Papers, etc.)

Encapsulation by predefined & EHR constrained Bubbling-up is bioinformatics System done continuously schemas by specialized CDS applications

Decision Support Applications Bubble up the most clinically-significant raw genomic data into specialized objects and link them with clinical data from the patient EHR

Continually reanalysis * Taken from the HL7 Clinical Genomics Specifications’ design principles. 6 Haifa Research Lab Main Concepts

3. Narrative  Structured ‘Reconciliation’

Health data standards need to accommodate unstructured data (e.g., clinician's narrative), while maintaining interlinks to structured data entries corresponding to contents that have been structured

7 Haifa Research Lab HL7/ISO CDA (Clinical Document Architecture)

³ EMR ³ Printed ³ Transcription ³ Bedside ³ … ³ …

Human-to-Human

³ Medical Records CDA ³ Transformation ³ … Machine-to-Machine

³ Clinical Decision Support ³ Patient held-records alerts ³ …

8 Haifa Research Lab CDA Overview

CDA Document § CDA – a generic specification Header § Could be used to represent various types of documents: Body § Consultation note § Visit / progress note Section § Referral letter Narrative § Discharge summary § Operative note BodyClinical Statement § … code codeCDA CDA Entrycode § A document type is also Entry called ‘template’ or CDA ‘implementation guide’ … Entry

9 Haifa Research Lab Main Concepts 4. Model-driven Constraining

The more expressive a standard is the less interoperate it is; addressing this tension is done by model-driven constraining of generic standards so that constrained specs remain semantically consistent 10

Haifa Research Lab Rich Expressiveness vs. Interoperability

§ Expressive structures lead to optionality § Possible solution: Constraining formalisms § Archetypes with EHR 13606 (European/ISO Standard) § HL7 Templates (no formalism has been agreed upon) § UML+OCL (Object Constraining Language) § GELLO (OCL-based) for clinical decision support § Public registries of archetypes / templates § Need dedicated IT to provide registry services

§ In research settings § Granularity, specificity and heterogeneity of data is higher § The same constraining technologies allows for capturing similarities while preserving disparities § Cohorts harmonization could be reassessed depending on analysis results, creating new computed fields and aggregates

11 Haifa Research Lab Section and Entry Level Templates – the CCD Example § CCD – a CDA Template for Continuity of Care Document

§ Bases on section templates: § Problems / Allergies § Medications § Immunizations § Vita Signs § Lab Results § Medical Procedures § Previous Encounters § Family and social history § Alerts and Plan of care Haifa Research Lab GTR Rendered – Summary Section

§ GTR – CDA constrained for Genetic Testing Reports

Draft that has not been clinically validated 13 Haifa Research Lab Standards-based Data Instance Generation

Data Source Template Model OWL Ontology Adapter

Instance CTS Generation Java API Engine

Mapping local Vocabularies

Representing constraints

Using MDHT with UML Standard-based +OCL to represent and Instances validate constraints (e.g., CDA) Conform to the Template Model 14 Haifa Research Lab Main Concepts

5. Family History as a Test Case

Family health history is an important test case of converging PHR, EHR and genetic data; it can make a difference as it is considered the most unused resource in clinical practice

15 Haifa Research Lab Family History: PHR-EHR-GEN Convergence

EHR PHR

Enable Decision Support e.g., risk analysis algorithms

Genomics Test Case of everything we

care about! 16 Haifa Research Lab The HL7 Pedigree Standard

§ HL7 v3 ANSI Normative (2007); under ISO ballot EHR

HL7 v3 § Selected by US HHS to Pedigree exchange FH between EHR and Clinical HL7 CCD Decision Support CDS (CDS) Applications HL7 v3 Pedigree § CCD provides the medical history of the PHR patient whose pedigree is exchanged

17 Haifa Research Lab Main Concepts

6. EHR – the optimal organizer!

Various healthcare and life sciences standards should be fused into a single & coherent information entity representing health information of an individual

18 Haifa Research Lab Extending Healthcare Federation Standards to serve EHR & Research

Topical data

Sensitivities | Diagnoses | Medications | etc. Non- redundant data Summative Info E H R Evidence

Temporal Data

Ongoing extraction Ongoing summarization and Medical records: charts, documents, lab results, imaging, etc. Haifa Research Lab The End

³ Thanks for your attention! Unlocking the Power ³ Questions? of Health Information. ³ Comments: [email protected]

20 Clinical Bioinformatics decision support through knowledge and data integration

Riccardo Bellazzi Laboratory for Biomedical Informatics Labs “Mario Stefanelli” Biomedical University of Pavia Informatics “Mario LISRC - IRCCS Fondazione S. Maugeri Stefanelli” Pavia Translational Bioinformatics

"It is the responsibility of those of us involved in today's biomedical research enterprise to translate the remarkable scientific innovations we are witnessing into health gains for the nation... At no other time has the need for a robust, bidirectional information flow between basic and translational scientists been so necessary."

--Dr. Elias Zerhouni, Director of the National Institutes of Health, 2005 Clinical Bioinformatics

Translational Cancer Research & Clinical Bioinformatics data

biological clinical patient tissue

proteomic transcriptomic histology images genomic demographic patient HETEROGENOUS DATA

DATAMINE biomarkers clinical trial

Paul Lewis, Cancer Informatics Group, Univ. Swansea Clinical Bioinformatics (i2b2)

Bedside to bench Knowledge discovery

Knowledge to practice Test new Knowledge

Over 80 drugs contains genec informaon on their drug labeling (warfarin)

Decision support -omics … l Disease subgroups l Mutations, gene expression markers l Slow translation of the discoveries into diagnostic guidelines l Personalized medicine l Targeted drugs / re-targeted drugs l Need changes in the paradigm of clinical trials l Availability of high throughput data at the point of care l Diagnostics in short time l Potential new mutations, uncertain evidence Knowledge and data integraon

GEO clinical data KEGG MeSH ontology

GO Annotaons

PubMed GO

organism P2P genome sequences taxonomy and annotaon

Mutaon databases -omics data (3rd generaon sequencing) Adapted from Zupan (2009)

Designing new clinical decision support systems

Annotation tools EHR

Web interface Text mining and literature search engines CBR Data plugin warehouse

Reasoning module I2b2/SMART environment

Semantic Wiki-based collaborative system Dilated cardiomiopathy

Dystroﬁnopathies Laminopathies Desminopathies Mitocondriopathies Epicardinopathies DCM Acnopathies Zaspopathies Desmosonopathies

Centre for Inherited Cardiovascular Diseases - IRCCS Policlinico San Matteo - Pavia Gene networks - cardiomyopathies Networks navigation

Interest propagation The Onco-i2b2 Project (D. Segagni, et al, BMC Bioinformacs, 2012)

HIS CRC Anonymized data Data

Match IDs i2b2 Researcher Paent

Samples Anonymized samples Laboratory Biobank Clinical paent Research management Kohane IS. Using electronic health records to drive discovery in disease genomics. Nat Rev Genet. 2011 Extending CBR

Case-based reasoning

Exploiting text mining, ontologies and terminologies - Retrieval from heterogeneous data bases - Retrieval of articles reporting similar patients Drugs repurposing The italian Amyloidosis network – community of practices The ST-model (Ramoni, Stefanelli et al)

Hypothese s

• The ST-Model: an epistemological model of scientific and medical reasoning • hypotheses selection/ generation phase: abstraction and abduction • hypotheses testing phase: ranking, deduction, eliminative induction

DATA Acknowledgements

Harvard Medical School

University of Ljubljana

IRCCS Fondazione C. Mondino

Bioinformatics and Data Mining group (http://bioinfo.unipv.it)

IRCCS Fondazione S. Maugeri Genec Individual Risk markers risk assessment

Life style

Genec markers Diagnosis Non- Diagnosis genec and molecular Drug therapy markers selecon planning and Paents therapy clinical planning ﬁndings Cognitive Foundations for Decision Support in Clinical Bioinformatics

Vimla L. Patel, PhD, DSc, FRSC Senior Research Scientist and Director Center for Cognitive Studies in Medicine and Public Health New York Academy of Medicine, New York City Professor, Biomedical Informatics Columbia University, New York City

Panel: Clinical Bioinformatics: New Challenges for Human- Centered Decision-Support Systems 2012 MIE, Pisa, Italy August 26-29, 2012 Knowledge Organization and Reasoning

• Structure and organization of available information drives the decision process • Paper-based charts and EHRs affect the way doctors reason with patient information • Nature of heuristics (strategies for making decisions) are different • Addition of genomic data will change information organization and thus will affect reasoning processes for patient- care decisions The EHR and Decision Making An electronic health record has a tightly structured format. Consequently when practitioners switch to the typical EHR, they no longer record the narrative thread of the patient history. Patel VL, Kushniruk AW, Yang S, Yale J. Impact of a computer-based patient record system on data collection, knowledge organization, and reasoning. J Am Med Inform Assoc. 2000; 7(6): 569-585. Challenge: • EHRs affect the way doctors practice • EHR changes with genomics data will alter practice • Opportunity to provide decision support at this level by developing representations that match mental models of decision makers Introductory History of a Patient’s Illness From a Representative Paper-Based Record (Before author’s use of an EMR)

This is a 74 year old woman, whose diagnosis of diabetes was made in February, as she had complained of polyuria/nocturia and fatigue for a few years. She was told her sugar was very high and she was sent to Dr. K., who started her on Diabeta 5 mg/d and sent her to Dr. S. in ophthalmology who reported normal retina. She lost weight, her polyuria improved, her bladder urgency got better, and her glucose values improved dramatically. She does no monitoring at home. She had to be hospitalized for an ankle fracture after falling on ice, for 3 months. At follow-up, Dr. K. seemed pleased with the results. Introductory History of a Similar Patient’s Illness: Representative Electronic Record

CHIEF COMPLAINT: Type II diabetes mellitus

PERSONAL HISTORY SURGICAL: cholecystectomy: Age 60 years old

MEDICAL: hypothyroidism: asymptomatic since 25 years

LIFE STYLE MEDICATION DIABETA (Tab 2.5 MG) Sig: 1 tab(s) Oral before breakfast SYNTHROID (Tab 0.125 MG) Sig: 1 tab(s) Oral before breakfast

HABITS: smoking: 0 alcohol: 0

Clinical Decision Making Process

Perception

Comprehension Cognitive Representation  Interpretation Support

Decision

Actions Genomic Information and Context of Interpretation • Only relevant genomic information is available to doctors; rest is inaccessible • Data loss: Interpretation of information in another context is not possible • Storage of data for future interpretation in another context will be valuable • Explicit retention of relational structures between laboratory data and associated genomic knowledge bases: as the knowledge gets updated, the nature of heuristics and practice will change Relationship Between Science and Practice • Practitioners do not use scientific information in routine daily practice • They only use heuristics learned through experience • They fall back on scientific explanations/reasoning under conditions of uncertainty or complexity • We accordingly need an understanding of the relationship between genomic science and how doctors use such information to make clinical decisions

Summary: Cognitive Challenges

• Increase in cognitive load: Load on working memory during information processing: Challenge to: • Provide decision support at the level of knowledge representation • Provide and retain relevant link to additional information to avoid information loss • Understand how doctors make patient-care decisions with new science and generated heuristics • Provide decision support at the point of care to support doctors’ clinical tasks Summary: Cognitive Challenges • Challenge: • To avoid extra formalization of rules to manage decisions, previously done informally • Leads to loss of both flexibility and resilience in practice • Human behavior is influenced by our thoughts and our social values • Difficult to change behavior • Science underlying clinical bioinformatics is constantly developing and being updated • Challenge to update decision heuristics to keep abreast of the evolving science Thank you

[email protected] http://ccsmph.nyam.org/