Wh

at’s the

role

of h

ealth

system

s?

Lucila Ohno-Machado, MD, PhDUniversity of California San Diego

12/7/16 HIMMS Chile

Disclosures

• No relevant financial relationships with commercial interests

• Grant funding from NIH, PCORI, VHA and the University of California

• Editor-in-Chief of the J Amer Med Inform Assoc

• On staff at the Veteran’s Administration

• Non-paid advisor for not-for-profit institutions

• Not speaking on behalf of any of these entities

Electro

nic H

ealth

Re

cord

s

Big Data

Electronic Health Records (EHRs) in most health systems can be used for research• Predictive analytics on big data help clinicians and

administrators, but

• Re-identification success from ‘de-identified’ data is

possible, and depends on many factors:

• What is disclosed, to whom

• Who an attacker is interested in identifying and how much

she is willing to spend

Patient Interaction

AnalyticsDistributed Analysis Machine Learning

Data StructuringData Modeling Natural Language Processing

Predictive ModelingEvaluation Methods

Decision Support ToolsGuidelines, Alert & Reminders

Data Collection ToolsClinical Data Warehouse

mHealthSensorsEnvironment

Data De-IdentificationPrivacy Technology

Communication StrategiesInformed ConsentConsumer Health Informatics

BioinformaticsGenomicsProteomics

Who owns the data?

Who can use the data?

• Patient

• Researcher

• Institution

The Evolution of Healthcare IT Systems (M Hogarth, UC Davis)

LaboratoryInformation

System

“Electronic Health

Record” System

RadiologyInformation

System

PharmacyInformation

System

Data Access/Exchange v1.0- Within institutional systems

Data Access/Exchange v2.0- EHR to EHR

Data Access/Exchange v3.0- Clinical Data Networks- Federated querying

LIS

EHR

LIS RX LIS

EHR

LIS RX

EHR

EHREHR

EHR

EHR

EHR

Example of Consent for Care

[institution]

[institution]

[institution]

[state]

[institution]

Knowledge & Tools

Privacy

Consent

Data

• Share access to data and computation

• Train the new generation of data scientists

• Provide innovative software, platform, and infrastructure

• Protect privacy

• Algorithms

• Tools

• Infrastructure

• Policies

iDASH

Knowledge

& Tools

ServicesPlatform

Data

Sensors

Genomic

Clinical

ServiceWWW

Apps

Exec.

Aggreg.Hosting

Sharing

Policies

Platform

Research

Develop.

Federation

Biomedical Data Science

Data Discovery Index Consortium

NIH BD2K Initiative

Do for data what PubMed did for the literature

The consortium has funded several institutions around the world

U24AI117966

DEMOTuesday 11:15amSalon A-3 Lower Level

MetadataIngestion

Terminology server• Query expansion• Result ranking

DataMed User InterfaceSearch Engine

Metadata Management• Mapping• Indexing

Repositories

Data Sets

Funding Agencies

Data Producers

Publishers

Dat

a so

urc

es

U24AI117966

DEMOTuesday 11:15amSalon A-3 Lower Level

DigestingIndexingFindingFiguring outInteroperating withCountingUsingLeveragingTesting

data

Metadata elements identified by combining the two complementary approaches

top-down approach bottom-up approach

The development process in a nutshell

Model serialized as JSON schemas and mapping to schema.org

(v1.0, v1.1, v2.0, v2.1)

S Asunta-Sansone 2016

Metad

ata

Data and Precision Medicine Components

Site 1

OMOP CDMData

Claims Data

SITE 2...N

…

Data transformation into PM CDM

EHR System

PatientReported

OMOP CDMData

EHR System

Patient reported data

Wearables

PMCDM Data

PM CDM Data

Data & Research Support Center

Data transfer > quarterly

Claims Data

M Hogarth 2016

Do

ing th

e R

ight Th

ing

Is it ethical to share?People have not been explicitly asked and don’t know who is sharing what

Could people choose?Is it practical?What if massive number of people withdraw their data?

Is it ethical not to share? New discoveries and acceleration of science depend on sharing

patient-centered SCAlableNational Network for Effectiveness Research

)

Supported by the Patient-Centered Outcomes Research Institute Contract CDRN-1306-04819

pSCANNER is a stakeholder-governed,

distributed clinical data network that aims to make health

data more accessible and usable for the generation of

scientific evidence that patients, clinicians, and other

stakeholders together use to make more informed health

decisions.

Set research

priorities and

design research

Stakeholder-governedInvolve stakeholders—patients, patient advocates, clinicians, and

researchers—as advisors in the pSCANNER work.

Develop

education and

communication

materials

Weight management/obesity

Congestive heart failure

Kawasaki disease

Governance

Patient-Centered Outcomes Research (PCOR)

Focuses on patients’ needs and preferences

and on outcomes most important to

them.

Helps patients and other healthcare

stakeholders, such as caregivers,

clinicians, insurers, policymakers and

others, make better-informed decisions

about health and healthcare options.*

*The word options implies the comparison of different types of treatments, medications, or healthcare practices.

Comparative Effectiveness Research (CER) is one type of PCOR

“Preparing for the Community Forum: Thinking about quality health care” AHRQ Community Forum pg. 8

Comparative Effectiveness Research (CER)

Medicine A Medicine Bvs.

Which of the two medications?

Comparative Effectiveness Research (CER)

Three common types of PCOR and CER research design

Randomized

Control Trials

Pragmatic

Trials

Observational

Studies

pSCANNER is part of PCORnet

PCORnet seeks to improve the nation’s capacity to conduct

clinical research by creating a large, highly representative,

national patient-centered network that supports more

efficient clinical trials and observational studies.

PCORnet embodies a “community of research” by uniting systems, patients & clinicians

11 Clinical Data Research

Networks

(CDRNs)

18 Patient-Powered Research Networks

(PPRNs)

PCORnet:

A national infrastructure for patient-centered clinical research

13 CDRNs and 20 PPRNs Funded

This map depicts

the number of

PCORI-funded

Patient-Powered or

Clinical Data

Research Networks

that have coverage

in each state.

24 Million Patients

9 health systems

Phase 1

UCD 2.3M

UCSF 3.2M

UCLA 4.3M

UCI 300k

UC ReX

SCANNER

USC LA

VA VINCI 11M

UCSD 2.3M

CTSA hubNetwork

AltaMed 607kTCC 240k QueensCare 19k

Phase 2

UC Davis 2.3M

UCSF 3.2M

UCLA 4.3M

UC Irvine 300k

USC Keck 2M

LA Children’s 200k

LA DHS 600K

UC ReX

SCANNER

LA

VA VINCI 11M

UCSD 2.3M

Cedars-Sinai 2M

University of Washington CTSA

CTSA hubNetwork

San Mateo Medical Center 77k

University of Colorado

Altamed 200kChidlren’s Clinic 24k Queenscare 19k

30 Million people

14 health systems

U Texas Houston

Rutgers

Emory U

Collaborating Patient-Powered Research Networks

PPRN Targeted Condition (IRB approved)

Health eHeart Heart Disease

AR-PoWER Inflammatory Arthritis

CENA Alström Syndrome, Dyskeratosis Congenita, Gaucher Disease, Hepatitis, Inflammatory Breast Cancer, Joubert Syndrome, Klinefelter Syndrome and Associated Chromosomal Anomalies, Metachromatic Leukodystrophy, Pseudoxanthoma Elasticum

DuchenneConnect Duchenne and Becker Muscular Dystrophy

iConquerMS Multiple Sclerosis

PPRN (in process) Targeted Condition

ImproveCareNow Inflammatory Bowel Disease

PRIDEnet Sexual and Gender Minorities

Crohn's & Colitis Foundation of America (CCFA)

Crohn's Disease and Ulcerative Colitis

MoodNetwork Mood Disorders

Slide from Dr. Maehara

International collaborations are welcome and needed• Federated data for distributed analytics

• Common data model

• Minimal computational infrastructure

• Compatible institutional policies

• Agreement on rules of engagement

• Shared ethics principles

How pSCANNER improves health researchResearch has focused on the priorities of scientists and

clinicians, which may be different from those of

patients, family members, and caregivers.

Health research often requires large numbers of study

participants but many individual healthcare

organizations do not have enough patients with a given

condition.

Evidence from health research takes years to make it

into practice because of the challenges in adapting and

communicating research findings.

1

2

3

http://urp.ucsd.edu/for-students/what-is-research.html (Accessed 12/5/2014)

Local data in EHRs, clinical,

administrative systems

Standardized data from public health and other sources

Data

21

pSCANNERstandardizes data

into Common Data Model for PCORnet

3

• Local Data are Cleaned and Harmonized for pSCANNER

pSCANNER supports big sciencepSCANNER will enable researchers to obtain data from distributed sites covering over

32 million patients in a privacy-preserving environment.

UCLA

4.1M

2.2M

UCDUCSF

3M2M

Cedar SinaiVA

9.1M

0.3M

ALTAMED

24K

TCC QUEENSCARE

19K 1.4M

UCI UCSD

2.1M

pSCANNER Hub

Population & Outcomes Characterization

What are the environmental and genetic determinants of Kawasaki Disease?

Do antibiotics contribute to

childhood obesity?

Which aspirin dose is better in coronary artery disease?

Can we use EHR data for research?

Preserving Privacy

Privacy Preserving Analytics for KD in African-Americans

Consent for Data and Biosample

Sharing in Underserved Populations

Partnership for Epidemiological

Research

Study on Latinos

Which DNA variants are implicated in KD susceptibility in this population?Emory, Genome Institute of Singapore, Imperial College

Does consent rate depend on who is obtaining the consent?Maricopa Health System, FQHS in Arizona

Do patients understand what they consented for?San Diego State University

What type of ‘sharing’ is acceptable?University of Oklahoma

StrongHeartStudy on American Indian Populations

Data and Biospecimen SharingPrivacy Preserving Computation

Personal monitorsno regulation of apps

My neighbor’s dataDoes she have Disease X?

Other databaseslinking data for re-identification

Public ‘de-identified’ database of Condition XResearch database with “de-identified” EHRs, genomes

EHRsDisease, Family History, etc

Health Insurance Portability & Accountability Act

34

HIPAA ‘De-identified’ data• removal of 18 identifiers,

such as dates, biometrics, names, etc.

• expert certification of low risk of re-identification

• ‘Limited’ data sets have ‘de-identified data’ plus dates

Biometrics and Protected Health Information

PHI requires HIPAA compliance

• Biometrics require HIPAA compliance

Biometrics are

Protected Health Information (PHI)

Genomes are Biometrics

PHI requires HIPAA compliance

• Genomes should be treated as HIPAA identifiers

Biometrics are

Protected Health Information (PHI)

New DNA tests on a secret sample collected from a relative of suspect Albert deSalvotriggered the exhumation after authorities said there was a “familial match” with genetic material preserved in the killing of Sullivan…

Authorities made the match through DNA taken from a water bottle thrown away by DeSalvo’s nephew…

But a lawyer for the DeSalvos told CNN the family was “outraged, disgusted and offended” by the decision to secretly take a DNA sample of one of its members…

Technology “solutions” (mitigation strategies)

Data-centric

• Add noise to data

(e.g., differential privacy)

• Operate on encrypted data

(e.g., homomorphic encryption)

• Multiparty computation

(e.g., distributed analytics)

Institution/People centric

• Data broker(e.g., clinical data research networks)

• Patient-defined data sharing permissions

(e.g., consent management)

Policies

Ohno-Machado L. To Share or Not To Share: That Is Not the Question. Science Translational Medicine, 2012 4(165)

homomorphic encryption

secure multiparty computation

Institutional and Data-Centric Strategiesdifferential privacy

NIH U54HL1084600

Statistical Data Release (macrodata release)

Mohammed N. Privacy Preserving Heterogeneous Health Data Sharing. J Am Med Inform Assoc 2013

Jiang XL. Differential-Private Data Publishing Through Component Analysis. Transactions on Data Privacy 2013

Courtesy of Li XiongME-1310-07058 (Xiong)

Original records Original histogram

Statistical Data Release: Disclosure Risk

Courtesy of Li Xiong

Original records Original histogramPerturbed histogram with differential privacy

Statistical Data Release: Differential Privacy

Courtesy of Li Xiong

Differential Privacy (Dwork et al)

A privacy mechanism A gives ε-differential privacy if for all neighbouring databases D, D’, and for any possible output S ∈ Range(A),

Pr[A(D) = S] ≤ exp(ε) × Pr[A(D’) = S]

D D’

• D and D’ are neighboring databases if they differ on at most one record

Courtesy of Li Xiong

Clinical Data Research Networks

30 Million people

14 health systems

UC Davis 2.3M

UCSF 3.2M

UCLA 4.3M

UC Irvine 300k

USC Keck 2M

LA Children’s 200k

LA DHS 600K

UC ReX

SCANNER

LA

VA VINCI 11M

UCSD 2.3M

Cedars-Sinai 2M

University of Washington CTSA

CTSA hubNetwork

San Mateo Medical Center 77k

University of Colorado

Altamed 200kChidlren’s Clinic 24k Queenscare 19k

U Texas Houston

Rutgers

Emory U

Patient-Centered Outcomes Research InstituteCDRN-1306-04819

Ohno-Machado L. To Share or Not To Share: That Is Not the Question. Science Translational Medicine, 2012 4(165)

homomorphic encryption

secure multiparty computation

Data sharing ecosystem

Sharing Data, Tools, Systemsdifferential privacy

indexing

User requests data

for Quality

Improvement or

Research

•Identity & Trust

Management

•Policy

enforcement

Trusted

Broker(s)

Diverse Healthcare Entities

in 3 different states (federal, state,

private)

Distributed Computing

Wu Y et al. Grid Binary LOgistic REgression (GLORE): Building Shared Models Without Sharing Data. JAMIA, 2012 Wu Y et al. Grid Multi-Category Response Logistic Models. BMC Med Inform Dec Making 2015

NIH U54HL1084600

Distributed Regression Model

Conclusion: no patient data needs to be sent from the sites, only aggregates

Distributed Analytics across Horizontal and Vertical Partitions

Patient Age Genome data

A1 45 ACTGACT

A2 32 ACTTAGT

Patient Age Genome data

B1 48 CCTGACT

B2 72 CCTTAGT

Patient Age Genome data

A1 45 ACTGACT

A2 32 ACTTAGT

Li Y, et al. VERTIcal Grid lOgistic regression (VERTIGO) J Am Med Inf Assoc. 2015

International Collaboration

Slide from Dr. Shuang Wang

2nd Genome Privacy Protection Challenge

• Task 1: Homomorphicencryption (HME) based secure genomic data analysis

• Task 2: Secure comparison among genomic data in a distributed setting

Focus on secure outsourcing and secure data analysis in a distributed setting

Technology “solutions” (mitigation strategies)

Data-centric

• Add noise to data

(e.g., differential privacy)

• Operate on encrypted data

(e.g., homomorphic encryption)

• Multiparty computation

(e.g., distributed analytics)

Institution/People centric

• Data broker(e.g., clinical data research networks)

• Patient-defined data sharing permissions

(e.g., consent management)

Policies

Consent

Manageme

nt System

Do I wish to

disclose

data D to

U?

Sharing

Look-up

Yes

Patient I

Patient Interface

I can check

that U

looked at

my data D

• Data use

agreements

• Study registry

Trusted

broker

Healthcare

Institutions

User U

requests

Data D on

individual I

People-Centered Strategies

NIH R01HG008802

Informed CONsent for Clinical data Use in Research

iCONCUR

Courtesy of H Kim

NIH R01HG008802

My Sharing Choices • 14 Data Classes

+ Sharing for stem cell research

Courtesy of H Kim

Sharing Preference Distribution

Supported by the NIH Grant U54 HL108460 to the University of California, San Diego

Courtesy of H Kim

Preliminary dataon >1k patients

NIH R01HG008802

Big D

ata are a B

ig De

al

App datano regulation

Social network can be used

Databasescan be purchased

Asking peoplemay help

EHRscan be linked

Informatics Training for Global Health

NIH D43TW007015

Thank you

Acknowledgements to a large team

of study participants, community engagement professionals, privacy technology colleagues, advisors, funding agencies

Fun

din

g Sou

rces

NIH R01HG008802

NIH U24AI117966 Department of Veterans AffairsI01HX000982

Patient-Centered Outcomes Research InstituteCDRN-1306-04819

NIH T15LM011271

NIH U54HL1084600