Making Health Data Work

for Patients and Populations

Iain Buchan

Farr Institute @ HeRC

& University of Manchester

www.herc.ac.uk

Koç University Hospital, 20th March 2017@profbuchan

#DataSavesLives

Ancient, yet Relevant, Public Health Statistics

Plague Of Justinian (541–542)

• 40% Constantinople dead

• 25-50 m dead globally over 200y

Black Death (1338/1346–1353)

• China to Constantinople by 1347

• 30-60% Europe’s population dead

Wagner DM et al. Yersinia pestis and the plague of Justinian 541-543 AD: a genomic analysis.

Lancet Infect Dis. 2014;14(4):319-26.

London 1600s: Plague; Fire; Data; SocietyParish (deaths)

Number of hearths (fireplaces) as a proxy for house size and over-crowding

1 2 3 4 5 6 7 8+

St James Clerkenwell (44)

St Botolph without Aldgate (41)

St Dunstan in the West (49)

St Michael Queenhithe (20)

St Saviour Southwark (42)

31.7

39.8

31.2

26.5

15.5

10.7

35.1

24.3

14.2

30.0

20.8

19.3

25.2

33.1

11.7

7.5

24.3

12.9

31.8

33.5

24.7

16.7

22.3

24.4

12.6

7.3

16.2

20.0

28.2

20.1

9.7

7.3

8.3

5.1

20.2

14.3

5.4

7.1

12.4

7.5

2.6

3.9

4.7

3.8

8.6

11.9

5.4

15.0

6.9

3.8

3.2

2.7

4.7

4.1

5.0

10.5

2.7

10.0

4.3

2.9

1.2

1.9

1.5

1.5

4.7

8.7

2.7

2.1

0.7

0.9

6.2

8.4

2.2

1.7

21.7

29.1

8.1

8.5

1.5

1.3

From: Epidemic Disease in London, ed. J.A.I. Champion (Centre for Metropolitan History Working Papers Series, No.1, 1993): pp. 35-52

…any man's death diminishes me, because I am involved in mankind,

and therefore never send to know for whom the bells tolls; it tolls for thee.

Epidemiology and politics preceding the Great Plague of London (1665; 25% population die)

followed by the Great Fire of 1666 then social and structural reform.

John Donne, London 1624

Health Data Computation: 1841

Farr

(1807-1883)

HPC for life tables public health reform

1700s: Bernoulli & DeMoivre introduce

probability theory to quantifying (health) risks

Early 1800s: Laplace then Louis apply

probability theory to showing some treatments to be ineffective

– rebuked by medical profession

– Quetelet’s concept of ‘the average man’ adds fuel to the fire

Letting the data speak computationally…

Babbage

(1791-1871)

Evidence Based Care

Mid-late 1800s: Lister uses statistical arguments and

Pasteur’s germ theory to

revolutionise surgery with carbolic spray

Early 1900s: Statistical Movement,

strong in Agriculture and emerging in Medicine

Mid 1900s: Experimental (statistical) discipline into Medicine

and NHS founded (1948)

1970-80s onward: Disciplined implementation

of evidence into practice

NHS: Learning System Legacy and Duty

30 years of GPs coding

in routine primary care

Needs-based

Constitution

NHS Computability: 1970s onward

• Administrative consistency: Körner to ICD to HES to QOF

• Clinical utility: GP home-grown IT to patient apps

• Research integration: VAMP to CPRD & registries to Farr

1988 AAH MEDITEL advertisement courtesy of Tim Benson

30 Years of Structured Primary Care Data

Schulz EB, Price C, Brown PJ. Symbolic anatomic knowledge representation in the Read Codes version 3:

structure and application. J Am Med Inform Assoc. 1997 Jan-Feb;4(1):38-48.

UK: 30 years of GPs coding in routine primary care

Healthcare Data Analytic Partnerships

MISSED OPPORTUNITIES DETECTOR

Find patients relevant to

care pathway

Exclude if target

inappropriate

e.g. A&E asthma terminal illness

Exclude if target

achieved

Follow-up < 48h

Identify how care could be

improved

Rx & social review

Integrated Care Record

BLIZZARD OF DATABASES

(Salford: 53 GP offices + 1 Hospital)

Salford Resident Population

Care Quality Management

Patients’ Decisions

ACTIONABLE INFORMATION

Actionable information

attracts: trust & traction

from patients, public and

practitioners… and better

data quality.

Brown B et al. Missed opportunities

mapping: computable healthcare

quality improvement. Stud Health

Technol Inform. 2013;192:387-91.

Theoretical Framework

Perception Acceptance

Desire Action

Clinical performance

Intention

Data analysis

Message

Data collection

Interaction

Organisational

Individual patient

Verification Unintended outcomes

Task Action Audit Message

PatientCo-

interventions Recipients Organisation

USABILITY/DESIGN

TEAM DELIVERY

ALGORITHM ACCURACY

DATA CREDIBILITY

ACTION PLANNING

NHS: a decade of dashboards

Business intelligence tools

Provider management led

No theoretical framework

Connecting Population Analytics with Care

• Audit & Feedback Theory• Eye Tracking Experiments• Field Trials in Salford, UK• Patients Asking for Safety Alerts• Now Targeting Antimicrobial Resistance

Instrumenting Alternative Views

Fraccaro, P et al. "Patients’ online access and interpretation of laboratory test results: a human computer interaction study”. Digital Health and Care Congress 2016: www.kingsfund.org.uk/events/digital-health-and-care-congress-2016

In search of a better conversation of healthcare over shared records, brokered by informatics that is context-aware.

Reusable Health Analytics: Trials & Audits

National Proteomics Centre:

Stoller Biomarker Discovery

Clinical Audits

and Service Planning

for the local population

Evidence Translation Singularity

Evidence

Practice

Evidence

Practice

Implementation

Evidence

(Co-)Practice

System

Asynchronous Healthcare Information Synchronous Translation

More Specialisation

More Underpinning Biology

Precision Medicine?

Ironing out Variation in Care

Real-world Evidence

Managed Self-care

Translation

Problem: Big Data & Blunt Evidence

Mental health team…

Primary care team…

Zak… 47y; asthma since early childhood; schizophrenia since teenage; overweight; smoker

Weight

Respiratory team… Evidence needed is theunion not sum of models

~ Mood

Primary Care

Respiratory Medicine

PsychiatrySmoking cessation;

social support; weight control; work

Inhaled steroid adherence

Antipsychotic medication adherence

Over-implementing Blunt Evidence

Current evidence-base predicts < 30% healthcare outcomes: so why try to “iron out variation”?

Primary

Care

Respiratory

Medicine

Better healthcare needs information

on how care works across diseases,

providers and daily-life contexts

Psychiatry

Valderas JM, Starfield B, Sibbald B, Salisbury C, Roland M. Defining comorbidity: implications for understanding health and health services. Ann. Fam. 2009;7:357–363.

Precision Medicine Horizons: Asthma(s)• Life-course complexity indicates multiple (sub-)diseases

• Usually starts young• May progress, remit or relapse over life

• Inconsistent gene-environment interactions indicates multiple (sub-)diseases• Variable effects of genetic polymorphisms, e.g. CD14• Variable treatment-setting interactions

C allele associated

T allele associated

No association

CD14 Endotoxin Receptor

Simpson A et al. Endotoxin exposure, CD14, and allergic disease: an interaction between genes and the environment.

Am J Respir Crit Care Med. 2006;174(4):386-92.

50-60% heritability in twin studies but < 2%

phenotype explained by current genomics

Seeking Sub-disease Patterns in Data

Mite

Cat

Dog

Pollen

Egg

Milk

Mold

Peanut

Sensitized

Age 1

Sensitized

Age 3

Sensitized

Age 5

Sensitized

Age 8

Skin Test

Age 1

Skin Test

Age 3

Skin Test

Age 5

Skin Test

Age 8

Blood Test

Age 1

Blood Test

Age 3

Blood Test

Age 5

Blood Test

Age 8

Sensitization Groupswitch group

P(Sens’n)

in year 1

P(Gain)

P (Loose)

Sens’n

3 intervals

P(+ skin)

Sens’

P(+ skin)

Not Sens’

P(+ blood)

Sens’

P(+ blood)

Not Sens’

Sens’n state

1,053 Children

8 Allergens

Machine-learning software& partial statistical models

ATOPY

(allergic tendency)

Crude clinical label

not explained by

genomic studies

New Risk Factor for Asthma Discovered

Allergic sensitisation

patterns ‘learned’ from data

Simpson A, Tan VY, Winn J, Svensén M, Bishop CM, Heckerman DE, Buchan I, Custovic A. Beyond atopy: multiple

patterns of sensitization in relation to asthma in a birth cohort study. Am J Respir Crit Care Med. 2010;181(11):1200-6.

Multiplying Analytic Capacity via eLab

MAAS

SEATON

ASHFORD

ALSPAC

IOW

Modelling

Data & Harmonized

Metadata from Cohorts

Data Extracts

Networking:

Ideas, Activities,

Results, Meanings

MRC STELAR and NIH CREW Consortia: www.asthmaelab.org

New US

New Au.

Custovic A et al. The Study Team for Early

Life Asthma Research (STELAR) consortium

'Asthma e-lab': team science bringing

data, methods and investigators together.

Thorax. 2015;70(8):799-801.

• Progression of allergy

Eczema → Asthma → Rhinitis

• Inferred from population summary →

• Assumed causal link between eczema – asthma & rhinitis

• Clinical response:

target children with eczemato reduce progression to asthma

Received Wisdom: Atopic March

Spergel & Paller, 2003

World Allergy Organization, 2014

Model-based Machine Learning

Probability Eczema Age 8

Children (n=9801)

Probability Eczema Age 5

Probability Eczema Age 3

Probability Eczema Age 1

Probability Eczema Age 11

Probability Wheeze Age 8

Probability Wheeze Age 5

Probability Wheeze Age 3

Probability Wheeze Age 1

Probability Wheeze Age 11

Probability Rhinitis Age 8

Probability Rhinitis Age 5

Probability Rhinitis Age 3

Probability Rhinitis Age 1

Probability Rhinitis Age 11

Eczema Class

Wheeze Class

Rhinitis Class

Latent Class Disease Profile

Start with a well-reasoned

(partial) model, not a

‘bucket of data’

Ecologic Fallacy Revealed

Belgrave et al. Developmental Profiles of Eczema,

Wheeze, and Rhinitis: Two Population-Based Birth

Cohort Studies.

PloS Medicine 2014;21;11(10):e1001748.

MRC STELAR consortium working at scale

across MAAS and ALSPACS cohorts

Better:

population analytics;

targets for ‘omic research

Biology-Behaviour-Environment Interaction

Life course

Developmental genetics

Disease risk environment

Treatment environment

‘Persistent’ genetics

Data

Measurement error

Mechanism knowledge

Missingness

ASTHMA

genes

* environments endotypes

Transient early wheeze

Late-onset wheeze

Persistent troublesome wheeze

Persistent controlled wheeze

My Health Data Ecosystem

My Health, My Data:

Where are the most predictive data?

Rhythms of Life, Health, Disease and Care

Low-cost ubiquitous technologies capturing

digital by-products of the life

High-cost medical devices

(regulated clinical algorithms)

Clin

ic v

isit 8

Clin

ic v

isit 9

Patterns of disease invisible to

infrequent clinical observation

Precision medicine may need data on

(sub)disease rhythms to realise its

potential

Future? My ‘health avatar’ says no to your care pathway

n-of-1 trials Average patient guidelines

• Who self-weighs?

• UK Withings smart-scale users vs. Health Survey for England (2011)

Ubiquitous (Almost) Technology

Difference in Mean BMI

Fatter men use smart-scalesSlimmer women use smart-scales

-2 1 0 1 2

Men point-estimate = 1.26 [95% CI: 0.84,1.69]Women point-estimate = -1.62 [95% CI: -2.22,-1.03]

Sperrin M et al. Who Self-Weighs and What Do They Gain From It? A Retrospective Comparison Between

Smart Scale Users and the General Population in England. J Med Internet Res. 2016;18(1):e17.

• What came first, weighing or weight-loss?

Complex Frequent Observation/Intervention

Engagement

Weight Loss

• An additional monthly weighing is associated with

an extra 1kg weight lost over the course of a year

• Recent weight loss encourages subsequent

measurement: a person who has recently lost 1kg is

twice as likely to reweigh on a given day compared

with someone who has remained the same weight

Sperrin M et al. Who Self-Weighs and What Do They Gain From It? A Retrospective Comparison Between

Smart Scale Users and the General Population in England. J Med Internet Res. 2016;18(1):e17.

Rethink Experimental Designs

Dwyer T et al. Objectively Measured Daily Steps and Subsequent Long Term All-Cause Mortality:

The Tasped Prospective Cohort Study. PLoS One. 2015;10(11):e0141274.

Co-produced Observations & OutcomesAim: To Reduce Relapse in Schizophrenia via Smartphone

Drug + behaviour (information * psychological endotype) = outcome

From J. Ainsworth & S. Lewis

Informatics enabled observation

Informatics intervention

www.clintouch.com

Generic:

• Self-measurement

• Symptom awareness

• Clinical workflow integration

• Self-efficacy / autonomy

• Alert-fatigue avoidance

Civic AND Clinical Analytics

Smartphone GPS data

infer social functioning

in patients with

schizophrenia

From Difrancesco et al. Out-of-home activity recognition from GPS data in schizophrenic patients. IEEE 29th International Symposium on Computer-Based Medical Systems (CBMS 2016).Sport

Swimming pool

Volleyball

1. Raw GPS data

2. Detection of geolocation visited

3. Geolocations visited

4. Identification of places visited

5. Places visited

6. Type of places and activities recognition

7. Out-of-home activities

Health & Care Context: PlaceTypical health & care data landscape: Greater Manchester

• £6 billion annual care budget

• >7000 care provider databases

• Local bye-law “duty to share” equal to “duty to protect”

• 2.7M people with low life expectancy and high inequalities

Data Sharing: Diameter of Trust Actionable

information for

health system

optimisation

National/Large Population

Audits/Registers/Monitoring

NOT SCALABLE

Excellence provider benchmarking e.g. strokeaudit.org but no learning

across disease areas and not integrated with clinical workflows

SCALABLE

Payer evidence, quality management, public health

intelligence and research share data, infrastructure and expertise

Large enough for

economy of scale

Small enough for

a conversation with

the citizenry

about data sharing

www.herc.ac.uk/get-involved/citizens-jury/

Trust in Predictive Analytics

Academia rewards

publishing papers on the

same topic every 10y or so

Law sees algorithms as

medical devices

(EU Directive 2007/47)

Industry has no trusted 3rd

party lab for validating

algorithms/models

EuroScore prediction

Calibration drift:

Typical of many

published models

Observed death rate

Hickey GL et al. Dynamic trends in cardiac surgery:

why the logistic EuroSCORE is no longer suitable for

contemporary cardiac surgery and implications for

future risk models. Eur J Cardiothorac Surg. 2013

Jun;43(6):1146-52.

Civic Health Data Analytics

Data

Public sectorencounters

Services

Targetedby need

TargetingTools

Ark

Involved CitizensProblem OwnersData Managers

Public Health AnalystsCare Service Analysts

StatisticiansInformaticians

Social ScientistsHealth Economists

Health Service Researchers

Communications Experts

Service Planning

and PolicyInsights

SME Global Corp.

Which services and how?

Spin-in/out Laboratory

Farr Institute & NIHR Centres

Connected Health CitiesPilots 2016-9North England

twitter.com/hashtag/datasaveslives

Ainsworth J, Buchan I. Combining Health Data Uses to Ignite Health System Learning.

Methods Inf Med. 2015 Nov 27;54(6):479-87.

www.connectedhealthcities.org

Time

Data Production

Analysis

Data ProductionData-Action Latency

Insight

ActionDataPreparation

Stroke Pathway Learning

1 3

2

Better decision support

tools for paramedics:

Recognise ‘stroke mimics’

Faster, more accurate triage

and improve access to

neurosurgery when needed

Enhanced workflows e.g.

medication vs. BP reviews to

prevent another stroke

• Smartwatch detects atrial fibrillation over a week,

otherwise missed by a GP in a 10 minute consultation:

then supports anticoagulant medication.

• Virtual rehab assistant (voice/AI appliance) and smart electricity meter

data alert rehab team to a change in daily living patterns.

• Subsidised public transport after cardiovascular risk screening

makes it cheaper to walk/tram than take the car to work:

increased physical activity sustained where exercise prescriptions fail.

Stroke Prevention: Civic Extensions

Borrowing Insights

De-identified Records

IdentifiedRecords

StudyProtocol

/Assessment

StudyRecruit

ClinicianResearcher

Commons of Metadata and Information Governance (Clinical & Research)

Clinical Care

Patient

Research Safe Haven

Encrypted (SHA1 & AES256);

Certified (ISO 27001)

System 1

System 3

System 2

System 4

Linkable Data Providers

Analytic

Objects

RAPID REPLICATION

• Study/audit protocol

• Codes for the data

• Statistical scripts

• Results in progress

• Report

• Slides etc.

Bechofer S, Buchan I et al. Why linked data is not enough for scientists.

Future Generation Computer Systems 2013;29(2):599–611.

Ainsworth J, Buchan I. e-Labs and

Work Objects: Towards Digital

Health Economies. Lecture Notes

of the Institute for Computer

Sciences, Social Informatics and

Telecommunications Engineering.

Springer, Berlin Heidelberg

2009;16:206-216.

Routine

Randomisation

“Learning Healthcare Systems”

are an illusion if restricted to

provider organisations.

Health(care) can’t be

optimised outside the civic

context of health.

Civic Imperative

@profbuchan

#DataSavesLives

• Grasp patient-led healthcare records/apps for self-care and clinical

workflow improvement: don’t focus on legacy clinical IT.

• Allow cities/regions to self-organise around healthcare workflow

optimisation, and join the global Connected Health Cities: prepare for

streams of place-based data that affect health and care.

• Exploit the diversity of biology-behaviour-environment interactions as a

global science and technology innovation asset as Turkish digital health

data start to join up.

Thoughts for Turkey’s Digital Health