1. Enabling Health E-Research at UK Scale Iain Buchan Director, Farr Institute @ HeRC 10 th July 2014 MRC Open Council, Newcastle

2. From Big Data to Big Scale DATA EXPERTISEMETHODS & MODELS Vast data volume, velocity, variety TSUNAMI Vast data volume, velocity, variety TSUNAMI Supra-linear growth in papers & tools BLIZZARD Supra-linear growth in papers & tools BLIZZARD Similar number of analysts DROUGHT Similar number of analysts DROUGHT Three Big Health E-Research Challenges 1.Assist hypothesis formation with data 2.Weed out non-reproducible findings early 3.Couple data-intensive healthcare and research

3. Shaping Hypotheses with 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 Group switch 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

4. New Asthma Risk Factor Found Allergic sensitisation patterns ‘learned’ from data

5. Myth Bust Working Across Cohorts From: MRC STELAR: MAAS & ALSPAC 2013 (D. Belgrave et al. pre-publication) Large-scale analysis: Atopic march < 5% From: Spergel & Paller, 2003 Received wisdom “Atopic March”

6. Searching Clinical Data Clinically Database Searches select distinct PatientID, GPPracticeCode from Patients where DeathDate is null and PatientID in (select distinct PatientID from Patients where Dob =1931 ) ) and PatientID in (select distinct PatientID from Patients where PatientID in (select distinct j.PatientID from Journal j where j.ReadCode in ('.C21.','.C2A.','.C2D.','.C2G2','C1021','C109.','C109 4','C1095','C1097','C1099','C109D','C109F','C109G',' C109H','C109J','C10F.','C10F4','C10F5','C10F7','C10F 9','C10FD','C10FF','C10FG','C10FH','C10FJ','L1806','X 40J5','X40J6','X40Jk','XaELQ','XaFn7','XaFn8','XaFn9','XaFWI','XSETp','XU70f','XU71F','XUKO0','XULXc','XUP Hn','XUSbx') and j.EntryDate<@p4DateLimit1 ) ) and PatientID not in (select distinct PatientID from Patients where PatientID in (select distinct j.PatientID from Journal j where j.ReadCode in ('.1226','.12C2','.12C3','.12C5','.12C8','.12CA','.12CB','.12CC','.12CD','.12CE','.12CF','.12CG','.12CH','.12C J','.12CL','.12CM','.12CN','.12CP','.12CR','.12CS','.12 CT','.12J3','.14A.','.14A3','.14A4','.14A5','.14A6','.14A D','.14AH','.14AJ','.14AL','.14AM','.14AN','.14AP','.14 AQ','.14AR','.14H1','.187.','.1I10','.1I3.','.1I5.','.1I6.','.1J 60','.1O1.','.2241','.679X','.68B2','.68B6','.6C0.','.7721','.7722','.865.','.8651','.8652','.8653','.8654','.865Z','.B1N Z','.G12.','.G121','.G122','.G123','.G12Z','.G131','.G2.. ','.G21.','.G211','.G212','.G213','.G21Z','.G22.','.G221','.G222','.G223','.G22Z','.G23.','.G231','.G232','.G233','.G234','.G235','.G236','.G23Z','.G2Z.','.G32.','.G34.','.G4..','.G41.','.G42.','.G420','.G43.','.G44.','.G440','. Analysts Drowning in Codes Interactive Clinical Searches Clinicians Exploring Data

7. Using Clinical Codes Reliably From E. Kontopantelis & T. Doran Consider the GP annotation on a diabetes code “DM r/o” Consider Vioxx MI risk detectable pre-2005 via text not coded data Bias from different: population samples; clinical IT; coding practices; data cleaning

8. Public Trust: Safe Haven Network De-identified Records Identified Records Study Protocol /Assessment Study Recruit Clinician Researcher NHS and Academic Health Science Information Governance Clinical Care Patient Research Safe Haven Encrypted (SHA1 & AES256); Certified (ISO 27001) Farr 1 Farr 3 Farr 2 Farr 4 Linkable Data Providers Research Objects RAPID REPLICATION Research protocol Codes for the data Statistical scripts Results in progress Draft manuscript Slides etc. RAPID REPLICATION Research protocol Codes for the data Statistical scripts Results in progress Draft manuscript Slides etc.

9. Scaling-up Replication via Farr Phenotyping Challenge: Automate classification of diabetic patients as “kidney protected” or not with regard to other risk factors

10. (Re-)Validating Clinical Algorithms Production line of clinical prediction models is broken EU Directive 2007/47: The law now sees algorithms as medical devices From G. Hickey & B. Bridgewater EuroScore Typical calibration drift

11. Challenging Models From/For NHS From G. Hickey & B. Bridgewater Relating mortality from cardiac surgery to lengths of service of surgeons

12. Data from Mobile Technologies Aim: To Reduce Relapse in Schizophrenia via Smartphone Drug + behaviour (information * psychological endotype) = outcome From J. Ainsworth & S. Lewis Informatics enabled observation Informatics intervention

13. Data-responsive Health Systems PUBLIC TRUST PATIENT TRACTION INTEGRATIVE INFORMATICS Patient / Mobile DataShared RecordCore OutcomesClinical Performance NetworksRoutine RandomisationLearning from Other SystemsDynamic Prediction Models Big Data & Big Sense Making From data castles… To large scale research over shared resources, as in Physics

14. Farr Institute Up North 15M population 20% greater risk of dying <75 years c.f. South England Farr-AHSN integration with NHS >75% of UK Health Informatics methods papers 2012-14 Potential N8 network of “data-responsive health systems”