Buchan he rc_uk_launch

255 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
255
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
3
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide
  • “Harnessing” is more than getting hold of data, it is about making the data work for patients and the public.HeRC is one of four health informatics research centres that are being established to provide the methods and expertise to harness such data at scale.This presentation is an overview of the mission of HeRC, and the national network, of harnessing health data for patient and public benefit.
  • Current medical evidence predicts less than a third of what will happen to patients when treated. This proportion is greater for those with multiple conditions. Specialism compounds the problem: consider Mrs Jones, a type 2 diabetic who sees her diabetologist who focuses on blood sugar control, which puts up her weight. More excess weight puts up her blood pressure, stressing her kidneys, worsening her chronic kidney disease. She sees her nephrologist who focuses on blood pressure control. Mrs Jones’ GP is more concerned by general cardiovascular risk factors and lifestyle. Mrs Jones is not the sum of three care pathways. Care interacts and realistic models of care are complex. Informatics is about harnessing such complexity to produce useful information.Fortin, M., Dionne, J., Pinho, G., Gignac, J., Almirall, J. & Lapointe, L. 2006 Randomized controlled trials: do they have external validity for patients with multiple morbidities. Annals of Family Medicine. 4:104–108.Valderas, JM., Starfield, B., Sibbald, B., Salisbury, C. & Roland, M. 2009 Defining comorbidity: implications for understanding health and health services Ann. Fam. Med. 7(4):357–363.
  • Three ingredients are needed to harness the complexity of care outcomes: data, models/methods, and expertise. We have a tsunami of data, and blizzard of models from the literature, but a drought in the human expertise to make sense of the growing data. So viewing ‘big data’ as a solution is a myth. Similarly it is a myth to consider that science will provide all of the models Informaticians need to turn healthcare data into decision support algorithms to improve patient outcomes – such models evolve as I will show later. The third myth is that clinicians will continue to be the main source of data, when, in a world of increasingly mobile technologies we all collect data about our daily lives as a matter of routine. Along with three myths are three main pipelines of mixing data, models and expertise together: for clinical audit or quality improvement; for public health and commissioning purposes; and for research. This dilution of effort, especially in human resource for distilling healthcare intelligence, is wasteful, and fails to borrow strength between these potentially synergistic areas of expertise.
  • Even with very messy data, given appropriate statistical modelling we showed an effect of starting glitazone therapy that is consistent with that seen in clinical trials, and it is biologically plausible. So messy data are very useful, especially when patients act as their own controls. This is true in long running electronic collections of records of people with long term conditions, which we have around Manchester.
  • Different sources of GP data lead to different results even over the simplest questions. Here we see large differences in diabetes prevalence between GPRD, THIN, Q Research and QoF by using each source’s data cleaning and coding rules. Informatics looks behind the codes, for example cleaning out the diabetes code with a GP annotation “DM r/o”. Colleagues at Stanford could not pick up the heart attack risk over Vioxx used in rheumatoid patients when they considered the coded data, but by text mining the data to encode more information the safety signal emerged – a signal that was discernable well before the drug was withdrawn.
  • A trial feasibility analysis and recruitment system rooted in health informatics research, winning the 2009 HealthGrid prize.Successful deployment in UK research networks through NorthWest eHealth, our NHS based services arm.
  • Another context is the law, which now treats software used to support clinical decisions as medical devices. Consider the software or algorithms used to predict the risk of dying while undergoing coronary procedures such as PCI or CABG. Since publication of the widely used Euroscore model in the late 90s there has been a drift from what the model predicts (in red) to what is actually observed in terms of patient deaths (in black).
  • The Salford Lung Study is a good example of borrowing strength. This is an open-label trial of Relovair, a combined steroid and beta agonist whose real world clinical effectiveness depends on improved adherence due to a more convenient dosing schedule (once rather than twice per day). Adherence is known to vary with socio-economic status. In Salford, the nationally accepted Index of Multiple Deprivation is unreliable, because the crime statistics were miscalculated. The Public Health team know this, so the researchers in Salford Lung Study can borrow strength from the Public Health team if they are connected in an eLab. Likewise the clinical teams auditing readmission risk with acute exacerbation of COPD can borrow better socio-economic metrics. The ideal health intelligence is a multi-disciplinary sense-making network.
  • Health informatics research leading to practical tools for connecting health professionals with their patients’ data for deeper, more evidence based clinical audit.
  • Data collected by patients, for example via smartphone apps, will provide a much richer longitudinal signal of patient experience and outcomes, and a platform for supporting self-care.One of the areas of decision support we are working on is schizophrenia medication adherence and the risk of relapse. Using the principles of cognitive behavioural therapy and experience sampling we have developed a mobile phone application that extends the ‘care loop’ from the usual six weekly checks by community psychiatric nurses to a more adaptive system. When patients’ responses indicate that their medication and symptoms are out of kilter the patient can be prompted to take action, the care team can also be prompted – in this way the scarce community resources can be targeted to those who need them most, and the overall schizophrenia relapse rates may be reduced.
  • Here is an example of sense-making that combines the reach of data mining with the focus of hypothesis driven research. We call it machine learned epidemiology. The challenge in this example was the bust the myth that children simply have an allergic tendency (diagnosed “atopic”) or not. By studying just over a thousand children in the Manchester Asthma and Allergies Study we searched for structure in the data pointing toward children with different patterns of something we can’t measure directly, i.e. allergic sensitisation. Indirect measures of sensitivity to mite, cat, dog etc. antigens were made from skin prick and blood tests. Bayesian inference algorithms were then used to find latent classes of children gaining and losing sensitisation to these allergens.
  • This machine learning exercise found a class of children with “multiple early” sensitisations, with an odds ratio of 31 of developing asthma by age 8. This dataset had very detailed lab, parent-reported and clinical measures of asthma. This group had not been hypothesised, and further studies of them may lead to discovery of important new targets for prevention.
  • A patient’s experience is not lost if their data are harnessed to improve care for future generations.This is intellectually challenging.MRC has led the charge to meet this challenge with critical masses of informatics.
  • Buchan he rc_uk_launch

    1. 1. Health eResearch: North EnglandEngineering scalable health science for patient and public benefitIain BuchanDirector, Health eResearch Centre (North England)UK Network Launch1st May 2013
    2. 2. Primary Care Secondary CarePublic Health Self CareSpecialist A Specialist BDiabetology:Glucose focusNephrology:Blood pressure focusLinked Self-Care:Weight controlPhysical activity…Evidence predicts <30% outcomes.Building usefully complex models of health & careneeds a much larger scale of research, ‘in the wild’. Weight  Blood pressureEvidence Engine Does Not Scale
    3. 3. New Evidence from Health Records?ResearchCommissioningGovernanceDiabetesCancerTSUNAMI of DATABLIZZARD ofMODELS/METHODSand LITERATURESilo k…DROUGHT of EXPERTISEPROLIFERATION of PIPELINES
    4. 4. Improved liver function (pooled) = 0.15 (se=0.009), p<1x10-59Similar effect alone or with other drugs {G = 0.12; G+M = 0.16; G+S = 0.16; G+S+M = 0.15G = glitazone, M = metformin, S = sulponylurea}First glitazone RxMean log(ALT) beforeMean log(ALT) afterFitted trends do not join  evidence of glitazone effect on ALTLiver Function vs. GlitazonesFrom M. SperrinTrends in liver function among patients with type 2 diabetes:messy real-world data; effect consistent with best evidence.
    5. 5. Harnessing Codes for Clinical TrialsBefore FARSITE…select distinct PatientID, GPPracticeCode from Patientswhere DeathDate is null and PatientID in (select distinctPatientID from Patients where Dob<=1991 and PatientIDin (select distinct PatientID from Patients whereDob>=1931 ) ) and PatientID in (select distinct PatientIDfrom Patients where PatientID in (select distinct j.PatientIDfrom Journal j where j.ReadCode in(.C21.,.C2A.,.C2D.,.C2G2,C1021,C109.,C1094,C1095,C1097,C1099,C109D,C109F,C109G,C109H,C109J,C10F.,C10F4,C10F5,C10F7,C10F9,C10FD,C10FF,C10FG,C10FH,C10FJ,L1806,X40J5,X40J6,X40Jk,XaELQ,XaFn7,XaFn8,XaFn9,XaFWI,XSETp,XU70f,XU71F,XUKO0,XULXc,XUPHn,XUSbx) andj.EntryDate<@p4DateLimit1 ) ) and PatientID not in (selectdistinct PatientID from Patients where PatientID in (selectdistinct j.PatientID from Journal j where j.ReadCode in(.1226,.12C2,.12C3,.12C5,.12C8,.12CA,.12CB,.12CC,.12CD,.12CE,.12CF,.12CG,.12CH,.12CJ,.12CL,.12CM,.12CN,.12CP,.12CR,.12CS,.12CT,.12J3,.14A.,.14A3,.14A4,.14A5,.14A6,.14AD,.14AH,.14AJ,.14AL,.14AM,.14AN,.14AP,.14AQ,.14AR,.14H1,.187.,.1I10,.1I3.,.1I5.,.1I6.,.1J60,.1O1.,.2241,.679X,.68B2,.68B6,.6C0.,.7721,.7722,.865.,.8651,.8652,.8653,.8654,.865Z,.B1NZ,.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,.G441,.G442,.G443,.G444,.G445,.G446,.G45.,.G451,.G452,.G45Z,.G46.,.G47.,.G48.,.G49.,.G4Z.,.G5..,.G51.,.G511,.G52.,.G52Z,.G5Z.,.G6..,.G61.,.G612,’.After FARSITE…Analysts drowning in codesResearchers exploring trialfeasibility interactivelyReuse this investment…www.opencdms.orgopenCDMS
    6. 6. 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 20110.020.030.040.050.060.070.08 New Laws for Clinical PredictionYearIn-hospitalmortalityproportionObserved UK (SCTS) deaths from cardiac surgeryExpected UK (EuroSCORE I model) deaths from cardiac surgeryDrifting EuroSCORE CalibrationNote EU Directive 2007/47From G. Hickey & B. Bridgewater
    7. 7. Service-Science Synergy“Users who selected thevariables in your basket alsoselected these variables andthese models…”Enabling the world’s largest ‘real world’respiratory clinical trialClinical Trial:1 vs. 2 per day dosing adherence (* deprivation)Clinical Audit:Depression vs. readmissionPublic Health:Recalibrated deprivation score
    8. 8. COCPIT: Revealing Care PathwaysInput: Linked health record data and clinical guidelinesOutput: Pictures of observed care contrasted with expected care:helping health professionals to identify missed opportunities forbetter care and support service improvement
    9. 9. Integrating Patient-reported DataCareLoop: Reducing Relapsein Schizophrenia via SmartphoneFrom J. Ainsworth & S. Lewis
    10. 10. MiteCatDogPollenEggMilkMoldPeanutShaping Hypotheses: AtopySensitizedAge 1SensitizedAge 3SensitizedAge 5SensitizedAge 8Skin TestAge 1Skin TestAge 3Skin TestAge 5Skin TestAge 8Blood TestAge 1Blood TestAge 3Blood TestAge 5Blood TestAge 8Sensitization Classswitch classP(Sens’n)in year 1P(Gain)P (Loose)Sens’n3 intervalsP(+ skin)Sens’P(+ skin)Not Sens’P(+ blood)Sens’P(+ blood)Not Sens’Sens’n state1,053 Children8 AllergensModel-based machine learning
    11. 11. Atopy ‘Stratified’Important new risk group for asthma discoveredFrom A. Simpson & A. Custovic
    12. 12. Social Machine of Health eResearchLocalCommunityIntegratedHealthRecordDe-identified recordsUnified localhealth intelligenceplatformLocal NHS 1ResearchObjectUni. + NHS 1Science leveraging datasize and heterogeneityLocal NHS 2Uni. + NHS 2Service multiplyinganalyst capacity• Query• Dataset• Statistical code• Report• Slide deck…
    13. 13. InformaticsTrainingPatient & PublicInvolvementEngineering& Methods ScienceClinical TrialsEndotypeDiscoveryLinked Self-careScalableEpidemiologyScalableHealthcare QI
    14. 14. Health eResearch Centre: N. EnglandScience and Industry(R&D)Improved Care forPatients and Communities(Service)LinkValueLinkIngredientsDatasets MethodsExpertsDataQualityInsightsSuccess = Informatics releasing patient and public benefit from health records at scale<New N8…>

    ×