This document summarizes a presentation about identifying deficiencies in long-term condition management using electronic medical records. It discusses using data mining of electronic medical records to analyze hypertension management and electronic referrals. Case studies show opportunities for improved monitoring and treatment of long-term conditions were identified. The presentation encourages using available electronic health record data to help improve healthcare processes and outcomes.

1. Identifying deficiencies in long-
term condition management
using electronic medical records
Prof Jim Warren
Chief Scientist, National Institute for Health Innovation
Chair in Health Informatics
The University of Auckland
The National Institute
for Health Innovation

2. Overview
• Why we do long-term condition data mining /
analysis
• Case study in hypertension management
– Context, tools/methods, findings and implications
• Case study in electronic referrals
• What’s it mean for you?

3. Why the interest?
• Burning platform
– Increased rates of chronic illness (that’s not entirely
bad, btw)
– Ageing population
– Cost / workforce / delivery meltdown!
• Lots of data
– Computerisation naturally lays down data
– Seems wrong not to use it
• Both for refinement of processes, and to guide the search
for more radical transformations in health delivery

4. Chronic conditions, esp. Hypertension
(or anything ‘vascular’)
• Amenable (and interesting) for analysis of long
sequences in transactional electronic health records
(EHRs)
• Blood pressure (BP) is a huge risk factor
– Cardiovascular disease (CVD) risk doubles for every 20/10
mm Hg
• Also implicated in kidney failure
– Very controllable with medication
– Several classes of medication with different side-effects
and benefits/indications (makes it interesting!)
• Statins/cholesterol and blood sugar control also very
relevant

5. General Practice Computing
• Highly computerized
– High percentage penetration and good maturity/depth of use
– In common with Australia and UK
• NHI
– Well-established National Health Index numbers for patients
– Allows research linkage to national data collections (notably PHARMAC
for dispensing)
• Variable (but good) rates of problem classification
– Dx often coded with Read Codes ver 2 (rogue former UK system)
– Lower quality entry of local observations (e.g., BP)
• Electronic test results and hospital discharge summaries
received as HL7 messages
– GP can accept lab result items into their database

6. Study context
• General practice PMS (practice management system) software data
include: electronic prescribing, lab test results review, problem
lists, observations (e.g. BPs), practice notes
• Work with West Fono
Health Care
– Pacific led practice in
West Auckland
– Iterative analysis of PMS
data to identify opportunity
for improvement in
management of long-term
conditions

7. Criteria model
• Abstracted audit classes from general practice
opportunities for quality improvement
Criterion
Failure to Sustained Contra-
Unsustained
Measure Failure to indicated
Treatment
Outcome Meet Target Treatment
Lapse, low MPR
(medication
possession ratio)

8. ChronoMedIt Architecture
PMS Cohort Report
1
Data
Canonical Query
Uptake Case Report
EMR Processor
App
PMS 4
database Timeline Graph
Criterion 5
3
Criteria
Domain Authoring
Ontology App
2
Oral presentation (full paper) – MIE 2011, 28-31 August 2011 Warren Slide 8

9. Visualisation: Bad pattern, low MPR
Oral presentation (full paper) – MIE 2011, 28-31 August 2011 Warren Slide 9

10. Good pattern

11. Ontology (OWL, maintained w/ Protégé)

13. State-transition model
• Times at which therapy changes indicate
opportunities to critique performance / decision
• Designed set of alerts with 26% sensitivity and 93%
specificity by GP panel assessment
Developing high-specificity anti-hypertensive alerts by therapeutic state analysis of
electronic prescribing records. JAMIA 14(1), 2007
Gadzhanova S, Iankov II, Warren JR, Stanek J, Misan GM, Baig Z, Ponte L.

14. Temporal reasoning on intervals
• There are a LOT of cases to consider when evaluating a
couple of transactions - each with a temporal ‘shadow’ -
against an interval
– Practical contribution of ChronoMedIt is to ‘funnel’ a wide range
of practical parameters into a small set of well-tested queries
A Prescription begins and ends before contraindication
B Treatment continues into contraindication
C Treatment and diagnosis begin together
D Treatment is contraindicated when it commenced
Classification (ongoing chronic condition)
Investigation Period 1 Investigation Period 2
Four temporal relationships (A-D) of a treatment and a problem classification that contraindicates it

15. Uses
• ChronoMedIt analysis of PMS data provides a
basis for
– Research cohort identification (how do low and
high MPR groups differ?)
– Intervention cohort identification (follow up to
raise MPR)
– Tracking of progress over time, and variation
between sites
– Interactive decision support
– Critique of criteria per se

16. Some ChronoMedit findings
• For 646 patients prescribed at least one of
simvastatin, metoprolol
succinate, bendrofluazide, felodipine, cilazapril and metformin
in a 15-month period, 50% had high adherence MPR
(Medication Possession Ratio) ≥80% to all (out of those 6) that
they were prescribed
– High adherence to individual medications was from 68%
(felodopine) to 55% (metformin)
• For patients prescribed ACEi or ABR with Dx of hypertension
and diabetes, non-adherent patients (MPR <80% or lapse >30
days in 12 months) are three times more likely to have
uncontrolled BP (odds ratio = 3.055; p = 0.012).
Mabotuwana T, Warren J, Harrison J, Kenealy T. What can primary care prescribing data tell us about individual
adherence to long-term medication?-comparison to pharmacy dispensing data. Pharmacoepidemiol Drug
Saf 2009;18(10):956-64.
Mabotuwana T, Warren J, Kennelly J. A computational framework to identify patients with poor adherence to
blood pressure lowering medication. Int J Med Inform 2009;78(11):745-56

17. The Lack of attention to medication
adherence (sometimes termed ‘compliance’)
• From a study of therapy intensification and
adherence of mid-Western VA patients
“Patients’ prior medication adherence had little impact on
providers’ decisions about intensifying medications, even at
very high levels of poor adherence…. suggests that providers are
simply not taking patients’ prior medication adherence into
account in making medication management decisions.”
Heisler M, Hogan MM, Hofer TP, Schmittdiel JA, Pladevall M, Kerr EA. When more is not better: treatment
intensification among hypertensive patients with poor medication adherence. Circulation
2008;117(22):2884-92.

18. Understanding adherence in the
Pacific population
• 20 Samoan patients (10 high MPR, 10 low)
– Lower adherence: ‘lack of transport’, ‘family
commitments’, ‘forgetfulness’, ‘church activities’,
‘feeling well’ and ‘priorities’
– High adherence: ‘prioritising health’, ‘previous event’,
‘time management’, ‘supportive family members’ and
‘relationship with GP (language and trust)’
– Common to both: ‘coping with the stress of multiple
co-morbidities’
Chang Wai K, Elley CR, Nosa V, Kennelly J, Mabotuwana T, Warren J. Perspectives on adherence to blood
pressure lowering medications among Samoan patients: qualitative interviews. Journal of Primary Health
Care 2010;2(3):217-224

19. Intervening – AIM-HI
• Adherence Innovation in Medication use for Health
Improvement
– Worked with West Fono Health Care
– Used ChronoMedIt to define a register of some 200 patients with
anithypertensive MPR<80% (for a 6 month period)
– Two nurses undertook Chronic Disease Management (CDM) on
these patients
– Significantly improved MPR for the intervention year as compared to
similar (low MPR) patients in a control Pacific-led practice
• Marginally significant improvement on systolic BP as measured ambiently at the
practice

22. Onward with medication adherence
• Adherence promotion deserves more attention
– By reminder
• Packaging, alerts/reminders, invoking ‘whānau’ (family)
– By mobile phone
• Assess and modify the belief model underpinning non-
adherence
• Continue to improve the epidemiology
– Planning study with larger cohorts (45andUp Study in
New South Wales; Auckland regional TestSafe) to
better assess statistical impact of MPR

23. Other users for ChronoMedIt?
• It applies to other long-term medications, too
– E.g. repeat short-term users of anti-depressants
• But who operates in a healthcare setting
where they really want clever tools to find
more work for them?!
– A truly rational healthcare system would be
seeking this information
– Know of any?

24. NIHI Evaluation
• National Institute for Health Innovation
– Based at School of Population Health
– Dedicated to innovative use of health IT to deliver
better, and more equitable, health outcomes
• Engaged in evaluation of the benefits, and areas
for improvement, around innovative use of IT in
the NZ health sector
– Electronic referrals
– Home telemonitoring
– General practice PMS functionality
– Shared care planning

25. NIHI commission for eReferral evaluation
• Conducted from Aug 2010 to Jun 2011
• Evaluate electronic referral (eReferral)
implementations
• Hutt Valley
• Northland
• Canterbury
• Auckland Metro region (entering pilot operation at time of
study)
• Stakeholder feedback and analysis of IT system
records

26. The Hutt Valley Solution
• Implemented in 2007
• 30 general practices referring in electronically from
Medtech32
• 28 services at Hutt Hospital receiving eReferrals
into Concerto
• 16 service-specific forms, 12 services using a generic
form
• Electronic management of workflow
• GP notified of receipt, triage/decline and FSA
• Hospital can see list of referrals w/ triage pending
• Any authorised user can see content on Concerto (e.g.
from ED)

27. Hutt Valley Uptake
GP referral volume by year (iSoft, Concerto)
• 1000 eReferrals per month in 2008; 1200/mo in 2010
• 56% of total referrals electronic by 2010
• 71% electronic from practices that sent at least one eReferral

28. Hutt Valley Triage Latency
1
0.9
0.8
0.7
Cumulative frequency
0.6
p07
0.5 e08
p08
0.4
e09
0.3 p09
0.2
0.1
0
0 5 10 15 20 25 30
Days
Time from GP letter date to date priority assigned at Hutt Hospital

29. Hutt Valley Upsides and Downsides
• Upsides
• Greater transparency
• Faster turnaround
• Downsides
• IT done once and left
• Some persistent
usability issues
• Slow attachment
opening
• Difficulty attaching
photos
• Never revisited form
content

30. The rest of the eReferrals evaluation…
• Well, actually there’s a talk on that in the main conference!
• One other data mining point, though…
– For Canterbury, we did a content analysis of decline messages
sent back to GPs
– Developed thematic clusters with frequencies and examples
– This is a powerful kind of analysis often possible with certain
sources of uncoded EHRs
• We’re doing more of this with Shared Care Planning
• Can do a random sample if you lack the resources to theme your
whole collection
• Characterising author role/qualifications, message length and
message variety are other aspects to this approach

31. NIHI’s Evaluation Framework for
Innovative Health IT
• Establish a relevant benefits framework
– What’s this technology supposed to be good for?
• Always use the transactional EMR data
– Uptake, cycle times
• Always ask the users and other stakeholders
– Action research, grounded theory, theoretical sampling
• Disseminate broadly and often
• After ~12 evaluation projects in NZ, would like to work
with partners overseas

32. So what have we learned?
• The data is there
– It can reveal opportunities for process improvement
– It can confirm benefits of recent innovations
– It can be used integrally with on-going innovations
• E.g. to track user uptake and types of uses
• It doesn’t necessarily do what we ‘want’
– It can be quite haphazard exactly what the ambient data
can/cannot usefully describe
• With general practice we were led to MPR and adherence
• With eReferrals we couldn’t completely comment on referral letter
quality (decline messages give a hint, but quite confounded)
– Carefully track the limitations of your data sources
• E.g. what is systematically absent

33. Planning for data mining
• Should we try to engineer the situation in advance so
that we have the right data?
– Sure, but don’t ruin the usability of the system by adding
user data entry just for evaluation
• Certainly it helps to plan what you’d want to know for
evaluation as you do the implementation
– Some fields, esp. ‘meta data’ are cost free for end users
• E.g. keeping a good audit log (which you should anyway!)
– openEHR might be a good answer
• Design a good ‘archetype’ for each data element, with good tools
and process to back up this data definition (talk to Dr Koray Atalag)

34. Go forth and analyse!
• So go mine your EHRs!
– Or hire NIHI to do it for you
• Postscript on research ethics approval
– Get it – so you can publish to journals and let
everybody know the findings
– Not too hard for this kind of work
• Don’t need names or addresses, encrypt the NHIs, avoid
getting precise dates of birth

35. Questions / further info
• Jim Warren, Professor of Health Informatics
– jim@cs.auckland.ac.nz
– Also, try PubMed on ‘Mabotuwana’
Questions ???!
The National Institute
for Health Innovation