Statistical opportunities and challenges

1
Statistical opportunities and challenges of
electronic health records Dr Alex Dregan Lecturer
in Epidemiology and Public Healthemail
alexandru.dregan_at_kcl.ac.uk
2
Electronic health records (EHRs)

• Capture and integrate data on all aspects of care
  over time
• various data types, from structured information
  such as condition diagnosis, lab tests,
  referrals, drug prescription data, to
  unstructured data such as clinical narratives
• Growing volume of data
• Prescribing, blood pressure, morbidity data is
  accurate and complete
• Data can be related to individual patients'
  characteristics (sex, age, social class) and
  practice aspects (ie practice size, region,
  number of GPs, auxiliary staff)
• Widespread use in UK primary care
• For effective communication, clinical care,
  service organisation, quality and audit,
  professional development and self-directed
  learning
• Potential to supporting undergraduate learning
  and teaching as it reflects the context in which
  the students will be ultimately working
• Graduates must be able to use different
  techniques to record, organise, analyse, and
  present information

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
3
Using EHRs for research example
Antibiotic prescribing for acute RTIs
CPRD practices
NICE guidelines Qualitative research
Control
Intervention
Electronic reminder to GP, no or delayed AB
prescribing
Subjects aged 18-59 years consulting for RTI, 60
prescribed AB in 2006
Proportion of consultations for RTI with AB
prescribed
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
4
Using EHRs for research example Analysis

• Sample size (Hayes and Bennett (1999)) 47
  practice per arm for an 0.8 power to detect 5
  difference (ICC0.23 Ashworth et al., 2005)
• Statistical analysis
• - Intention to treat (ITT) principle -
  difference in outcome between intervention and
  control practices
• - Cluster-level (practice as a unit)
  analysis analysis of covariance framework
• - Minimum variance weights (Kerry and
  Bland, 2001) used to allow for varying number of
  participants and consultations per practice

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
5
Prompt Utilisation
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
6
RTI consultation and AB prescribing per 1,000
registered participants and proportion () of RTI
consultations with AB prescribed. Figures are
mean (interquartile range) of practice-specific
values for 12 months before- and after-
intervention.
  Intervention Trial Arm Intervention Trial Arm Control Trial Arm Control Trial Arm Adjusted mean differenced (95 confidence interval)   P value
  Before     After Before After   Adjusted mean differenced (95 confidence interval)   P value
RTI Consultation rate 219 (181 254) 209 (176247) 216 (186 246) 218 (184244) -9.10 (-21.513.30)   0.148
             
Antibiotic Prescription rate 116 (91 131) 108 (87129) 111 (86 135) 114 (85 128) -9.69 (-18.63 -0.75)   0.034
             
AB Prescriptions Per RTI Consultation () 53 (46 60) 52 (45 58) 52 (4560) 52 (4559) -1.85 (-3.59-0.10)   0.038
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
7
Intervention utilisation and AB prescribing by
quartile of intervention utilisation.
  Control practices Lowest Quartile of Utilisation (13) Second Quartile (13) Third Quartile (13) Highest Quartile of Utilisation (13)
Intervention Utilisation (per 1,000 consultations for RTI) Intervention Utilisation (per 1,000 consultations for RTI) Intervention Utilisation (per 1,000 consultations for RTI)      
Prompt Views Not applicable 0 (0 0) 16 (0 22) 77 (0 117) 174 (68 248)
Leaflets Printed Not applicable 0 (0 0) 6 (0 0) 18 (0 21) 15 (0 0)
Proportion () of RTI consultations with antibiotics prescribed Proportion () of RTI consultations with antibiotics prescribed Proportion () of RTI consultations with antibiotics prescribed      
Before Intervention 52 (45 59) 55 (49 61) 53 (46 59) 55 (51 63) 50 (41 57)
After Intervention 52 (45 59) 54 (46 63) 54 (51 60) 53 (5261) 48 (42 54)
Unadjusted mean difference (95 confidence interval) 0.7 (-0.6 2.0) -1.2 (-5.1 2.8) -1.0 (-2.9 0.9) -1.4 (-3.91.0) -1.6 (-5.0 1.7)
Adjusted test for trend across categories (95 confidence interval) Adjusted test for trend across categories (95 confidence interval) -0.64 (-1.23 -0.05) P0.034 -0.64 (-1.23 -0.05) P0.034 -0.64 (-1.23 -0.05) P0.034 -0.64 (-1.23 -0.05) P0.034
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
8
Using EHRs for research - example
(cont.)Inflammatory disorders and risk of
cardiovascular diseases

• Outcomes
• New diagnoses of stroke, CHD, and T2DM.
• Multiple morbidity was defined as the occurrence
  of 2 outcomes in a participant.
• Mean of CRP values (biomarker)
• Exposure
• Chronic inflammatory disorders including
  psoriasis, Crohns disease, Bullous skin disease,
  ulcerative colitis, systemic lupus, inflammatory
  arthritis, and vasculitis
• Statistical analysis
• Cox proportional hazards model
• Sensitivity analyses using competing risk
  analysis
• Missing indicator variables to deal with missing
  data
• Random-effects meta-analysis

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
9
Forest plot displaying random effect
meta-analysis of the influence of diverse chronic
inflammatory conditions on multiple
cardiovascular. HRHazard ratios CIConfidence
intervals.
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
10
Using EHRs for research - example
(cont.)Sodium Valproate and risk of stroke - A
nested case-control study

• A nested case control study was implemented using
  data from the Clinical Practice Research Datalink
  (CPRD) (www.cprd.com).
• The study population consisted of a cohort of
  epilepsy (N15,001) patients treated with at
  least one AED who were registered with 653 CPRD
  practices between 1 January 1992 and 31 January
  2013.
• Exposure Sodium valproate treatment represented
  the primary exposure of interest for the present
  study.
• Outcomes Ischemic stroke
• Analysis Conditional logistic regression
  analysis

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
11
Using EHRs for research - example (cont.)
Case (N2,002) Case (N2,002) Control (N13,098) Unadjusted model OR (95CI) p Fully adjusted model OR (95CI) p
Ever prescribed 681(34) 681(34) 4,407(34) 1.03(0.93,1.14) 0.555 1.01(0.91,1.12) 0.875
Pre-stroke year 555(28) 555(28) 3,106(24) 1.27(1.14,1.41) 0.001 1.22(1.09,1.38) 0.001
Number of SV prescriptions Number of SV prescriptions Number of SV prescriptions
None None 1,321(66) 8691(66) Reference Reference
Lowest quarter Lowest quarter 227(11) 1,075(8) 1.47(1.26,1.72) 0.001 1.22(1.02,1.45) 0.025
Second quarter Second quarter 198(10) 1,062(8) 1.28(1.09,1.59) 0.003 1.21(1.02,1.45) 0.033
Third quarter Third quarter 166(8) 1,100(9) 0.99(0.83,1.18) 0.924 1.00(0.83,1.21) 0.972
Highest quarter Highest quarter 90(5) 1,170(9) 0.49(0.39,0.61) lt0.001 0.59(0.46,0.74) lt0.001
Time on SV prescriptions Time on SV prescriptions Time on SV prescriptions
None None 1,321(66) 8,915(68) Reference Reference
Lowest quarter Lowest quarter 256(13) 962(7) 1.97(1.68,2.29) lt0.001 1.62(1.37,1.92) lt0.001
Second quarter Second quarter 194(10) 1,023(8) 1.35(1.14,1.60) 0.001 1.28(1.07,1.54) 0.007
Third quarter Third quarter 146(7) 1,068(8) 0.92(0.76,1.11) 0.373 0.95(0.78,1.15) 0.584
Highest quarter Highest quarter 85(4) 1,130(9) 0.48(0.38,0.60) 0.001 0.57(0.44,0.72) lt0.001
12
Using EHRs for research - example
(cont.)Validity of cancer diagnosis in a primary
care database compared with linkedcancer
registrations in England. Population-based cohort
study

• Population-based cohort study
• The eligible cohort comprised 42,556
  participants, registered with English general
  practices in the CPRD that consented to CR
  linkage.
• Read and ICD cancer code sets were reviewed and
  agreed by two authors
• The positive predictive value (PPV), sensitivity,
  and specificity were estimated using CR as the
  reference data. Median and interquartile ranges
  for the difference in date of cancer diagnosis
  between CPRD and CR databases were estimated for
  four cancer groups. Because the available CR data
  included only month and year of cancer diagnosis,
  a day of diagnosis for each CR case was imputed.

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
13
Using EHRs for research - example (cont.)
  CPRD Cancer registry Cancer registry CPRD total Grand total PPV Sensitivity Specificity
    Recorded Not recorded          
Colorectal Recorded 1732 43 1775   0.98 0.92 0.99
  Not recorded 150 40631          
  CR total 1882     42556      
                 
Lung Recorded 1659 65 1724   0.96 0.94 0.99
  Not recorded 104 40626          
  CR total 1763     42556      
                 
Oesophageal Recorded 872 27 899   0.97 0.92 0.99
  Not recorded 74 41583          
  CR total 946     42556      
                 
Urological Recorded 953 78 1031   0.92 0.85 0.99
  Not recorded 166 41359          
  CR total 1119     42556      
                 
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
14
Using EHRs for research - opportunities

• Mining of electronic health records (EHRs) -
  establishing new patient-stratification
  principles and for revealing unknown disease
  correlations
• Identify persons at very high (e.g. gt99th
  percentile, risk scores) risk for a given
  condition
• Identify novel risk/protective factors for
  disease onset and progression
• Integrating EHR data with registry data
• Link primary care data with genetic data (UK
  Biobank and CPRD linkage)
• Link primary care with registry data (CPRD with
  National Cancer Registry linkage)
• Developing predictive models for
• Therapeutic interventions effectiveness and
  safety (pharmacovigilance) use of propensity
  score matching to adjust for confounding
• Decision support systems
• Synthesize large amounts of information to
  provide alerts related to adverse events, patient
  safety, treatment course

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
15
Using EHRs for research Challenges - data

• Concepts
• Probability, randomness, variability, statistical
  errors, central limit theorem
• Data
• Reporting ie CONSORT, TREND, STROBE
• Accessing and visualizing - ie manipulation,
  graphical representation
• Interpretation - clinical vs statistical
  significance, effect size
• Sources of bias
• Incomplete data - EHR data are captured at the
  point of care by GPs, patients who do not
  regularly interact with the health system may
  have incomplete data
• Sampling bias, protopathic bias, measurement
  error, residual bias, confounding by indication
• Prediction models
• Uncovering patterns in patient trajectories
  through disease and intervention nodes (ie
  medication) in a clinical context is
  statistically and computationally challenging
• Inferential methods for clustered, matched,
  paired, or longitudinal studies
• Multiple testing common in EHRs research

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
16
Using EHRs for research Challenges students
teachers

• Motivation and interests
• Most students will be clinicians not researchers
  - focus on design, choice of analytical methods,
  and interpretation of findings?
• Statistics is not seen as a core subject for
  medical training use real-world examples, uses
  and abuses of statistics
• Aptitudes
• Differences in prior exposure to statistics
  group work
• Learning disabilities greater use of technology
• Differences in teaching abilites
• Assessment
• Formative vs summative assessment
• Use of quizzes at the end of each
  lecture/tutorial? Peer assessment?
• Teaching methods
• Online modules/ youtube type learning
• Lab-based teaching/use of personal computers
  during tutorial
• Staff shortage training the teacher?

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
17
Using EHRs for research Challenges

• EHRs are becoming common and viewed as a
  potential tool for healthcare quality assessment,
  clinical trials and health outcomes research
• Visualizing the data in clinical settings is a
  challenge, much less applying standard
  statistical methodology (standard errors and
  p-values) that may have little or no meaning in
  very large sample sizes
• Where biostatistics will fit in future education?
  
• Biostatistics is often viewed as a separate
  entity, and much of it is not directly
  statistical in nature, as the issue of how to
  process such large datasets is a dominating
  consideration
• Public health also requires the analysis of large
  databases, both specifically and in relation to
  issues affecting the ongoing restructuring of the
  NHS and is also a key area of potential research
  for biostatistics
• Integrate statistics teaching within the context
  of epidemiological analysis, medical-decision
  making, computing, and policy development

Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre
18
Thank youQuestions?
Guys and St Thomas NHS Foundation Trust and
Kings College Londons comprehensive Biomedical
Research Centre