1. Can CER and Personalized
Medicine Work Together?
Zhaohui (John) Cai, MD, PhD
Biomedical Informatics Director, AstraZeneca
Medical Informatics World Conference
Boston, MA
April 9, 2013
– Where Generalization Meets
Personalization

2. Disclaimer
This presentation mainly represents my
personal views of how CER and PM should
work together and how Biomedical Informatics
could help achieve that. It does not constitute
any positions of AstraZeneca or any other
organizations

3. Presentation Outline
•Introduction: CER, generalizability of CER vs.
RCT, CER and Personalized Medicine (PM)
• An external example of CER study for PM
• Analytical approaches: subgroup analysis and
predictive learning
• Internal examples of predictive learning
• Proposed personalized CER in drug
development and in clinical practice
Author | 00 Month Year3 Set area descriptor | Sub level 1

4. CER
The definition of CER proposed by the
Congressional Budget Office:
“An analysis of comparative effectiveness is
simply a rigorous evaluation of the impact of
different treatment options that are available
for treating a given medical condition for a
particular set of patients.”
4
For more definitions of CER, see the IOM report on Initial National Priorities for Comparative Effectiveness
Research

5. RCT, CER, and Generalizability
• RCTs, by their nature, limit the
applicability of their findings to
homogeneous populations
because of inclusion and
exclusion criteria and the tightly
controlled way they are generally
conducted
• CER studies can involve very
few inclusion or exclusion criteria
in order to make the results as
generalizable as possible to
‘real-world’ patients with various
co-morbidities, also can be
conducted in naturalistic settings
5
Real World
population
CER
population
RCT
population

6. CER and Personalized Medicine
• A tension between Personalized Medicine (PM) and CER can be created
when pressure is placed on CER to conform to the prevailing RCT model
• Concerns have been raised that CER will not take into consideration
individual patient differences and may impede the development and
adoption of PM
• CER studies can include a wide range of patient populations common to all
healthcare provider environments
• Taking advantage of a variety of epidemiological and informatics research
methods can help non-randomized CER studies address PM concepts
6

7. CYP2C9 and VKORC1
genotyping
Quasi-
experimental
design
Medco-Mayo Warfarin Effectiveness Study (MM-
WES)
7
Epstein et al, 2010

8. ATE and HTE
• Randomized controlled trials (RCTs)
usually report an average treatment
effect (ATE), which is critically for
regularity approval
• Heterogeneity of treatment effect
(HTE) is defined as nonrandom
variability in the direction or
magnitude of a treatment effect, in
which the effect is measured using
clinical outcomes
• Understanding HTE is critical for
decisions that are based on knowing
how well a treatment is likely to work
for an individual or group of similar
individuals, and is relevant to
stakeholders including patients,
clinicians, and policymakers
8
Kravitz, Duan, Braslow 2004
As defined by AHRQ User Guide to Observational CER, 2013

9. HTE Analysis for PM
• HTE implies applicability of findings from RCT or observational
CER to individual patients (i.e. PM)
• HTE analyses
- Subgroup analysis to estimate treatment effects in clinically relevant
subgroups , one variable at a time, usually a baseline or pretreatment
variable (e.g. genetic variants or gender)
• Exploratory subgroup analysis
• Confirmatory subgroup analysis
- Predictive learning to predict whether an individual might benefit from a
treatment
• Can take a multivariate machine learning approach
• Can be a pre-step for subgroup analysis
• Applicable to both healthcare setting and drug development
9

10. Predictive Learning: Identify Responders Early
in Treatment Course
Subgroup 1
(predicted non-
responders at
baseline)
Subgroup 1
(predicted
responders)
Treatment period I Treatment Period 2
Prediction Prediction
ContinueBaseline Outcome
Measure
Prediction algorithm based
on biomarker(s) and/or
simply clinical disease
activity score(s)
Prediction
Drop and
alternative
treatment
Baseline
Subgroup 2
(predicted non-
responders at
early time points )
Treatment period I
Prediction Prediction
DiscontinueBaseline
10

11. Internal Example: Predictive Learning for PM Clinical
Decision Tool
• Question: Can we predict responders
early, and use the predictions in clinical
practice?
• Data & Method: model Phase II data
using ~30 clinical variables to identify an
early predictor of individual response at 6
months, using Random Forests models
• Result: A combination of 4 clinical
variables are predictive at month 1 to
identify responders at month 6 with close
to 80% accuracy
• Benefit: Clinical Decision Tool for patient
selection that may double response rate
identified, to be validated using phase III
and real world data (subgroup analysis)
Accuraciesofearlypredictions
Predicting month 6 endpoint 1
Predicting month 6 endpoint 2
Time of Prediction
11

12. Current Personalized Medicine Strategy in Drug
Development: drug-test co-development
In vitro/vivo studies
Data/literature mining
Candidate biomarker(s)
(predictive learning)
Validated biomarker(s)
(subgroup analysis)
Marker based design
(subgroup analysis)
Hypothesis &
initial modeling
Phase 2b
Design and analysis
Phase 3
Design and analysis
Outcome (a PM product)
Which patients will benefit most from the therapy?
Explore
Confirm
Preclinical/
Phases 1 & 2a
12

13. Proposed Personalized Comparative Effectiveness
Research (PCER) in Drug Development
RealWorldData
In vitro/vivo studies
Data/literature mining
Candidate biomarker(s)
(predictive learning)
Validated biomarker(s)
(subgroup analysis)
Marker based design
(subgroup analysis)
Hypothesis &
initial modeling
Phase 2b
Design and analysis
Phase 3
Design and analysis
Outcome (a PM product)
Who will benefit most from treatment A (i.e. candidate drug) and
who will benefit most from treatment B (i.e. standard of care)?
Explore
Confirm
Observational CER
Preclinical/
Phases 1 & 2a
13
Clinical /payer decision
support
Predictive learning
Explore
Conform

14. New informatics challenge
Personalized Comparative Effectiveness
Research (PCER) for Healthcare Decisions: an
informatics challenge
14
A real patient
Retrospective real-world
database
Search for similar patients
A cohort of similar, previously
treated patients
Different outcomes from
different treatment pathways
Retrospective CER study
(subgroup analysis)
Predictive
learning
Drug A Drug B
Decision
point 2
Outcome
Decision
point 1
Drug B Drug A
Diagnosis
Personalized
treatment Current informatics challenge
Outcome

15. Drug-test co-
development
PCER
Achieving Personalized Healthcare in Real
World Using CER
Personalized
healthcare
Decision
support for
payers
Decision
support for
cliniciansGenomic
research
data
Integrated
healthcare
data
Healthcare cost
(insurance claims)
data
15
• CER and PM can and have to work together
• “Drug A is better than drug B for disease X” kind of
general comparative effectiveness may not be
applicable to individual patient care
• PCER will answer “to what patient subgroup, what
disease stage, what treatment pathway, and where in
the treatment pathway, a comparative effectiveness
evidence is applicable”
Clinical (RCT,
EHR, PHR/PRO,
Registry) data

16. 16
Confidentiality Notice
This file is private and may contain confidential and proprietary information. If you have received this file in error, please notify us and
remove it from your system and note that you must not copy, distribute or take any action in reliance on it. Any unauthorized use or
disclosure of the contents of this file is not permitted and may be unlawful. AstraZeneca PLC, 2 Kingdom Street, London, W2 6BD, UK,
T: +44(0)20 7604 8000, F: +44 (0)20 7604 8151, www.astrazeneca.com
Thank you
&
Questions?