Transcript Burman FDA_Industry final

A Quantitative Approach to
Clinical Development
Carl-Fredrik Burman, PhD
Statistical Science Director
AstraZeneca R&D, Sweden
A new paradigm (?)
How should we get there?
Alternative designs
(adaptive, cross-over, “traditional”)
To where
do we want
to go?
Where
are we?
Modeling
Decision Analysis (DA)
to optimize design,
based on model
& preferences
Simulations
Preferences
Study design
decisions
How statisticians used to design trials
— A caricature
Medic (M): “What sample size do we need?”
Statistician (S): “Could you tell me the least clinically
relevant effect, D, please?”
M: It’s 20.
S: “… and the standard deviation?”
M: “It was 100 in the last trial”
S: “Then it’s simple. N=1053 gives 90% power.”
M: “Oh, we cannot afford that. Say that D=30 instead.
S: “Then the required sample size is 469.
M: Excellent
The medics have taken care of
population, duration, variable, etc.
How should we get there?
Alternative designs
(adaptive, cross-over, “traditional”)
To where
do we want
to go?
Where
are we?
Modeling
Decision Analysis (DA)
to optimize design,
based on model
& preferences
Simulations
Preferences
Example of a
study design
decision
Thanks to Claes Ekman & Björn Bältsjö
Background
• Loosely based on experiences from
• AZD7009 project (atrial fibrillation)
• Compound in early phase II
• Potential side effect X
• New results for stopped competitor drug, say.
• Competitor drug-induced AE rate about 10%
• Placebo rate likely to be about 1%
• Minor AEs, no ethical complications
• Should a specific safety trial be added before
entering next phase?
AE probabilities
• q = P( AE | placebo )
• p = Drug-induced rate of X
• p>0 will hit sales
• no approval if p>5%
• P( AE | drug ) = 1–(1-p)(1-q) = q+p(1-q)  q+p
Will trial results be interpretable?
• “Standard” design
• n=30 subjects get active treatment
• m=30 receive placebo
• Say that the number of AEs found are
• x=2 on active treatment
• y=0 on placebo
• Far from statistically significant
Single-arm trial
• Historical data exist for placebo group
• Alternative trial with n=60, m=0
Formulation of priors
• Prior for drug-induced AE probability
• P(p=0.00) = 0.6
Excellent
• P(p=0.03) = 0.3
2nd line treatment
• P(p=0.10) = 0.1
Not a viable treatment
• Prior for placebo AE probability
• P(q=0.01) = 0.9
• P(q=0.05) = 0.1
• Independence in prior distribution
• NB! Model is too simplistic for practical use,
but may have pedagogical value
Prior distribution
100%
p=0.10
80%
p=0.03
60%
40%
p=0.00
20%
0%
Single-arm safety trial
n=60 pat’s; x=3 AEs
Posterior = Prior + Data
Prior distribution
100%
p=0.10
80%
p=0.03
60%
40%
p=0.00
20%
0%
Posterior if x=3, n=60
100%
p=0.10
80%
60%
p=0.03
40%
20%
0%
p=0.00
1
100%
Before trial / Prior
p=0.10
p=0.03
80%
60%
40%
p=0.00
20%
0%
After n=60 patients
100%
p=0.10
80%
p=0.03
60%
p=0.00
40%
20%
0%
x=0
x=1 x=2 etc
100%
Before trial / Prior
100%
80%
80%
60%
60%
40%
40%
20%
20%
0%
0%
After n=20 patients
x=0
After n=60 patients
Ideal (infinite info)
100%
100%
80%
80%
60%
60%
40%
40%
20%
20%
0%
0%
x=0
x=1 x=2 etc
etc
Economic assumptions
• (Expected Net Present) Value V(p) before
dose-finding:
• V(p=0.00) = 1000
• V(p=0.03) = 100
• V(p=0.10) =
0
• Planned dose-finding trial cost K = 500
Total value of
suggested safety trial (n=60)
x
0
1
2
3
4
…
•
•
•
Probability
32.2%
24.9%
14.4%
9.2%
6.1%
…
E[Value] = …
Project value
433
280
16
-169
-243
…
E[ Value | Data ]= E[ E[ Value | Data ] ]= 130
Terminate project if value<0
NB! The trial is useful only if it separates positive and
negative values.
After n=60 patients
After n=20 patients
100%
100%
80%
80%
60%
60%
40%
40%
20%
20%
0%
0%
x=0
Value
x=1 x=2 etc
After n=60 patients
x=0
Value
600
600
400
400
200
200
0
0
-200
-200
-400
-400
-600
-600
etc
After n=20 patients
How to choose n and m?
• Add cost of safety trial
• Maximizing E[Value] over all possible n’s, m’s
• Do we need a placebo group?
• Adaptive design of safety trial
• allocation fraction to placebo group may
depend on data
• Adaptive design of next phase
• checking for AE X during study
Dose-response
example
A new drug
• has pros and cons
• … and some uncertainty in the assessment
thereof
• It is important to study each dimension
(efficacy, different types of safety issues)
separately
• But a combined analysis may also be useful
• May this help sponsor-regulator
communication?
Rate /
Loss fcn
0,2
0,15
Weighted
net loss
0,1
0,05
Net loss
0
0
1
2
3
Exposure
Inspired by Marie Cullberg’s PhD thesis
4
5
AE
Lack of
effect
Don’t trust your DA blindly!
• Check robustness
• Question the assumptions
• Let the decision-makers, not the DA model,
determine the final decision
• DA helps decision-makers
• by structuring the problem
• exploring logical consequences of
assumptions
• facilitate communication