Date post: | 22-Dec-2015 |
Category: |
Documents |
Upload: | piers-chase |
View: | 214 times |
Download: | 0 times |
1
Statistical Designs for Stratified Medicine
Cindy BillinghamProfessor of Biostatistics, School of Cancer Sciences
Director of Statistics, Cancer Research UK Clinical Trials Unit
Lung Cancer Research Stratified Medicine Educational EventBirmingham, June 22nd 2015
2
Agenda
Early phase clinical trials for stratified medicine Illustrated using National Lung Matrix Trial Bayesian adaptive design Example of an ‘umbrella trial’
Later phase clinical trials for stratified medicine Typical randomised controlled designs (RCTs) Umbrella and basket trials Do we always need an RCT to change clinical practice in
stratified medicine?
3
Actionable targets (biomarkers) and targeted drugs = 17 drug-biomarker cohorts
Prev A:AZD4547
B:AZD2014
C:Palbociclib
D:Crizotinib
E:Sel+Doc
F:AZD5363
G:AZD9291
A1: FGFR2/3 mutation-NSCLC 4.0%
B1: TSC1/2 mutation-NSCLC 2.7%
B2: LKB1 TIER1 mutation-NSCLC 6.4%
C1: Proficient Rb & p16 loss-SCC 29.0%
C2: Proficient Rb & CDK4 amp-NSCLC 7.0%
C3: Proficient Rb & CCND1 amp-NSCLC
7.3%
C4: Proficient Rb & KRAS mutation-ADC 25.8%
D1: Met amplified-NSCLC 2.3%
D2: ROS1 gene fusion-NSCLC 1.7%
E1: NF1 mutation-SCC 5.8%
E2: NF1 mutation-ADC 4.6%
E3: NRAS mutation-ADC 1.0%
F1: PIK3CA mutation-SCC 11.0%
F2: PIK3CA amp-SCC 15.0%
F3: PIK3/AKT deregulation-NSCLC 4.5%
F4: PTEN loss & mutation-SCC 20.0%
G1: EGFR mutation & T790M+-NSCLC 8.0%
4
National Lung Matrix Trial SchemaFinal biopsy result (diagnostic and/or repeat) mandated for National Lung Matrix Trial entry:
Biopsy failure
(diagnostic and repeat)or ineligible
No actionable genetic
aberration
Multiple actionable target with
open treatment arms
Single actionable target with
open treatment arm
Outcome Measures (common set for all arms with treatment-arm-specific primary): Best objective response rate (ORR), Change in total target lesion size (PCSD), Progression-free survival time (PFS), Time-to-
progression (TTP), Overall survival time (OS), Toxicity
Allocated to single
treatment arm relevant to actionable
target prioritised by CI if eligible
Allocated to single
treatment arm relevant to actionable
target if eligible
Standard treatment
OR recruitment to another relevant
trial
Allocated to no actionable genetic change (NA) cohort if
eligible
Standard Clinical
Outcome Measures
Actionable target but no
target therapy
arms open
5
Typical Single Arm Phase II Trial
Eligible Patients
NEW Treatment
Response rate
Historical data / clinical
experience of standard treatment
Benchmark response rate
0% 100%
?
Common Designs:A’Herns single stageSimon’s two-stage
6
Flexible design that embraces study complexity and facilitates decision-makingNeed efficient and flexible design that: has potential to stop early for lack of efficacy allows for differing prevalence rates of biomarkers allows continued recruitment to any sample size as appropriate has potential to incorporate relevant information from other biomarker
cohorts within each drug arm (‘borrowing information’) has potential to incorporate pre-existing evidence and emerging
external evidence enables cumulative learning
Question that we really want to answer: What is the probability that the TRUE signal of efficacy is above x% Make Go-NoGo decision for further research based on probability
Bayesian Adaptive DesignRef: Berry, Carlin, Lee, Muller; Bayesian Adaptive Methods for Clinical Trials, Chapman and Hall /CRC Biostatistics Series 2011
7
What is a Bayesian Approach to Statistical Analysis? Alternative method of statistical analysis to the
classical / frequentist approach ‘The explicit quantitative use of external evidence in the design,
monitoring, analysis, interpretation of a health-care evaluation’ Spiegelhalter et al 2004
Based on theorem devised by Reverend Thomas Bayes (1702-1761)
Basic maths:
Prior x Data → Posterior
)(
)()/()/(
Ap
BpBApABp
Bayes Theorem:
Posterior probability distribution
P(HR<1)=0.75
8
Start Recruitment
Interim Analysis 1 N=15
if P(<30%) ≥ 0.9STOP early @ Interim
Final AnalysisN=30
STOP @ Final
if P(>30%) < 0.5
if P(>30%) ≥ 0.5
GO
Arms A, B, D, F & GBayesian Adaptive Two-Stage Design
if P(<30%) < 0.9
Cohorts: A1, B1, B2, B3, D1, D2Primary outcome measure: objective response rate (ORR): true ORR
9
Illustrating Statistical Analysis Plan Example: GO-GO
Prior: Beta(1,1)
Interim Posterior:Beta(4,13)
Interim analysis: 3/15 = 20%Final analysis: 13/30 = 43%
P(<30%)=0.75
Final Posterior:
Beta(14,18)
P(>30%)=0.95
Beta-Binomial conjugate analysisPrior: ~ Beta(a0, b0)Posterior: |r,n ~ Beta(a0+r,b0+n-r)
10
Cumulative Learning Using Bayesian Analysis
n=1,r=1 n=2,r=1 n=3,r=1 n=4,r=2 n=5,r=2
n=6,r=2 n=7,r=2 n=8,r=2 n=9,r=2 n=10,r=2
n=11,r=2 n=12,r=2 n=13,r=2 n=14,r=2 n=15,r=3
n=16,r=4 n=17,r=5 n=18,r=5 n=19,r=6 n=20,r=7
n=21,r=7 n=22,r=7 n=23,r=8 n=24,r=8 n=25,r=8
n=26,r=9 n=27,r=10 n=28,r=11 n=29,r=12 n=30,r=13
11
How well does this design work? Are other designs better?Desirable operating characteristics
Sample size criteria: Need minimum of 10 and maximum of 15 at interim Need minimum of 20 and maximum of 40 at final
Interim analysis criteria: Need p(STOP early|=10%)>80% (true stopping rate) Need p(STOP early|=40%)<5% (false stopping rate)
Final analysis criteria: Need p(GO at final|=20%)<10% (false positive rate) Need p(GO at final|=40%)>80% (true positive rate)
Of all the designs that satisfy these criteria, the optimal design is that which maximises the true positive rate
12
Start Recruitment
Interim Analysis 1 N=15
if P(<30%) ≥ 0.9STOP early @ Interim
Final AnalysisN=30
STOP @ Final
if P(>30%) < 0.5
if P(>30%) ≥ 0.5
GO
Arms A, B, D, F & GDesign and Operating Characteristics
2/15=13% or less
8/30=27% or less
3/15=20% or more
9/30=30% or more
if P(<30%) < 0.9
=10% =20% =30% =40% =50%
STOP early 81.6% 39.8% 12.7% 2.7% 0.4%
STOP at final 18.2% 47.8% 31.8% 7.7% 0.7%
GO at final 0.2% 12.4% 55.5% 89.6% 99.0%
Operating characteristics based on exact binomial probabilities
: true ORR
13
Start Recruitment
Interim Analysis 1 N=15
if P(<40%) ≥ 0.9STOP early @ Interim
Final AnalysisN=30
STOP @ Final
if P(>40%) < 0.5
if P(>40%) ≥ 0.5
GO
Arm E (Selumetinib+Docetaxel)Design and Operating Characteristics
if P(<40%) < 0.9
=10% =20% =30% =40% =50%
STOP early 94.4% 64.8% 29.7% 9.1% 1.8%
STOP at final 5.6% 34.3% 54.7% 34.7% 8.7%
GO at final 0.0% 0.9% 15.6% 56.2% 89.5%
Operating characteristics based on exact binomial probabilities
Cohorts: E1, E2, E3 - objective response rate (ORR): true ORR
14
Start Recruitment
Interim Analysis 1 N=15
if P(<3mths) ≥ 0.8STOP early @ Interim
Final AnalysisN=30
STOP @ Final
if P(>3mths) < 0.5
if P(>3mths) ≥ 0.5
GO
Arm C (Palbociclib)Bayesian Adaptive Two-Stage Design
if P(<3mths) < 0.8
Cohorts: C1, C2, C3, C4Primary outcome measure: progression-free survival time (PFS): true median PFS (months)
15
Arm C (Palbociclib)Design and Operating Characteristics
True median = 1 month
True median = 2 months
True median = 3 months
True median = 4 months
True median = 5 months
True median = 6 months
STOP early 97.2% 49.9% 16.8% 5.4% 2.2% 1.0%
STOP at final
2.8% 49.0% 43.7% 11.4% 1.8% 0.3%
GO at final <0.1% 1.2% 39.4% 83.2% 96.0% 98.6%
Operating characteristics estimated through simulation
Cohort C1, recruiting at 93 patients per annum when all recruitment centres are open
16
Using Bayesian Hierarchical Modelling as Secondary Analysis
E1: NF1 mutant -
SCC
E2: NF1 mutant -
ADC
Arm E: Selumetinib+ Docetaxel
25/30 (83%)
2/15 (13%)
Ensures borderline decisions err on positive if drug has shown potential in other cohorts
Secondary analysis to aid decision-making,
particular when decisions are borderline
Can build in expected level of association
Never be used to negate a primary analysis that shows a potentially
positive result
External Evidence
17
National Lung Matrix Trial SchemaFinal biopsy result (diagnostic and/or repeat) mandated for National Lung Matrix Trial entry:
Biopsy failure
(diagnostic and repeat)or ineligible
No actionable genetic
aberration
Multiple actionable target with
open treatment arms
Single actionable target with
open treatment arm
Outcome Measures (common set for all arms with treatment-arm-specific primary): Best objective response rate (ORR), Change in total target lesion size (PCSD), Progression-free survival time (PFS), Time-to-
progression (TTP), Overall survival time (OS), Toxicity
Allocated to single
treatment arm relevant to actionable
target prioritised by CI if eligible
Allocated to single
treatment arm relevant to actionable
target if eligible
Standard treatment
OR recruitment to another relevant
trial
Allocated to no actionable genetic change (NA) cohort if
eligible
Standard Clinical
Outcome Measures
Actionable target but no
target therapy
arms open
18
No Actionable Genetic Change Cohorts (NA)
No actionable genetic
aberration
Drug NA1
Drug C
Drug NA2
Drug E
Drug B
Etc
Test in NEGATIVES once signal demonstrated in POSITIVES
ExamplePipeline of options that become available sequentially
Drug NA1: MEDI4736Statistical design: adaptive Bayesian design in line with actionable target cohortsTwo co-primary outcomes: ORR and PFS6 Initial sample size for interim analysis: N=20 (determined by drug supply)
No comparison will be made between treatments
19
Using NA Cohort for Decisions about Next Steps in Research Pathway
E.g. Drug D - Crizotinib
NACohort
15/30 (50%)
15/30 (50%)
10/20 (50%)
0/20 (0%)
All-comers design next
Enrichment design next
D1: Met amplified
- mixed
D2: ROS1 Gene fusions
- mixed
20
Measure Biomarkers
Biomarker+ Biomarker-
RANDOMISE
BM+Drug Control ControlBM+Drug
RANDOMISE
Stratified Trial Design
Marker-Based Strategy Design RANDOMISE
Marker-based treatment strategy Standard Care
Biomarker+ Biomarker-
Standard Care
Measure Biomarkers
BM+Drug
21
22
Basket Trial Umbrella Trial
23
When Might RCTs Not Be Needed / Ethical? Oxford / Sackett All or none criterion
E.g. Without intervention ALL patients die within 6 months VS With intervention NONE die within 6 months
Nick Black criteria Experimentation may be unnecessary when the effect is
so dramatic that unknown confounding factors could be ignored
Glasziou P, Chalmers I, Rawlins M, McCulloch P. When are randomised trials unnecessary? Picking signal from noise BMJ 2007;334:349-351 "10 x rule" – data from non-RCTs can be trusted if the
ratio of treatment effects between two alternative therapies > 10. In all other circumstances, the real treatment effects cannot be reliably separated from the effects of biases and random errors without employing RCT design.
http://personal.health.usf.edu/bdjulbeg/oncology/NON-RCT-practice-change.htm
24
Compelling Evidence From Single Case Studies / Clinical Experience
25
Compelling Evidence From A Single Arm Trial
26
What Treatment Can Create Such A Dramatic Effect? ALK Inhibitor
Patients whose tumour driven by ALK
27
Regulatory perspective on non-randomised evidence
99 trials supported approvals for 45 drugs for 68 rare cancer indications
28
Summary• More innovative statistical designs may be
needed as trials become complex• Umbrella and basket trials are an efficient
approach to stratified medicine research• Stratified medicine creates multiple rare cancers
that challenge conventional statistical designs • Developing the right drugs for the right patient to
target specific molecular drivers may create dramatic and biologically plausible treatment effects that do not require RCTs to change practice