trends and innovations in clinical trial statistics ...11 adaptive optimal design addresses the...
TRANSCRIPT
Application of optimal model-based
design in dose ranging
Sergei Leonov
April 23, 2014
Durham, NC
R&D | Biometrics & Information Sciences
Trends and Innovat ions in
Clinical Trial Stat ist ics
2
Outline
• What “optimality” means • Examples of optimal designs - Adaptive dose-ranging - Measuring efficacy and toxicity responses - PK/PD studies
• Treatment vs experimentation
• References on optimal model-based design - Classic monographs: Fedorov (1972), Silvey (1980), Pázman (1986), Pilz (1991), Atkinson
and Donev (1992), Pukelsheim (1993), Fedorov and Hackl (1997) - Recent books: Atkinson et al. (2007), Berger and Wong (2009), Morris (2010), Goos
and Jones (2011), Rasch et al. (2011), Pronzato and Pázman (2013), Fedorov and Leonov (2013) FL2013
3
Optimal model-based design: components
• Model • Controls and its admissible values • Utility function • Cost/penalty • Estimation method • Optimality criteria • Computing optimal designs • Sensitivity analysis
• Key for design/estimation: individual information matrix µ(x,θ) of a predictor x
- dose or sequence of doses in dose finding - sequence of sampling times in PK studies
4
Optimal Designs
5
Optimality criteria
Optimality criterion
(utility): • Single parameter:
size of confidence
interval
• Several parameters:
“size” of confidence
region
D-criterion: |D| = λ1 ∙λ2 = (OA ∙OB)2 ; area (V) = π (λ1 ∙λ2 )1/2 ~ ( det )1/2
A-criterion: tr D = λ1 + λ2 = (OC)2 = D11 + D22
E-criterion: λ1 = (OA)2
FL2013, Ch. 2.2
6
Cost-based (penalized) designs
FL2013, Ch. 4.1
7
Engineering locally optimal designs
DESIGN ENGINE
Model, prior information, optimality criterion, individual
information matrix, constraints, cost, penalty, …
Optimal design, comparison of different designs, sample size,
robustness, scenario analysis, benchmarking, …
FL2013, Ch.3
8
1. Adaptive design: Alzheimer’s disease study
• First-Time-in-Human, patient population • Biomarker
- Reduction expected after a single dose - Goal: find a dose with 90% reduction - Inhibition = Percentage decrease post-dose wrt baseline: 100*(Bpre-dose - Bpost-dose) / Bpre-dose
FL2013, Ch. 8.1
0
20
40
60
80
100
0.000 0.001 0.010 0.100 1.000 10.000 100.000
Dose
% I
nh
ibit
ion
at
day 2
1mean
Emax
9
Model: Emax
Measurements: Yi = η(xi , γ) + εi , i = 1,…, N {xi} - doses, N - number of patients, η – response function
Response parameters γ = (Emax , ED50 , β): – Emax : maximal response
– ED50: dose at which the response is half of Emax
– β : slope parameter
Variance model: Var εi = σ2A + σ2
M ηi (Emax – ηi ) (largest in the middle near ED50)
10
• Measurement/statistical model (Emax) • Long half-life: each new dose given to a new cohort • Cohort-dependent design region: limit on dose-escalation
− 10-fold for “small” PD effect − 5-fold once “larger” PD effect is observed
• Subject numbers: safety is a priority (FTIH)
− All cohorts: 2 subjects on placebo − First cohorts: “fixed” dose escalation
• Criterion: find doses that are “best” for estimation
− All model parameters → D-criterion − Dose ED90 → c-criterion, Var(ED90)
Adaptive optimal design with constraints
11
Adaptive optimal design
Addresses the problem of limited information at the start of the study: “locally” optimal designs, depend on model parameters
1. Collect some data using an initial design 2. Estimate model parameters using current data 3. Find optimal design (doses/weights) for current parameter estimates 4. Collect more data using this new design
Repeat steps 2-4 given available resources (# of subjects)
Box, Hunter (1965)
Size of pre-adaptation stage is critical!
12
Simulation stand-alone tool (Matlab)
Parameter
panel
Simulation
settings
Variance
model
Criterion
13
Single simulation: estimated curves Estimated curves for first cohorts: far
from “true” (doses too low to
properly estimate parameters)
After higher doses are tested, curves
start to approach the true curve
After first 4 cohorts (fixed doses) After 8 cohorts
14
Multiple simulations Explore empirical distribution of adaptive designs
• Cohort 8: “cluster” weights close to locally optimal (0.3, 0.3, 0.4)
optimal doses
15
The background normal distribution naturally incorporates correlation between responses
FL2013, Ch.6.5
2. Dose-ranging, efficacy and toxicity
16
Utility functions and optimality criteria
• Utility functions - Location of “best” dose x* and value of p10(x*)
- Knowledge of all model parameters
• Optimality criteria - Precision of estimators for maximum of the utility function and its location
Var {estimated p10(x*)} + Var {estimated location of max[p10(x)]}
- Function of the variance-covariance matrix of estimated parameters
FL2013, Ch. 2.2, 9.2.1
17
Penalties and constraints
Restricted design region: no observations beyond
X = { x: Pr{efficacy} > QE , Pr{toxicity} < QT }
FL2013, Ch.6.4.2
Pr{efficacy w/o toxicity} 1 - Pr{toxicity}
QE QT
18
To learn or to treat?
To treat:
• Increase CE and CT
• Increase QE and decrease QT
DESIGN ENGINE
To learn:
• Decrease CE and CT
• Decrease QE and increase QT
19
From locally optimal to other designs
• Locally optimal designs: guessing values of unknown parameters - "To guess is cheap, to guess wrongly is expensive“
• Bayesian designs: incorporating uncertainty of our prior knowledge
• Multistage/adaptive designs
- Size/design of initial stage is important (earlier/later examples)
• Fully adaptive (many adaptations)
- Complicated logistics, blinding, potential operational bias
- Little use when enrollment is fast and treatment results are not readily
available
- Can be inferior to two-three stage designs (Fedorov, Wu, Zhang (2012))
20
3. Design of population PK/PD studies
• What we optimize/control • When to take samples
• Number of sampling times per patient
• Goal: find the most “informative” sampling schedule
• Key: individual information matrix µ(x,θ) of a k-dimensional
predictor x (sequence of sampling times)
• Population Optimum Design of Experiments (PODE)
• Annual workshop, nonlinear mixed effects models, started in 2006
• Population optimal design tools: PFIM (INSERM, Université Paris 7, France),
PkStaMp (started at GlaxoSmithKline), PopDes (CAPKR, Univ. of Manchester),
PopED (Uppsala Univ., Sweden), WinPOPT (Univ. of Otago, New Zealand)
• 2014 PODE meeting: Basel, September 11
• 2015 PODE meeting: Cambridge, INI, July
• Software comparison: Nyberg et al. (2014, Brit. J. Clin. Pharm.)
21
Nonlinear mixed models
22
PkStaMp: input screen, 2-compartment model
FL2013, Ch. 7.5
23
0
20
40
60
80
100
0.001 0.010 0.100 1.000 10.000 100.000
DoseR
esp
on
se
EDp
p
Adaptive designs,
two approaches • Optimal design, efficient
experimentation to gather information for future patients
Robbins, Lai (1979),
Bartroff, Lai (2010)
• Allocate next patient(s) to current “best” dose, eg. current estimate of MTD
CRM, O’Quigley et al (1990) Shu,O’Quigley (2008)
Treatment vs experimentation dilemma
24
Treatment vs experimentation
(1) Original CRM: one-parameter models, appropriate for estimating a single target dose (Bayesian c-optimal design)
(2) Some extensions: two-parameter models for estimating a single target dose, examples referenced in Shu, O’Quigley (2008)
Results from optimal/adaptive control: designs of type (2)
- may converge to the “wrong” dose
- may not converge at all
Bozin, Zarrop (1991), Azriel et al. (2011)
Designs of type (2): “Best-Intention” designs (BI)
Fedorov et al. (2011)
25
“Best intention” designs
BI example:
Robbins-Monro (1951)
- finding a dose d* that
gives target response f *
yn = f(dn,θ) + εn ,
monotone f , {ε} – noise
Next dose:
dn+1 = dn – an (yn - f *)
Why it works (“on average”): if dn > d*, then yn > f *→ dn < dn+1
if dn < d*, then yn < f *→ dn > dn+1
Simplest case – linear,
f = θ1 + θ2d
26
BI: Robbins-Monro
Procedure
dn+1 = dn – an (yn - f *)
• “Small” an ~ 1/n (to avoid “big” jumps)
• Fastest convergence: an = 1/(θ2n)
Parameters unknown: how about using estimate θ*2,n?
Shu, O'Quigley (2008): “being optimal for anything other than the
best estimated treatment for the next patient, or group of
patients, to be included in the study is not acceptable''
Next plots: FL2013, Ch. 9.2
27
Linear model, optimal doses
Left panel:
Robbins-Monro (BI)
Right panel:
penalized adaptive design (PAD)
Penalty φ(x) = (x-x*)2
# of iterations:
- left columns: 100
- right : 400
Initial design:
- top row: 2-point
- bottom : 8-point
convergence to wrong doses
BI PAD
28
θ1
Linear model, parameter estimates
BI PAD
θ1 θ1 θ1
θ2
θ2
# of iterations:
- left column: 100
- right : 400
Initial design:
- top row: 2-point
- bottom : 8-point
29
Quadratic model f(x,θ) = θ1 + θ2 x + θ3 x2
convergence to wrong doses
Goal: find maxx f(x,θ)
x* = -θ2 /(2θ3 )
BI PAD
Left panel: BI
Right panel: PAD
# of iterations:
- left columns: 100
- right : 400
Initial design:
- top row: 3-point
- bottom : 12-point
30
Quadratic model, parameter estimates
BI PAD
θ2 θ2 θ2 θ2
θ3
θ3
# of iterations:
- left columns: 100
- right : 400
Initial design:
- top row: 3-point
- bottom : 12-point
31
Summary
• Optimal design: finding most “informative” levels of controls
(doses in dose-finding, sampling times in PK/PD studies, etc.)
• Validating standard/alternative designs (benchmarking)
• Test robustness of optimal designs (sampling windows, PK/PD)
• Can incorporate costs/penalties
• Can accommodate various practical constraints
• Practical, user-friendly GUI-based applications
32
References
• Azriel, D., Mandel, M., Rinott, Y. (2011). The treatment versus experimentation dilemma in dose finding studies. J. Statist. Plann. Inf., 141 (8), 2759–2768.
• Aliev A, Fedorov VV, Leonov S, McHugh B, Magee M (2012). PkStaMp library for constructing optimal population designs for PK/PD studies. Comm. Statist. Simul. Comp. 41 (6), 717–729.
• Atkinson, A.C., Donev, A. (1992). Optimum Experimental Design. Clarendon Press, Oxford. • Atkinson, A.C., Tobias, R., and Donev, A. Optimum Experimental Designs, with SAS. Oxford
University Press, Oxford. • Berger, M.P.F., Wong, W.K. (2009). An Introduction to Optimal Designs for Social and
Biomedical Research (Statistics in Practice). Wiley, Chichester. • Box, G.E.P., Hunter, W.G. (1965), Sequential design of experiments for nonlinear models. In:
Korth, J.J. (ed.), Proceedings of IBM Scientific Computing Symposium, IBM, White Plains, New York, pp. 113-137.
• Bozin, A., Zarrop, M. (1991). Self-tuning extremum optimizer – convergence and robustness properties. In: Proceedings of ECC 91, First European Control Conference, pp. 672–677.
• Elfving, G. (1952). Optimum allocation in linear regression theory, Ann. Math. Statist., 23, 255-262.
• Fedorov, V.V. (1972), Theory of Optimal Experiment. Academic Press, NY. • Fedorov, V.V., Flournoy, N., Wu, Y., Zhang, R. (2011). Best intention designs in dose–finding
studies. Preprint NI11065-DAE, Isaac Newton Institute for Mathematical Sciences, Cambridge, UK. http://www.newton.ac.uk/preprints/NI11065.pdf.
• Fedorov, V.V., Gagnon, R., Leonov, S., Wu, Y. (2007). Optimal design of experiments in pharmaceutical applications. In: Dmitrienko A, Chuang-Stein C, D’Agostino R (Eds), Pharmaceutical Statistics Using SAS. SAS Press, Cary, NC, pp. 151-195.
• Fedorov, V.V., Hackl, P. (1997). Model-Oriented Design of Experiments. Springer, NY • Fedorov, V.V., Leonov, S. (2005), Response driven designs in drug development. In: Wong,
W.K., Berger, M.P.F. (Eds.), Applied Optimal Designs, Wiley, Chichester, pp. 103-136. • Fedorov, V.V., Leonov, S.L. (2013). Optimal Design for Nonlinear Response Models. Chapman
& Hall/ CRC Biostatistics Series, Boca Raton, FL.
33
• Fedorov, V.V, Wu, Y., Zhang, R. (2012). Optimal dose-finding designs with correlated continuous and discrete responses. Statist. Med., 31 (3), 217–234.
• Goos, P., Jones, B. (2011). Optimal Design of Experiments: A Case Study Approach. Wiley, New York.
• Lai, T.L., Robbins, H. (1979). Adaptive design and stochastic approximation. Ann. Statist., 7, 1196–1221.
• Leonov, S., Miller, S. (2009), An adaptive optimal design for the Emax model and its application in clinical trials. J. Biopharm. Stat. 19 (2), 360-385.
• Lindstrom, M.J., Bates, D.M. (1990). Nonlinear mixed effects models for repeated measures data. Biometrics, 46, 673-687.
• Magnus, J.R., Neudecker, H. (1988). Matrix Differential Calculus with Applications in Statistics and Econometrics, Wiley, NY.
• Mentré, F., Mallet, A., Baccar, D. (1997). Optimal design in random-effects regression models, Biometrika, 84 (2), 429-442.
• Mentré, F., Nyberg, J., Ogungbenro, K., Leonov, S., Aliev, A., Duffull, S., Bazzoli, C., Hooker, A. (2011). Comparison of results of the dierent software for design evaluation in population pharmacokinetics and pharmacodynamics. In: Abstracts of the Annual Meeting of the Population Approach Group in Europe (PAGE). ISSN 1871-6032. http://www.page-meeting.org/default.asp?abstract=2066.
• Morris, M. (2010). Design of Experiments. An Introduction Based on Linear Models. Chapman & Hall/CRC, Boca Raton, FL.
• Neumann AU, Lam NP, Dahari H, Gretch DR, Wiley TE, Layden TJ, Perelson AS (1998). Hepatitis C viral dynamics in vivo and the antiviral efficacy of interferon therapy. Science, 282, 103-107.
References (cont.)
34
• Nyberg. J., Bazzoli, C., Ogungbenro, K., Aliev, A., Leonov, S., Duffull, S., Hooker, A., Mentré, F. (2014). Methods and software tools for design evaluation in population pharmacokinetics-pharmacodynamics. Brit. J. Clin. Pharm., early view. http://onlinelibrary.wiley.com/doi/10.1111/bcp.12352/abstract
• O’Quigley, J., Conaway, M. (2010). Continual reassessment and related dose-finding designs. Statistical Science, 25 (2), 245–257.
• O’Quigley, J., Pepe, M., Fisher, L. (1990). Continual reassessment method: a practical design for phase I clinical trials in cancer. Biometrics, 46, 33–48.
• Pázman, A. (1986). Foundations of Optimum Experimental Design. Reidel, Dordrecht. • Pilz, J. (1991). Bayesian Estimation and Experimental Design in Linear Regression Models.
Wiley, New York. • Pronzato, L. and Pázman, A. (2013). Design of Experiments in Nonlinear Models: Asymptotic
Normality, Optimality Criteria and Small-Sample Properties. Lecture Notes in Statistics, Volume 212, Springer, New York.
• Pukelsheim, F. (1993). Optimal Design of Experiments. Wiley, New York. • Rasch, D., Pilz, J., Verdooren, L.R. and Gebhardt, A. (2011). Optimal Experimental Design
with R. Chapman & Hall/CRC, Boca Raton, FL. • Robbins, H., Monro, S. (1951). A stochastic approximation method. Ann. Math. Statist., 22,
400-407.
• Shu, J., O’Quigley, J. (2008). Dose-escalation designs in oncology: ADEPT and the CRM.
Statist. Med., 27, 5345–5353.
• Silvey, S.D. (1980). Optimal Design. Chapman & Hall, London.
• Wynn, H.P. (1970) The sequential generation of D-optimal experimental designs. Ann. Math.
Statist., 41, 1655-1664.
References (cont.)
35
Back-up, Ex.1: Comparing designs, D-efficiency
1/Eff: by how much sample size has to be increased to get
same precision (information) compared to optimal design ξ*
Composite (multi-stage) design: uses placebo & fixed doses from
4 fixed cohorts
• Total 58 patients (as adaptive)
• ξ0 = {16 on placebo, 3 each on [0.001, 0.01], 6 each on [0.05, 0.25] }
• Remaining 24 patients: allocated wrt composite D-optimality,
complementing existing design (analog of Bayesian design)
Ratio of utility functions:
Efficiency: adaptive - 0.89 (st.dev. 0.02)
FL2013, Ch. 2.7
36
Combination drug for treating chronic hepatitis C (HCV) infection Neumann et al. (1998), Mentré et al. (2011)
Back-up, Ex.3, complex models: HCV dynamics
FL2013, Ch. 7.5.7