Title: Designs in nonlinear mixedeffects for a model of HIV viral load decrease: evaluation, optimisation a
1Designs in nonlinear mixed-effects for a model of
HIV viral load decrease evaluation, optimisation
and determination of the power of a treatment
effect
- France MENTRE, Sylvie RETOUT, Adeline SAMSON
- and Emmanuelle COMETS
- INSERM U738
- Dpt of Epidemiology, Biostatistics and Clinical
Research - University Hospital Bichat-Claude Bernard ,
Paris, France
2OUTLINE
- Introduction
- Model of viral load decrease
- Predicted standard error of treatment effect
- Power of the test of the treatment effect
- Design optimization with Federov-Wynn algorithm
- Conclusion
31. Introduction
- Nonlinear mixed-effects models (NLMEM)
increasingly used in drug development - pharmacokinetic / pharmacodynamic studies
- analysis of longitudinal data collected during
clinical trials - Several algorithms for maximum likelihood
estimation - NONMEM, nmle (Splus/ R), Proc NLINMIX (SAS)
- based on first-order linearization of the model
- Monolix (Matlab) Stochastic approximation EM
algorithm - (Kuhn Lavielle, CSDA, 2005)
- MCMC method to find maximum likelihood estimates
- accurate likelihood and SE obtained by importance
sampling - http// mahery.math.u-psud.fr/lavielle/monolix/lo
giciels.html
4Nonlinear mixed-effects model
- Individual model
- yi f(?i, ?i) ei
- ?i individual parameters (p components)
- e gaussian zero mean random error
- var (e) sinter2 sslope2 f(q, x) 2
- Random effects model with two treatment groups
- ?i µ b T i bi , bi N (0, ?)
- Ti 0 if group A, Ti 1 if group B
- ? diagonal wk2 Var(bik)
- Population parameters Y (P components)
- µ, b (fixed effects)
- unknowns in ? (variance of random effects)
- sinter and/or sslope (variance of residual
error)
5Population design
- Before the experiment choice of the population
design - number of patients N
- elementary design in each patient xi
- number ni of samples
- allocation of the sampling times
- General population design X x1, ..., xN
- total number of observations ntot Sni
- Grouped population design
- 1 to Q groups of elementary designs xq
- each elementary design to be performed in Nq
patients
6Fisher Information Matrix for NLMEM
- (Mentré, Mallet Baccar, Biometrika, 1997
Retout, Mentré Bruno, Stat Med, 2002)
- Fisher information matrix of population design X
- Nonlinear model f
- no analytical expression for MF (x, Y)
- first order expansion of f about b taken at 0
- assumption of independence between Var(bi) and
fixed effects - analytical expressions for MF(x, m) and MF(x, W,
s)
7Population design evaluation and optimisation
- MC simulation
- time consuming and to be repeated for each design
- Direct evaluation and optimisation using Fisher
Information matrix - Designs comparison
- For a given model, a given Y, and for each X
- predict MF, det(MF)1/P
- predict SE associated with each parameter
- PFIM in Splus/R and Matlab
- http//www.bichat.inserm.fr/equipes/Emi0357/downlo
ad.html - Design optimisation maximisation of det(MF)
- Local planification find X for a given Y
8Optimization of exact or statistical designs
- Exact design
- Fixed group structure (Q, Nq, number of samples
in xq) - Optimization of the sampling times in xq, q 1,
, Q - General algorithms simplex, simulated annealing,
NARS, (Duffull, Retout Mentré, Comput
Methods Programs Biomed, 2002) - Implementation of Simplex in Splus/R PFIMOPT
- (Retout Mentré, J Pharmacokin Pharmacodyn,
2003) - Statistical design
- ?q ( ? Nq/N) proportions of subjects in ?q
- Fedorov-Wynn algorithm for optimization of
- group structure (Q, ?q, nq)
- sampling times in xq among a given set (discrete
times) - (Mentré, Mallet Baccar, Biometrika, 1997)
- developed in C and linked with PFIMOPT
-
92. Model of viral load decrease
- (Wu, Ding de Gruttola, Stat Med, 1998 Wu
Ding, Biometrical J, 2002)
- Viral load decreases after initiation of
antiretroviral treatment in HIV1-infected
patients - Decrease can be described by a bi-exponential
model - f(?, t) Log10 (P1 exp(- l1 t) P2 exp(- l2 t)
) - y log viral load, e with constant variance
- Four parameters with additive random effects on
their log - Fixed effect b for treatment effect on first
slope - log (l1) B log (l1) A b
- Parameter values
- Fixed effects log(P1)12, log(P2)8, log(l1)
-0.7, log(l2) -3.0 - Variance of random effects 0.3
- SD of residual error 0.065
10Treatment effect on first slope
- First rate-constant 30 greater in group B vs A
- b 0.262
- Similar design in all patients 6 samples
- weeks 1, 3, 7, 14, 28, 56 after treatment
initiation
Group B 40 patients
Group A 40 patients
113. Predicted standard error of treatment effect
- Evaluation of SE of b under H0 b 0
- Two groups of 100 patients, same elementary
design - Predicted SE two approaches
- Extension of PFIM for covariates (in R)
- (Retout Mentré, J Biopharm Stat, 2003)
- Monolix (SAEM in Matlab)
- one simulation of 5 000 patients per group
(closer to asymptotic) - estimated SE divided by 501/2
- Empirical SE from replicated simulations under H0
- 100 simulations of 100 patients per group
- Distribution and SD of estimated b
- Estimation algorithm 1. nlme (in R) and 2.
Monolix
12Results on SE of b
- Predicted SE
- PFIM 0.079
- Monolix 0.078
- Close results of PFIM (with linearization) and
SAEM (exact approach but time consuming) - Empirical SE (on 100 replications)
- nlme 0.085
- Monolix 0.086
- Close results (on this example) of the two
estimation algorithms - Predicted SE smaller than empirical SE
- MF-1 lower bound of estimation variance
13Results on other estimation CV ()
Predicted
Empirical
144. Power of the test of the treatment effect
- Wald test performed to test H0 b 0
- For a given population design
- compute PFIM predicted SE under H1
- evaluate the power of Wald test under H1 (upper
bound) - number of subjects needed for a given power
- SE proportional to N1/2
155. Design optimization with FW algorithm
Total of 480 samples in set 0,1,2,3,5,7,10,14,21,
28,42,56
b Rounded optimal statistical design
a Previous non-optimized design
16Evaluation of the power by simulation
Total of 480 samples in set 0,1,2,3,5,7,10,14,21,
28,42,56
b Rounded optimal statistical design
a Previous non-optimized design
17Example of PFIMOPT 2.0 output with Fedorov-Wynn
algorithm (1)
18Example of PFIMOPT 2.0 output with Fedorov-Wynn
algorithm (2)
196. Conclusion
- Optimal design in NLMEM using FIM
- Interesting and growing field with great
potential applications - Avoid extensive repeated simulations
- Power of Wald test and number of patients needed
- Optimal allocation of sampling times
- Main limitations
- Based on asymptotic results
- evaluate best designs by simulation
- Local planification based on Y
- sensitivity analysis
- robust or Bayesian planification (Han Chaloner,
Biometrics, 2004) - Global D-optimality for all parameters in Y
- Ds optimality on subset of parameters
- optimal compound design
20Future developments
- Optimisation with IOV
- balance number of occasions / number of samples
per occasion - Optimisation with covariates
- given distribution
- optimal designs across patients
- optimal designs with respect to covariates values
- optimisation of distribution
- find best designs and best covariate distribution
- Optimisation in PK/PD (more than one variable)
- Optimal design for subset of parameters
(DS-optimality) - Optimal compound designs
- Need for more exact evaluation of FIM witha SAEM?