Sunteți pe pagina 1din 16

This article was downloaded by: [Aston University]

On: 23 August 2014, At: 14:42


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer
House, 37-41 Mortimer Street, London W1T 3JH, UK

Journal of Biopharmaceutical Statistics


Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/lbps20

OPTIMAL SAMPLING TIMES IN BIOEQUIVALENCE TESTS


a a
Fan Hui Kong & René Gonin
a
Westat , 1441 West Montgomery Avenue, Rockville, Maryland, 20850, U.S.A.
Published online: 16 Aug 2006.

To cite this article: Fan Hui Kong & René Gonin (2000) OPTIMAL SAMPLING TIMES IN BIOEQUIVALENCE TESTS, Journal of
Biopharmaceutical Statistics, 10:1, 31-44, DOI: 10.1081/BIP-100101011

To link to this article: http://dx.doi.org/10.1081/BIP-100101011

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of
the Content. Any opinions and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied
upon and should be independently verified with primary sources of information. Taylor and Francis shall
not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other
liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or
arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
Journal of Biopharmaceutical Statistics, 10(1), 31–44 (2000)

OPTIMAL SAMPLING TIMES IN


BIOEQUIVALENCE TESTS
Downloaded by [Aston University] at 14:42 23 August 2014

Fan Hui Kong and René Gonin

Westat
1441 West Montgomery Avenue
Rockville, Maryland 20850

Key words. Bioequivalence; Bioavailability; Optimal sampling design;


Compartment model; Concentration-time curve

Abstract

In bioequivalence studies, drug formulations are compared in terms of


bioavailability parameters such as the area under the concentration-time
curve (AUC), the maximum concentration (C max ), and the time to maxi-
mum concentration (t max ). Accuracy in measuring these parameters di-
rectly affects the accuracy of bioequivalence tests. Because the number
of blood draws per patient is limited, the blood collection times must
be spaced so that concentration-time curve measurements can produce
accurate bioavailability parameter estimates. This paper describes an
optimization approach for calculating optimal time designs for one-
compartment models, but is sufficiently general for other compartmental
models. Simulation indicates that the optimal design improves the accu-
racy of AUC estimation.

1. Introduction

Bioequivalence studies are one of the most frequently performed investigations


in clinical trials. In such studies, two or more formulations are compared by means
of a randomized crossover design, and pharmacokinetic variables are measured
from the drug concentration in blood, serum, or plasma. The most frequently used
pharmacokinetic variables or bioavailability parameters are the area under the

31

Copyright  2000 by Marcel Dekker, Inc. www.dekker.com


ORDER REPRINTS

32 Kong and Gonin

concentration-time curve (AUC), which measures the total amount of drug ab-
sorbed in the body; the maximum concentration (C max ); and the time to reach max-
imum concentration (t max ).
The ability to measure pharmacokinetic variables accurately is thus of critical
importance in a bioequivalence study. However, most methods for conducting
bioequivalence tests assume that the pharmacokinetic variables have been mea-
sured accurately. In fact, the values of these variables rely on the concentration-
time curve, C(t), which can be measured only at certain predetermined blood
sampling times. For example, AUC is calculated by a trapezoidal or other approxi-
mation method using the values of C(t) at those time points (see Bailer and Pie-
gorsh (1)). Because only a limited number of time points can be chosen, the
appropriate design of these time points becomes crucial to the accuracy with which
Downloaded by [Aston University] at 14:42 23 August 2014

pharmacokinetic variables are measured. Wei (2) has suggested a simple proce-
dure for selecting sampling time points to reduce the mean squared error of the
trapezoidal approximation of the AUC. However, his example shows that his pro-
cedure does not produce an optimal design. Chang and Wong (3) proposed proce-
dures to select optimal blood sampling time points in phase II or III clinical trials
in order to investigate pharmacokinetic/pharmacodynamic relationships. Because
their procedure only searches within the previously fixed time points, it therefore
may not locate globally optimal time points.
Optimal sampling time designs have been proposed in pharmacokinetic stud-
ies as a means to estimating pharmacokinetic model (PK model) parameters such
as the rate of absorption k a , the rate of elimination k e , and the apparent volume
of distribution V (D’Argenio (4); Tod and Rocchisani (5)). In a bioequivalence
study, however, only the bioavailability parameters AUC, C max, and t max are to be
compared; we therefore assume that the PK model for each drug formulation is
given so that the corresponding concentration-time curve C(t) is well defined. For
a given population of study subjects, the PK model parameters V, k a , and k e are
assumed to follow certain probability distributions with certain known parameters.
Of course, the estimation of these parameters would require that some pharmaco-
kinetic studies of the given population have been undertaken.
The main objective of this paper is to provide an optimal selection procedure
that substantially improves the accuracy of AUC estimation in a bioequivalence
test. Our procedure is easy to describe to practitioners, and our simulation study
indicates that AUCs approximated by the trapezoidal method using our time de-
sign improve the corresponding AUC estimations using conventional times de-
signs in most cases.

2. Optimal Time Design in Bioequivalence Studies

Suppose the drug concentration-time curve C(t) has an analytic form that depends
on parameters V, k a , and k e , which represent certain characteristics of the study
ORDER REPRINTS

Optimal Sampling Times 33

subjects. Thus, for a specific population and formulation, these parameters follow
some joint distribution. We assume that this distribution can be obtained by early
experience or certain pharmacokinetic studies. For a fixed set of time points t 0 ⫽
0 ⱕ t 1 ⱕ ⋅ ⋅ ⋅ ⱕ t m , the profile, Ĉ(t), consists of piecewise straight lines that connect
C(t i⫺1 ) and C(t i ) for i ⫽ 1, . . . , m. Then the squared deviation between C(t) and
Ĉ(t) for t i⫺1 ⱕ t ⱕ t i is


ti
{C(t) ⫺ Ĉ(t)} 2 dt (1)
t i⫺1

The total sum of squared deviations is the summation for i ⫽ 1, . . . , m, which


depends on the parameters V, k a , and k e . Taking its expectation with respect to
the distribution of these parameters, one obtains
Downloaded by [Aston University] at 14:42 23 August 2014

冤冱 冮 冥
m ti
E {C(t) ⫺ Ĉ(t)} 2 dt (2)
t i⫺1
i⫽1

The optimal time points are determined by minimizing this expectation subject
to the order restrictions t 0 ⫽ 0 ⱕ t 1 ⱕ ⋅ ⋅ ⋅ ⱕ t m . To avoid the degenerate case
(all t i → 0), we determine a positive number T f and require that t m ⱖ T f .
This is a linearly constrained nonlinear programming problem, and a gen-
eral theory and methodology have been developed in the optimization literature
(Fletcher (6)). A summary of optimality conditions are provided in the Appendix.
Although a general theory of optimization is available, great difficulties may be
encountered in implementation, especially when parameters V, k a , and k e have a
three-dimensional joint distribution. Numerical integration and differentiation are
required in the minimization of Expression (2).

2.1 A One-Compartment Model

Although the above approach applies to general PK models and parameter distri-
butions, we shall consider the one-compartment open model as a special case. It
is assumed that the distribution of a drug follows a one-compartment model with
first-order absorption, which is the situation most frequently encountered in prac-
tice. For this model, the concentration-time curve has the form
Dk a
C(t) ⫽ {e ⫺ke t ⫺ e ⫺k a t} (3)
V(k a ⫺ k e)
where D, V, k a , and k e are positive constants and 0 ⱕ t ⬍ ∞. In particular, D
represents the dose of drug administered with V, k a , and k e as before. It is reason-
able to assume that D and V are constants in bioequivalence tests conducted in
healthy volunteers. Although k a may depend on the formulation and the individual,
it usually has low variability, and we thus assume it to be constant. The parameter
with the largest variation, k e , directly determines the total AUC, which equals
ORDER REPRINTS

34 Kong and Gonin

D/(Vk e ) if model (3) is valid. The total AUC is approximated using the trapezoidal
method. Because k e ⬎ 0, it is reasonable to assume a lognormal distribution for
k e , with density function
1 1 µ 2
e ⫺2 冢logx⫺σ冣 (4)
σ√2π x
for 0 ⬍ x ⬍ ∞, where µ and σ 2 are the mean and variance of the logarithm of k e .
Given that ln k e ⬃ N(µ, σ 2 ), then k e ⬃ lognormal(µ k e , σ 2ke ), where mean µ k e ⫽
e µ⫹0.5σ 2 and variance σ 2ke ⫽ (e σ2 ⫺ 1)µ 2ke .
For the one-compartment model, the squared deviation for each i in Expres-
sion (1) has a closed-form expression, as we shall see subsequently. The expecta-
tion (2) with respect to distribution (4) is computed by means of a numerical
Downloaded by [Aston University] at 14:42 23 August 2014

integration.

2.2 Determination of the Time Interval [0, Tf ]

The next step is to choose a positive number Tf so that the optimal time points
satisfy the restriction Tf ⱕ t m . The selection of Tf affects the accuracy of the ap-
proximated AUC. Given the values of D, V, k a , and k e , we find that high accuracy
is achieved if one chooses Tf so that the AUC on the right-hand side of Tf is
approximately 3% of the total AUC. In general, let p be the percentage of the
AUC on the right-hand side of Tf , then Tf is determined by solving the equation
k a e ⫺ke T ⫺ k e e ⫺kaT
⫺p⫽0 (5)
ka ⫺ ke
for T. Note that T → ∞ as p → 0. Table 1 gives the solution, Tf , for various values
of k e , k a , and p.
From Table 1 we see that Tf depends heavily on k a and k e , and for most of
k a and k e the values of T f are too large to be of practical use. In fact, in many
bioequivalence studies, subjects may only be kept in the clinical center for at most
one day, so T f ⱕ 24. In that case the overall AUC is more difficult to determine,
hence we find the optimal time points so that the AUC over the period [0, 24]
can be determined more accurately.

Table 1. Tf for Various Values of k a , k e , and p


k a ⫽ 0.50 k a ⫽ 0.70
k e ⫽ 0.05 k e ⫽ 0.10 k e ⫽ 0.15 k e ⫽ 0.05 k e ⫽ 0.10

p 0.03 0.05 0.03 0.05 0.03 0.05 0.03 0.05 0.03 0.05
Tf 72.2 62.0 37.3 32.2 25.8 22.4 71.6 61.4 36.6 31.5

Tf is the number of hours above which the AUC equals p.


ORDER REPRINTS

Optimal Sampling Times 35

2.3 Optimal Times with Both k e and ka as Fixed Constants

For a one-compartment model, the nonlinear objective function (2) entails integra-
tion since k e follows a lognormal distribution that complicates the optimization
procedure. We first consider the special case where k e is assumed to be constant
so the objective function does not involve integration.
For ease of analytical detail, define
C 0 (t) ⫽ e ⫺k et ⫺ e ⫺kat (6)
ignoring the constant Dk a /V(k a ⫺ k e ) for the time being. Let the straight-line
estimate of concentration be
Downloaded by [Aston University] at 14:42 23 August 2014

Ĉ 0 (t) ⫽ a i ⫹ b i t, for t i⫺1 ⱕ t ⱕ t i


where
C 0 (t i )t i⫺1 ⫺ C 0 (t i⫺1 )t i
ai ⫽ ⫺ (7)
t i ⫺ t i⫺1
C 0 (t i ) ⫺ C 0 (t i⫺1 )
bi ⫽ (8)
t i ⫺ t i⫺1
The constant Dk a /V(k a ⫺ k e ) will now be taken into consideration by multiplying
all expressions by this factor. Thus, the objective function to be minimized be-
comes

冦 冧 冱冮
2 m ti
Dk a
{(e ⫺ke t ⫺ e ⫺k a t ) ⫺ a i ⫺ b i t} 2 dt (9)
V(k a ⫺ k e ) i⫽1
t i⫺1

subject to constraints:
c i ⫽ t i ⫺ t i⫺1 ⱖ 0, for i ⫽ 1, . . . , m
c m⫹1 ⫽ t m ⫺ Tf ⱖ 0
Under these restrictions, one always has t m ⱖ Tf . Integration of Expression (9)
with respect to t yields


ti
(e ⫺ke t ⫺ e ⫺ka t ⫺ a i ⫺ b i t) 2 dt
t i⫺1

e ⫺2k a t e ⫺2ke t (2k a a i ⫹ 2b i ) e ⫺ka t 2b i te ⫺ka t 2b i te ⫺ke t


⫽⫺ ⫺ ⫺ ⫺ ⫹ (10)
2k a 2k e k 2a ka ke


ti
(2k e a i ⫹ 2b i )e ⫺ke t b2 t3 2e ⫺(ke⫹k a)t
⫹ ⫹ a t ⫹ i ⫹ ai bi t2 ⫹
2
i
k 2e 3 ke ⫹ ka t i⫺1
ORDER REPRINTS

36 Kong and Gonin

Table 2. Optimal Sampling Time Designs for Model (3) with Fixed k e and k a for m ⫽ 6, 7, 8, 9,
and 10 Design Points a

AUC Total
m t1 t2 t3 t4 t5 t6 t7 t8 t9 t 10 to Tf AUC b

6 1.59 3.47 5.79 8.94 20.06 24.0 — — — — 81.80 98.31


7 1.29 2.76 4.48 6.59 9.40 19.61 24.0 — — — 82.16 98.66
8 1.08 2.29 3.65 5.24 7.17 9.70 19.39 24.0 — — 82.37 98.87
9 0.93 1.95 3.08 4.36 5.83 7.61 9.90 19.27 24.0 — 82.50 99.00
10 0.94 1.98 3.14 4.44 5.96 7.79 10.20 18.18 21.09 24.0 82.38 98.87
a
V ⫽ 40, D ⫽ 400, k a ⫽ 0.2, k e ⫽ 0.1, total AUC ⫽ 100, Tf ⫽ 24, and truncated AUC ⱕTf ⫽ 82.68.
b
AUC is the trapezoidal approximation of AUC.
Downloaded by [Aston University] at 14:42 23 August 2014

which is evaluated for each partition, (t i⫺1 , t i ), i ⫽ 1, . . . , m, thus the objective


function has a closed form. To perform the optimization procedure, we first calcu-
late a i , b i using formulæ (7) and (8). These quantities are then substituted into
Expression (10) to derive the definite integral for each partition before the final
summation is made.
We used a general nonlinear programming algorithm that solves a quadratic
programming problem (Appendix) at each iteration of the algorithm. Powell’s
(7) implementation of sequential quadratic programming was used to solve prob-
lem (9). The algorithm converges quite well for different starting values of
t1, . . . , tm.
So far we have calculated only the approximate AUC on interval [0, Tf ] and
ignored the AUC beyond time point t m ⫽ Tf . Note that over [Tf , ∞), the AUC
approximately equals C(Tf )/k e . Adding this to the approximate AUC over [0, Tf],
an estimate of AUC over [0, ∞) is obtained.
To demonstrate the optimization process, we use two examples and as-
sume typical values V ⫽ 40 and D ⫽ 400 as used in practice. In the first exam-

Table 3. Optimal Sampling Time Designs for Model (3) with Fixed k e and k a for m ⫽ 6, 7, 8, 9,
and 10 Design Points a

AUC Total
m t1 t2 t3 t4 t5 t6 t7 t8 t9 t 10 to Tf AUC b

6 0.74 1.71 3.16 11.38 16.91 24.0 — — — — 89.82 100.18


7 0.55 1.23 2.10 3.35 11.10 16.74 24.0 — — — 90.26 100.63
8 0.57 1.27 2.18 3.51 10.11 13.96 18.48 24.0 — — 89.70 100.07
9 0.46 1.00 1.65 2.49 3.67 9.90 13.80 18.39 24.0 — 89.92 100.29
10 0.46 1.01 1.68 2.54 3.77 9.37 12.34 15.66 19.48 24.0 89.66 100.03
a
V ⫽ 40, D ⫽ 400, k a ⫽ 0.8, k e ⫽ 0.1, total AUC ⫽ 100, Tf ⫽ 24, and truncated AUC ⱕTf ⫽ 89.63.
b
AUC is the trapezoidal approximation of the AUC.
ORDER REPRINTS

Optimal Sampling Times 37

ple we assume that k a ⫽ 0.2, k e ⫽ 0.1, and Tf ⫽ 24.0 so the true AUC is 100
(⫽ D/(Vk e )), and the true truncated AUC up to time Tf is 82.68. Table 2 re-
ports the optimal time designs for 6 ⱕ m ⱕ 10 and the corresponding approxi-
mate AUCs. In the second example, we choose k a ⫽ 0.8 so the rate of absorption
is higher and the true truncated AUC to Tf is 89.63. Table 3 reports the optimal
time designs and the corresponding approximate AUCs.

2.4 Optimal Times with k a , a Fixed Constant, and k e ,


a Random Variable
2.4.1 Optional Time Points
In general, k e is a random variable that varies with study subjects, and an optimal
Downloaded by [Aston University] at 14:42 23 August 2014

time design for one subject may not be optimal for another. By assuming that k e
follows a lognormal distribution with known parameters µ ke and σ ke , we can mini-
mize the expected objective function (2) instead of Expression (9), where (2) is
the expectation of (9) with respect to the distribution of k e . First, note that we
have a class of concentration-time curves, but that only one Tf value is needed.
We therefore apply the method in Section 2.2 only to the curve with mean values
of µ k e and σ ke to derive the last sampling time point Tf . Thus the AUC over [Tf ,
∞) is p100% of the total AUC for that particular curve only, and the remainder
of the total AUC must be estimated for the remaining curves. Because k e must
be smaller than k a in Expression (9) in order for the objective function to be finite,
we truncate the lognormal distribution of k e to x ⬍ k a . Once t m ⫽ Tf is fixed, the
objective function can be minimized in terms of the unknowns, t 1 , . . . , t m , and
the constants k a , V, D, µ ⫽ µ ke and σ k e . It has the functional form

冮 冦冱 冮 冧
ka m ti
D 2 k 2a
{(e ⫺xt ⫺ e ⫺ka t ) ⫺ a i (x) ⫺ b i (x)t} 2 dt
V 2 σ√2π 0
i⫽1
t i⫺1

1 1 log x⫺µ 2
⫺ 冦冢
⫻ e 2 σ 冣冧 dx
x(k a ⫺ x) 2

Table 4. Optimal Sampling Time Designs for Model (3) with Fixed k a and Random k e for m ⫽
6, 7, 8, 9, and 10 Design Points a
m t1 t2 t3 t4 t5 t6 t7 t8 t9 t 10

6 1.13 2.61 4.85 10.71 16.53 24.0 — — — —


7 0.89 1.99 3.43 5.60 11.36 16.91 24.0 — — —
8 0.76 1.67 2.80 4.32 6.77 11.93 17.26 24.0 — —
9 0.71 1.54 2.55 3.86 5.76 9.92 13.90 18.44 24.0 —
10 0.61 1.32 2.14 3.15 4.47 6.43 10.47 14.29 18.68 24.0
a
V ⫽ 40, D ⫽ 400, k a ⫽ 0.5, µ ke ⫽ 0.1, σ ke ⫽ 0.05, and Tf ⫽ 24.
ORDER REPRINTS

38 Kong and Gonin

subject to constraints:

c i ⫽ t i ⫺ t i⫺1 ⱖ 0, for i ⫽ 1, . . . , m
c m⫹1 ⫽ t m ⫺ Tf ⱖ 0

Given the values of the constants V, D, k a , and µ k e , σ ke , the optimization process


consists of an inner iteration that performs the numerical integration and an outer
iteration in which the objective function is decreased. A pharmacokinetics study
is usually performed for each formulation of the drug in order to estimate the
model parameters V, D, k a , µ ke , and σ k e before bioequivalence studies can be car-
ried out.
To estimate the AUC for each curve over [Tf , ∞), note that it can be approxi-
Downloaded by [Aston University] at 14:42 23 August 2014

mated by C(Tf )/k e . However, k e is usually unknown for a fixed curve C(t), but
can be estimated by means of linear regression on the last three values of ln C(t).
Adding the result to the trapezoidal approximation over [0, Tf ] produces an esti-
mate of the total AUC.
Because no optimization method is guaranteed to converge globally for gen-
eral nonlinear objective functions, the optimal sampling times obtained may only
represent a local solution. Our suggestion is to try different starting values of the
parameters and compare the values of the objective functions at their respective
optimal points and select the best one. Our simulation study indicates that the
optimal time points with our procedure are quite robust to the starting values. As
a general rule, one should include the designs that are used in practice and are
believed to be good among starting values.

2.4.2 Numerical Optimization Results


As examples of optimal sampling designs for constant k a and random k e , we
choose D ⫽ 400, V ⫽ 40, and assume that µ ke ⫽ 0.1 with the coefficient of
variation (C.V.), σ ke /µ ke ⫽ 0.5, so that the standard deviation of k e equals 0.05,

Table 5. Optimal Sampling Time Designs for Model (3) with Fixed k a and Random k e for m ⫽
6, 7, 8, 9, and 10 Design Points a
m t1 t2 t3 t4 t5 t6 t7 t8 t9 t 10

6 0.78 1.83 3.48 9.61 15.70 24.0 — — — —


7 0.61 1.38 2.41 4.03 9.74 15.79 24.0 — — —
8 0.60 1.34 2.32 3.83 8.30 12.43 17.53 24.0 — —
9 0.49 1.08 1.81 2.78 4.28 8.53 12.62 17.64 24.0 —
10 0.48 1.05 1.75 2.67 4.05 7.49 10.66 14.31 18.68 24.0
a
V ⫽ 40, D ⫽ 400, k a ⫽ 0.8, µ ke ⫽ 0.1, σ ke ⫽ 0.05, and Tf ⫽ 24.
ORDER REPRINTS

Optimal Sampling Times 39

Table 6. Optimal Sampling Time Designs for Model (3) with Fixed k a and Random k e for m ⫽
6, 7, 8, 9, and 10 Design Points a
m t1 t2 t3 t4 t5 t6 t7 t8 t9 t 10

6 1.68 3.70 6.24 9.79 16.73 24.0 — — — —


7 1.41 3.04 4.99 7.45 10.91 17.63 24.0 — — —
8 1.22 2.59 4.18 6.09 8.50 11.96 18.19 24.0 — —
9 1.09 2.30 3.66 5.25 7.15 9.60 13.21 18.76 24.0 —
10 0.99 2.06 3.26 4.62 6.22 8.16 10.69 14.59 19.37 24.0
a
V ⫽ 40, D ⫽ 400, k a ⫽ 0.2, µ ke ⫽ 0.1, σ ke ⫽ 0.05, and Tf ⫽ 24.
Downloaded by [Aston University] at 14:42 23 August 2014

as is typical in practice. In this case, the mean and standard deviation of ln k e are
µ ⫽ ⫺2.4142 and σ ⫽ √ln(1 ⫹ 0.5 2 ) ⫽ 0.4724, respectively.
We chose the last sampling point to be 24.0 as is common practice. For k a ⫽
0.5, the optimal time values for m ⫽ 6, . . . , 10 are reported in Table 4. For k a ⫽
0.8 and 0.2, the results are reported in Tables 5 and 6. Dramatic differences be-
tween optimal time values for different values of k a were observed.

3. A Simulation Study

The purpose of this simulation study is to conduct a sensitivity analysis of AUC


estimation under different designs, i.e., the optimal design we proposed and some
practically used designs. Furthermore, we would like to determine the minimal
number of sampling time points beyond which the gain in the accuracy of trapezoi-
dal AUC is negligible.

3.1 Comparing Optimal and Regular Designs

We would like to see how much of the AUC is missed if a trapezoidal method
is used for each of these designs. We used µ ⫽ ⫺2.4142 and σ ⫽ 0.4724, thus
µ ke ⫽ 0.1 and σ ke ⫽ 0.05, as before, and generated 1000 random deviates from
this lognormal distribution. For a fixed value of k a this yields 1000 concentration-
time curves. Simulation studies were conducted using various values of k a and
Tf . For most of these situations, the optimal time design we proposed improves
the regular design. There were a few exceptions in which the regular design out-
performed the optimal design. This indicates that our procedure of finding the
optimal sampling times could be improved further.
We now provide a typical example of such a simulation. For k a ⫽ 0.5, m ⫽
6 design points, and Tf ⫽ 24.0, the optimal time points are T * ⫽ (1.1, 2.6, 4.9,
10.7, 16.5, 24.0) (Table 4) and for the chosen regular design T R ⫽ (1, 2, 4, 8,
ORDER REPRINTS

40 Kong and Gonin

16, 24), which is quite close to the optimal one. This might not be the case if
different values of k a or µ k e were selected, in which case quite different optimal
designs would be obtained. For each value of k e , we calculated the true AUC up
to 24 hours and its trapezoidal approximations under T * and T R. For the first 15
simulated values of k e , the corresponding true AUC over [0, 24.0] and trapezoidal
approximations using both designs are reported in Table 7. The second column
in Table 7 reports the true AUC up to 24.0. Columns 3 and 4 report the approxi-
mate AUC and the percentage of unexplained AUC under the regular design T R.
Columns 5 and 6 report the corresponding approximations and percentages under
the optimal time design T *. The mean percentage of the unexplained AUC for
regular design T R is ⫺1.1% (SD ⫽ 1.5%) with a median of ⫺0.7% and range
[⫺6.4%, 0.9%]; the mean percentage of unexplained AUC for the optimal design
Downloaded by [Aston University] at 14:42 23 August 2014

T * is 0.001% (SD ⫽ 0.9%) with median of 0.3% and range [⫺5.1%, 1.05%].
The simulation study indicates that the trapezoidal approximation for the AUCs
up to 24 hours under the optimal design is an improvement over the regular time
design for m ⫽ 6. In particular, the regular design has a nonnegligible bias while
the optimal design reduces it.

Table 7. Simulated True AUCs, Their Trapezoidal Approximations, and Per-


centage Differences Under Regular and Optimal Time Designs for Model (3) with
Fixed k a and Random k e for Six Sampling Time Points a
True
AUC
ke to Tf AUC R % diff R AUC* % diff *
0.1136 80.58 81.46 ⫺1.1 80.51 0.1
0.1629 59.57 61.11 ⫺2.6 60.09 ⫺0.9
0.0590 122.91 122.43 0.4 121.83 0.9
0.0719 110.18 110.07 0.1 109.37 0.7
0.1990 49.56 51.33 ⫺3.6 50.34 ⫺1.6
0.0487 134.60 133.81 0.6 133.29 1.0
0.1047 85.70 86.40 ⫺0.8 85.50 0.2
0.2113 46.81 48.62 ⫺3.9 47.66 ⫺1.8
0.1106 82.26 83.08 ⫺1.0 82.15 0.1
0.1027 86.94 87.60 ⫺0.8 86.71 0.3
0.2424 41.01 42.88 ⫺4.6 42.0 ⫺2.4
0.0505 132.43 131.70 0.6 131.17 1.0
0.0837 100.19 100.41 ⫺0.2 99.62 0.6
0.0924 93.81 94.23 ⫺0.5 93.39 0.4
0.2114 46.79 48.61 ⫺3.9 47.64 ⫺1.8
a
V ⫽ 40, D ⫽ 400, k a ⫽ 0.5, k e ⬃ lognormal (0.1, 0.05 2 ), and Tf ⫽ 24.0. The regular time
design is T R ⫽ (1.0, 2.0, 4.0, 8.0, 16.0, 24.0), the optimal time design is T * ⫽ (1.13, 2.61,
4.85, 10.71, 16.53, 24.0). The table only reports the first 15 simulation results.
ORDER REPRINTS

Optimal Sampling Times 41

We note that both the regular and the optimal designs tend to overestimate
the small AUCs and underestimate the large AUCs. We investigated this phenom-
enon further and performed additional simulations for different means and vari-
ances of k e (yielding different AUCs) and compared the biases between the regular
and optimal designs. The results indicate that almost all the finite time designs
have this imperfection. For larger elimination rates k e (total AUC becomes
smaller), the regular design exhibits more bias and the optimal design reduces the
bias more. Examples are shown in Table 8. In real applications, the magnitude
of error may be larger. In fact, the values of C(t i ) are usually measured in labora-
tories and therefore subject to measurement error. Such measurement errors will
distort the profile of C(t) and inflate the error in the estimated trapezoidal AUC.
Downloaded by [Aston University] at 14:42 23 August 2014

3.2 Determining the Number of Sampling Times

In practice, choosing the appropriate number of sampling times m is of great inter-


est. Although increasing m will in general increase the accuracy of the trapezoidal
AUC, it will also increase the workload of the clinic center and the inconvenience
to the study subject. On the other hand, there may be some value of m beyond
which the gain in accuracy is negligible. To determine the value of m, we con-
ducted a series of simulations. We used the model in which k a is fixed and k e is
random with k e ⬃ lognormal (0.1, 0.05 2 ). For each combination of k a (0.2, 0.5,
0.8) and m (6, 7, 8, 9) and its corresponding optimal design, we generated 1000
deviates of k e and calculated the true and trapezoidal AUC and their corresponding
percentage differences. We evaluated the differences by computing their means,
standard deviations as well as medians and ranges. The results are depicted in
Table 9. We observe that the accuracy of the trapezoidal AUC depends on both
k a and m. For k a ⫽ 0.2, it seems that there is little gain for m beyond 6. For k a ⫽
0.5 and 0.8, it seems that beyond m ⫽ 8 there is little gain since the reduction
in both the standard deviation and the range of percentage differences is small.
So, based on the limited simulation, we suggest that the proper choice of the
number of sampling times should range from six to eight.

Table 8. Comparison of Percentage Unexplained AUC Between the Regular and the Optimal Time
Designs for k a ⫽ 0.5 and m ⫽ 6
Regular design Optimal design

µ ke σ ke Mean (SD) Median (range) Mean (SD) Median (range)


0.2 0.1 ⫺3.08 (1.85) ⫺3.03 (⫺6.4, 0.7) ⫺1.64 (1.56) ⫺1.39 (⫺6.2, 1.0)
0.3 0.15 ⫺4.49 (1.60) ⫺4.84 (⫺6.4, 0.3) ⫺3.0 (1.47) ⫺3.01 (⫺6.0, 0.6)
0.4 0.2 ⫺5.14 (1.35) ⫺5.57 (⫺6.4, ⫺0.1) ⫺0.79 (0.21) ⫺0.80 (⫺1.1, 0.3)
ORDER REPRINTS

42 Kong and Gonin

Table 9. The Range and Standard Variation of Percentage Unexplained AUC for Different Number
of Time Points
k a ⫽ 0.20 k a ⫽ 0.50 k a ⫽ 0.80

m Mean (SD) Median (range) Mean (SD) Median (range) Mean (SD) Median (range)
6 1.41 (0.2) 1.59 (0.6, 1.6) ⫺0.09 (0.9) 0.26 (⫺5.1, 0.6) ⫺1.08 (1.8) 0.48 (⫺12.0, 0.5)
7 1.18 (0.2) 1.27 (0.5, 1.4) ⫺0.50 (1.0) ⫺0.18 (⫺5.0, 0.4) ⫺1.44 (1.7) ⫺0.96 (⫺10.0, 0.2)
8 1.05 (0.2) 1.12 (0.5, 1.2) ⫺0.55 (0.7) ⫺0.34 (⫺3.2, 0.2) ⫺0.51 (0.9) ⫺0.24 (⫺5.5, 0.2)
9 1.0 (0.1) 1.06 (0.6, 1.1) ⫺0.18 (0.5) ⫺0.02 (⫺2.2, 0.2) ⫺0.7 (0.8) ⫺0.4 (⫺4.9, 0.1)
Downloaded by [Aston University] at 14:42 23 August 2014

4. Conclusion

We have proposed an optimal sampling time procedure for bioequivalence studies


in which the profile obtained is closest to the true drug concentration-time curve
profile. Although our procedure can still be improved, it does have distinct advan-
tages. Minimization of the expected total sum of squared deviations exhibits fast
convergence, while yielding stable results. The profile based on T * resembles the
shape of the true concentration-time curve; thus, other parameters (e.g., C max and
t max ) can also be estimated using the profile based on T *. To implement the proce-
dure, one must correctly specify the drug distribution model as well as its model
parameters. Optimal sampling times based on an incorrect model may yield poor
estimates of AUC. The robustness of AUC approximation for different time de-
signs requires further study. Correct model construction and parameter estimation
are usually performed in a pharmacokinetics study prior to the bioequivalence
study. Given the difference in pharmacokinetic parameter values for different for-
mulations, we suggest that pharmacokinetics studies be performed for each drug
formulation so that an optimal time design can be obtained for each formula-
tion.
In more general situations, pharmacokinetic parameters V, k a , and k e are
random variables following some trivariate joint distribution. Thus the expectation
(2) should be taken with respect to this joint distribution. The optimization proce-
dure will therefore include more complicated numerical integration. An evaluation
of how this optimal sampling time design affects bioequivalence tests is beyond
the scope of this paper. However, our simulation results indicate that the proposed
optimal time design yields small bias in estimating the total AUC, and the bioequi-
valence test based on the calculated AUC should therefore be more reliable. The
simulation study also indicates that the proper number of sampling times should
range from six to eight.
ORDER REPRINTS

Optimal Sampling Times 43

Appendix: Optimality Conditions for Constrained


Optimization Problems

Given an objective function f (x), which is defined on x ∈ ℜ n, the linearly con-


strained nonlinear programming problem is defined as
min f(x)
(11)
s.t. a iT x ⱖ b i for i ⫽ 1, . . . , m
for a certain integer m, with a i an n ⫻ 1 vector and b i some real number.
The optimum solution can only be obtained under certain conditions, which
we call optimality conditions. To state these conditions, we first introduce
some concepts: The set of active constraints is defined as Ꮽ(x) ⫽ {i |aTi x ⫽ b i ,
Downloaded by [Aston University] at 14:42 23 August 2014

i ⫽ 1, . . . , m}, the inactive set is Ᏽ ⫽ {i|a iT x ⬎ b i , i ⫽ 1, . . . , m} and the


feasible region is Ᏺ ⫽ {x ∈ ℜ n |a iT x ⱖ b i , i ⫽ 1, . . . , m}. A point x* is a regular
point if a Ti , i ∈ Ꮽ(x*) are mutually linearly independent vectors.

A.1 First-Order Optimality Theorem

Suppose that the objective function f (x) is continuously differentiable and that x*
is a regular local minimum point of (11). In that case, the following Karush–
Kuhn–Tucker conditions (Karush (8); Kuhn and Tucker (9)) are necessary for a
local minimum of problem (11).
1. Feasibility: x* ∈ Ᏺ
2. Complementary slackness: There exist unique multipliers λ*i ⱕ 0, i ⫽
1, . . . , m, such that
λ*i (aTi x ⫺ b i ) ⫽ 0, i ⫽ 1, . . . , m
3. Stationarity: At (x*, ␭*), the first derivative of the Lagrangian function
L(x, ␭) with respect to x satisfies
m

∇x L(x*, ␭*) ⫽ ∇f (x*) ⫹ 冱 λ* a


i⫽1
i
T
i

A.2 Sequential Quadratic Programming

Define the quadratic programming problem (QP) (iteration k)


1 T 2 k
min d ∇ f (x ) d ⫹ ∇ f(x k ) T d ⫹ f (x k )
2
(12)
s.t. a iT d ⱖ b i for i ⫽ 1, . . . , m
ORDER REPRINTS

44 Kong and Gonin

with ∇ f (xk ) and ∇ 2 f(x k ), the gradient vector and Hessian matrix of f(x), respec-
tively. For k ⫽ 1, 2, . . . , then
1. Solve problem (12) to determine d k and let ␭ k⫹1 be the vector of La-
grange multipliers associated with the constraints of problem (12).
2. Set x k⫹1 ⫽ x k ⫹ d k. Return to Step 1 with k :⫽ k ⫹ 1. Repeat until
certain convergence criteria are met.

References

1. Bailer AJ, Piegorsh WW: ‘‘Estimating Integrals Using Quadrature Methods with an Application
Downloaded by [Aston University] at 14:42 23 August 2014

in Pharmacokinetics,’’ Biometrics, 46, 1201–1211, 1990.


2. Wei GCG: ‘‘Experimental Design for Estimating Area Under the Curve by Trapezoidal Approx-
imation in Destructive Sampling,’’ Drug Information J., 31, 1237–1242, 1997.
3. Chang CT, Wong RL: ‘‘Optimal Selection Procedures for Abbreviated Area Under the Curve
(AAUC) of Blood Concentration Versus Time for Drug Blood Concentration Levels,’’ J. Bio-
pharm. Stat., 8, 407–416, 1998.
4. D’Argenio DZ: ‘‘Optimal Sampling Times for Pharmacokinetic Experiments,’’ J. Pharmacoki-
netics and Biopharmacentics, 9, 739–756, 1981.
5. Tod M, Rocchisani J-M: ‘‘Implementation of OSPOP and Algorithms for the Estimation of
Optimal Sampling Times in Pharmacokinetics by the ED, EID and API Criteria,’’ Comp. Meth-
ods and Programs in Biomed., 50, 13–22, 1996.
6. Fletcher R: Practical Methods of Optimization, 2nd Ed., John Wiley & Sons, New York, 1989.
7. Powell MJD: ‘‘VMCWD: A FORTRAN Subroutine for Constrained Optimization,’’ Report
DAMTP-1982/NA4, University of Cambridge, Cambridge, UK, 1982.
8. Karush W: ‘‘Minima of Functions of Several Variables with Inequalities as Side Conditions,’’
MS Thesis, University of Chicago, Chicago, Illinois, 1939.
9. Kuhn HW, Tucker AW: ‘‘Nonlinear Programming,’’ In: Proceedings of the Second Berkeley
Symposium on Mathematical Statistics and Probability, J Neyman, Ed., UCLA Press, Berkeley,
California, 481–492, 1951.
Request Permission or Order Reprints Instantly!

Interested in copying and sharing this article? In most cases, U.S. Copyright
Law requires that you get permission from the article’s rightsholder before
using copyrighted content.

All information and materials found in this article, including but not limited
to text, trademarks, patents, logos, graphics and images (the "Materials"), are
the copyrighted works and other forms of intellectual property of Marcel
Dekker, Inc., or its licensors. All rights not expressly granted are reserved.

Get permission to lawfully reproduce and distribute the Materials or order


Downloaded by [Aston University] at 14:42 23 August 2014

reprints quickly and painlessly. Simply click on the "Request


Permission/Reprints Here" link below and follow the instructions. Visit the
U.S. Copyright Office for information on Fair Use limitations of U.S.
copyright law. Please refer to The Association of American Publishers’
(AAP) website for guidelines on Fair Use in the Classroom.

The Materials are for your personal use only and cannot be reformatted,
reposted, resold or distributed by electronic means or otherwise without
permission from Marcel Dekker, Inc. Marcel Dekker, Inc. grants you the
limited right to display the Materials only on your personal computer or
personal wireless device, and to copy and download single copies of such
Materials provided that any copyright, trademark or other notice appearing
on such Materials is also retained by, displayed, copied or downloaded as
part of the Materials and is not removed or obscured, and provided you do
not edit, modify, alter or enhance the Materials. Please refer to our Website
User Agreement for more details.

Order now!

Reprints of this article can also be ordered at


http://www.dekker.com/servlet/product/DOI/101081BIP100101011

S-ar putea să vă placă și