a bayesian approach for estimating detection times in horses:...
TRANSCRIPT
A Bayesian approach for estimating detection times in horses: exploring the
pharmacokinetics of a urinary acepromazine metabolite
J. M. MCGREE*
G. NOBLE�
F. SCHNEIDERS�
A. J. DUNSTAN�
A. R. MCKINNEY�
R. BOSTON§ &
M. SILLENCE*
*Mathematical Sciences, Queensland
University of Technology, Brisbane, Qld,
Australia; �School of Animal and Veterinary
Sciences, Charles Sturt University, Wagga
Wagga, NSW, Australia; �Australian
Racing Forensic Laboratory, Kensington,
NSW, Australia; §School of Veterinary
Medicine, New Bolton Center, University of
Pennsylvania, Kennett Square, PA, USA
McGree, J. M., Noble, G., Schneiders, F., Dunstan, A. J., McKinney, A. R.,
Boston, R., Sillence, M. A Bayesian approach for estimating detection times in
horses: exploring the pharmacokinetics of a urinary acepromazine metabolite.
J. vet. Pharmacol. Therap. doi: 10.1111/j.1365-2885.2012.01389.x.
We describe the population pharmacokinetics of an acepromazine (ACP)
metabolite (2-(1-hydroxyethyl)promazine) (HEPS) in horses for the estimation
of likely detection times in plasma and urine. ACP (30 mg) was administered to
12 horses, and blood and urine samples were taken at frequent intervals for
chemical analysis. A Bayesian hierarchical model was fitted to describe
concentration–time data and cumulative urine amounts for HEPS. The
metabolite HEPS was modelled separately from the parent ACP as the half-life
of the parent was considerably less than that of the metabolite. The clearance
(Cl/FPM) and volume of distribution (V/FPM), scaled by the fraction of parent
converted to metabolite, were estimated as 769 L/h and 6874 L, respectively.
For a typical horse in the study, after receiving 30 mg of ACP, the upper limit of
the detection time was 35 h in plasma and 100 h in urine, assuming an
arbitrary limit of detection of 1 lg/L and a small (�0.01) probability of
detection. The model derived allowed the probability of detection to be estimated
at the population level. This analysis was conducted on data collected from only
12 horses, but we assume that this is representative of the wider population.
(Paper received 2 August 2011; accepted for publication 30 January 2012)
James M. McGree, Mathematical Sciences, Queensland University of Technology, 2
George St, Brisbane, Qld 4000, Australia. E-mail: [email protected]
INTRODUCTION
Acepromazine (ACP) is a phenothiazine tranquilizer commonly
used in horses to reduce stress during transportation and to calm
excitable horses during veterinary procedures or in training
(Smith et al., 1996). As with most therapeutic substances, ACP
is prohibited in competition by most racing organizations (Chou
et al., 2002). If a sufficiently long washout period is not observed
before a race after ACP has been administered, the trainer runs
the risk of detection and associated penalties.
Drug screening is regulated by racing authorities who are in the
process of establishing limits of detection in blood and urine for a
range of therapeutic substances, based on either the parent drugs
or their metabolites. These limits are being established in order to
protect the integrity of the sport without adversely affecting the
welfare of the horse by restricting legitimate therapeutic treat-
ment, as discussed by Toutain (2010). Therefore, there is interest
in assembling detailed pharmacokinetic data to enable the rational
setting of analytical detection limits and corresponding with-
drawal times for particular prohibited substances.
Approaches to the detection of prohibited substances in horse
racing differ from country to country, such that estimated
withholding times may vary depending upon which country the
horse is racing in. For example, urine samples are favoured for
drug detection in Australia, whereas plasma samples are
preferred in the USA. Nevertheless, both countries consider a
drug test to be positive if the parent compound or one of its
metabolites is detected on race day. Therefore, to provide precise
estimates of withholding times, an understanding of the
metabolism and the renal excretion of the drug is required. This
understanding can be facilitated through the modelling of the
pharmacokinetics (PK) of the drug, that is, what the body does to
the drug after administration (Benet, 1984). In general, such
models describe how the drug is metabolized and then eliminated
from the body over time.
In this study, we consider ACP as the parent drug and
2-(1-hydroxyethyl)promazine (HEPS) as the principle metabolite.
This leads to the consideration of a parent–metabolite model that
describes the conversion of ACP to HEPS, and the elimination of
both substances from the body via the urine. The individual (or
horse) specific PK for ACP have been studied previously (Ballard
et al., 1982; Hashem and Keller, 1993; Marroum et al., 1994).
To our knowledge, limited research has appeared in the
literature in regard to the PK of HEPS, see Schneiders et al.
(2012), who suggest that detection of HEPS could be used as an
indicator for the use of ACP. Furthermore, detection times are
J. vet. Pharmacol. Therap. doi: 10.1111/j.1365-2885.2012.01389.x
� 2012 Blackwell Publishing Ltd 1
often reported in studies of a few (3–6) horses and in absolute
terms (i.e. number of hours or days) with no probability values
attached to them. In the present study, we sought to determine
more accurate detection times by using a larger number of
horses (12) and to explore a means to judge, not just the time of
detection, but the probability of detection at a given time.
We developed an understanding of the PK and renal excretion
of HEPS at a population level through the implementation of a
Bayesian hierarchical model (BHM). Covariates were available
and included breed, body weight (kg) and age (years), and these
were explored as potential effects to minimize unexplained
between-subject variability (BSV). Also, plasma concentrations
were subject to measurement error as two measurements were
taken for each sample, and analysed separately. Given the two
approaches to measure plasma concentration from each sample
generally gave differing results, this source of variability needed
to be considered in the model. Once a parsimonious statistical
model had been developed, simulation techniques were used to
estimate the detection times in plasma and urine.
ANALYTICAL METHODS
Acepromazine and HEPS were analysed by LC-MS following the
extraction from blood plasma and urine using a solid-phase
extraction method. To prepare the samples for analysis, aliquots
were spiked with internal standards (propionylpromazine hydro-
chloride and HEPS-d4 maleate), then diluted with either acetic
acid (urine) or acetic acid plus methanol (plasma).
The samples were centrifuged to remove proteinaceous
material before the supernatant fractions were loaded onto IST
Isolute HCX solid-phase extraction cartridges that had previously
been conditioned with methanol and acetic acid. The cartridges
were washed with acetic acid and methanol, dried briefly and
then eluted with an mixture of ethyl acetate, methanol and
ammonium hydroxide. The eluates were evaporated to dryness
at 60 �C under a stream of nitrogen and reconstituted in
isopropanol and ammonium acetate (20 mM) for analysis.
The samples were analysed using a Waters Acquity UPLC
system equipped with a Waters Acquity UPLC BEH C18
analytical column and a Phenomenex Security-Guard C18
guard column interfaced to an Applied Biosystems 4000 Q-Trap
mass spectrometer Melbourne, Vic., Australia. The mobile phase
varied from 100% aqueous ammonium acetate to 40% ammo-
nium acetate/60% acetonitrile, which was run as a gradient.
The instrument was operated in positive ion atmospheric
pressure chemical ionization (APCI) mode with ion source
conditions optimized for ACP.
Calibration standards and quality control samples were
prepared by spiking pooled blank equine plasma and urine with
ACP maleate and HEPS maleate. The specificity of the method
was assessed by analysing 12 blank equine plasma and 12 blank
equine urine samples. No significant matrix interference was
observed, and the calibration curves for both ACP and HEPS in
plasma and urine were linear over the full calibration ranges.
Correlation coefficients for ACP were 0.9889 or greater and for
HEPS 0.9996 or greater in both matrices. The limit of
quantification (LOQ) for both analytes in both matrices was
1 ng/mL. The limit of detection (LOD) based on a signal-to-noise
ratio >3 was <1 ng/mL in all cases.
Accuracy and precision data were collated from the quality
control samples prepared at low and high concentrations. The
accuracy of detection (% of nominal) for ACP in plasma ranged
from 95% to 109%, with a precision (% CV) of 3.6–5.3%. In urine,
values for accuracy and precision for ACP detection were 89% to
104% and 4.4% to 7.7%, respectively. The corresponding values
for HEPS were as follows: accuracy 88–104%, precision 6.4–7.9%
(plasma); and accuracy 103–119%, precision 3.1–4.5% (urine).
DESIGN AND DATA COLLECTION
Data were collected on twelve horses (geldings) after an
intravenous administration of ACP maleate equivalent to
30 mg dose of ACP, denoted as D. Injections were given into
the jugular vein after the insertion of an indwelling catheter on
the morning of the experiment. A summary of the 12 horses is
given in Table 1. All horses were weighed the day before drug
administration and at the end of the study, and it was found that
the weight of each horse had changed slightly over the duration
of the study, despite being fed a diet that was designed to
maintain a stable body weight. If this change was linear (this is
the only change we can estimate given these data), then there
was <6% change in weight across all horses. Hence, weight as a
time-varying covariate was not considered because of such small
changes, and the average (or median) of the two measurements
was used for covariate exploration.
ACP and HEPS concentration–time data were derived from the
chemical analysis (as discussed in the previous section) of blood
and urine samples collected at frequent intervals; 10-mL blood
samples were collected into lithium heparin vacutainers at the
following time points: 0.08, 0.17, 0.33, 0.67, 1, 1.5, 2, 2.5, 3, 4,
6, 8, 12 and 24 h after administration. Urine samples were
collected approximately 2, 4, 6, 8, 12, 24, 36 and 48 h after
administration. The horses were trained to urinate to a whistle.
Table 1. Summary of horses analysed
Horse
Median
weight (kg)
Age
(years) Sex Breed
Dose
(mg/kg)
1 570 11 Gelding Thoroughbred 0.053
2 540 15 Gelding Thoroughbred 0.056
3 670 10 Gelding Thoroughbred 0.045
4 544 9 Gelding Standardbred 0.055
5 550 8 Gelding Thoroughbred 0.055
6 574 9 Gelding Standardbred 0.052
7 538 7 Gelding Standardbred 0.056
8 566 8 Gelding Standardbred 0.053
9 478 4 Gelding Standardbred 0.063
10 552 5 Gelding Standardbred 0.054
11 474 7 Gelding Standardbred 0.063
12 560 9 Gelding Standardbred 0.054
2 J.M. McGree et al.
� 2012 Blackwell Publishing Ltd
The urine was collected into a 2-L plastic container that was
attached to each horse. Each horse also wore a specially designed
harness known as a ‘horse nappy’ between collection times to
ensure that all urine was collected. The midpoint between
collection times was used as the average urine amount. Hence,
the average urine amount was modelled over each of the time
spans. Parent ACP was not detected in urine and hence could
not be modelled. All samples were refrigerated directly after
collection and stored at )20 �C until analysed.
MODEL
Realistic statistical models for PK data are often complex systems
incorporating auxiliary variables in hierarchies or strata. Bayes-
ian statistical methodology provides powerful methods for
modelling such data, combining prior knowledge with informa-
tion from the data to yield a posterior distribution from which all
inference about unknowns can be made. Unfortunately, sampling
from this posterior distribution can be difficult in practice as no
closed-form solution may be available. This led to the develop-
ment of a collection of proven computational methods used
across a wide area of applied sciences and technology. The most
widely used computational method is Markov chain Monte Carlo
(MCMC), a technique that can be used to sample from a target
distribution (here, a posterior) that is difficult to sample from
directly. This facilitates posterior inference in complex settings.
A full-conditional hierarchical Bayesian analysis was under-
taken using WinBUGS (Lunn et al., 2000) where MCMC tech-
niques are used to sample from the posterior distribution of
estimable parameters. Reviews of Bayesian modelling of phar-
macokinetic data are given by Duffull et al. (2005) and Lunn
et al. (2002), and an example is given by Dansirikul et al. (2005).
In our analysis, two MCMC chains were run simultaneously,
each with different initial parameter estimates for each param-
eter. Convergence of the MCMC chains to the stationary
distributions was assessed in two ways. Firstly, the Gelman
and Rubin convergence diagnostic (Kass et al., 1998) was
calculated for each parameter. This diagnostic essentially
compares the between-chain variability with the within-chain
variability. If this variability is similar, the value of the diagnostic
will be close to 1. In our analysis, if values were <1.1 for all
parameters, then it was assumed that the stationary distribution
had been reached under this diagnostic. As a further check, the
trace history for each parameter was plotted to ensure that the
‘fuzzy caterpillar’ was observed (Lunn et al., 1999).
The two MCMC chains were run for 20 000 samples with the
first 10 000 being discarded as a ‘burn-in’. If the above
convergence criteria were met, the two chains were pooled to
provide samples from the stationary/posterior distribution. It is
this distribution upon which all inferences were drawn.
Parent and metabolite model
The full data set consisted of ACP and HEPS plasma concentra-
tions and urine amounts over time. Therefore, a multivariate
response model needed to be derived. A full parent–metabolite
model based on mass balance is given in Fig. 1.
Figure 1 shows a five-compartment parent–metabolite model.
Initially, the parent (P) enters into the first of two compart-
ments, where it can move between this compartment (P1) and
another (P2). A two-compartment model was proposed for this
part of the model as this was fitted in the analysis of ACP given
by Ballard et al. (1982), Hashem and Keller (1993) and
Marroum et al. (1994). A fraction of the parent is either
converted to the metabolite (FPM), excreted renally (FEP) or
eliminated nonrenally (1 ) FPM ) FEP). In the second case, the
triangle represents the bladder (P3), and the parent will
continue to accumulate here until the horse empties the
bladder. For the fraction of the parent that is converted to the
metabolite (M), a fraction of M is either excreted renally (FEM) or
nonrenally (1 ) FEM). In this first case, the metabolite accumu-
lates in the bladder (M2, represented by a triangle) until the
horse empties the bladder. The ks denote rate constants for
movement between compartments.
A simplification of this parent–metabolite model was made.
Given the half-life of the parent was considerably shorter than
that of the metabolite, we can conclude that the kinetics of the
metabolite were not being driven by the parent. In such cases, it
should be sufficient to model the metabolite alone, particularly
given we are only interested in estimating detection times for the
metabolite; notably, this is a key assumption in our approach,
and further research would be required to determine whether
this biased our analysis.
Metabolite model
Figure 2 shows the proposed model for plasma concentration
and cumulative urine amount for the metabolite only. Initially,
ACP appears in compartment P1 and is converted to HEPS that
appears in M1 (the molecular weights of ACP and HEPS are 326
and 345, respectively). Here, M1 represents the amount of HEPS
in plasma, with a fraction being excreted renally (FE) to the
bladder (M2) and a fraction excreted nonrenally (1 ) FE). The
metabolite will then accumulate in M2 until the horse empties
their bladder.
Fig. 1. Parent–metabolite model for acepromazine (P) and 2-(1-hy-
droxyethyl)promazine (M) in horses.
A Bayesian approach for estimating detection times in horses 3
� 2012 Blackwell Publishing Ltd
From Fig. 2, the following is true
k22 ¼FECl
V
k23 ¼ð1� FEÞCl
V
FE ¼R1
0 k22M1dt
D;
where ClR ¼ FECl and ClNR ¼ (1)FE)Cl are renal clearance and
nonrenal clearance, respectively; Cl ¼ ClR + ClNR and V is the
volume of distribution.
As we only model HEPS, estimates of clearance and volume of
distribution are scaled by the fraction of the parent converted to
the metabolite, that is, Cl/FPM and V/FPM. To facilitate the
estimation of the parameters in the model, the assumption is that
FPM ¼ 1. We note, however, Ballard et al. (1982) show that
approximately 50% of injected ACP binds to red blood cells. Of
this 50%, an estimated fraction of 0.99 (±0.01) is converted to
HEPS, see Schneiders et al. (2012). Even though our parame-
terization may not reflect the true biological process, it does
allow an accurate mathematical representation of the data, see
Duffull et al. (2000).
The ordinary differential equations (ODEs) that describe the
kinetics of this model for the amount of the metabolite are as
follows. We note that these are linear ODEs and hence can be
solved analytically.
dP1
dt¼� k12P1
dM1
dt¼k12P1 � k22M1 � k23M1
dM2
dt¼k22M1;
where P1(0) ¼ D,M1(0) ¼ 0,M2(0) ¼ 0.
The above ODEs provide the structural model for the analysis.
When solved, they provide predicted responses for each of the
three compartments. For modelling the metabolite data, we are
only interested in the solution to compartments M1 and M2.
Denote the predicted responses in each of these compartments as
f1(.) and f2(.), respectively.
The following describes the full probabilistic model. In model
development, we consider a number of alternative models in an
attempt to derive a parsimonious description the multivariate
data. Instead of describing all models considered, we describe a
‘full’ model such that all other rival models are parametrically
nested within this full model.
The hierarchical structure of the probabilistic model can be
described in the following stages:
Stage 1.
yijkljkij; cijkl � Nðfijkl; cijklÞ; ð1Þ
where kij ¼ (Clij,FEij,Vij)¢, fijkl ¼ fl(kij,k12,tijkl,D) represents the
model prediction for the kth response on the jth measurement
type on the ith individual in compartment Ml, and the variance
of this response is given by cijkl ¼ r2add;l þ r2
prop;lf ðkij; tijkl;DÞ2 for
i ¼ 1,…,N (the total number of horses), j ¼ 1,2 (the two sources
for the chemical analysis), k ¼ 1,…,nijl (the total number of
observations from Ml on individual i on measurement type j) and
l ¼ 1,2.
Stage 2.
log kijj log hi;U � MVNðlog hi;UÞ; ð2Þ
where hi ¼ (Cli,FEi,Vi)¢ represents the ith individual’s mean PK
behaviour and U is the measurement error variance–covariance
matrix of PK parameters.
Stage 3.
log hij log l; Zi;X � MVNðlog l;XÞ; ð3Þ
where Zi represents the covariate values for the ith individual, Xthe between-horse variability and l ¼ (Cl,FE,V)¢ the population
PK parameters.
Stage 4 (priors).
r2add;l �Unið0;100Þ
r2prop;l �Unið0;100Þ
logðCl;VÞ0 �MVNð0;1000� IÞk12 �HNð0;1000Þ
log FE �Unið�300;0ÞU �Wð0:001� I;5ÞX �Wð0:001� I;5Þ;
where l ¼ 1,2, and Uni(), MVN, HN and W denote the uniform,
multivariate normal, half-normal and Wishart distributions,
respectively. These priors were chosen to be essentially uninfor-
mative, meaning that they should have minimal influence in
determining the parameter estimates when compared to the
observed data. It should be noted that, when estimating FE, it is
possible [through BSV and measurement error variability (MEV)]
that this estimate may be larger than 1. This is impossible in
Fig. 2. A simplified pharmacokinetic model for metabolite (2-(1-hy-
droxyethyl)promazine) after an intravenous administration of acepro-
mazine to horses.
4 J.M. McGree et al.
� 2012 Blackwell Publishing Ltd
practice, so the constraint min{1,FE} was imposed in the
estimation. As can be seen, k12 was assumed to be a fixed-
effects parameter, so no BSV (or MEV) was estimated.
RESULTS
The observed (log) concentration–time data and accumulated
urine amounts of HEPS for the 12 horses are shown in Fig. 3.
The demographic data collected for the 12 horses were breed,
weight (kg) and age (years). Of the 12 horses, four were
thoroughbreds and eight were standardbreds. The mean (± SD)
body weight of all horses was 544 (±48) kg, and the mean (±
SD) age was 8 (±2.8) years.
Structural and residual error model
The model specified in Fig. 2 was considered with additive,
proportional and a mixture of additive and proportional residual
errors on each of the two responses. These model fits can be
compared by considering deviance, calculated as minus twice
the log-likelihood. Given we typically wish to select parameter
estimates that maximize the log-likelihood, we would like to
select the model with the smallest deviance. However, as more
parameters are included, the model will generally fit the data
‘better’ and will therefore result in a smaller deviance (even
though this decrease may not be statistically significant).
Therefore, we would ideally like to select the model that has
small deviance and a small number of parameters.
The results of model fits are summarized in Table 2. The
smallest median deviance can be seen for the model that has a
proportional residual error term on plasma concentrations and
an additive residual error on the cumulative urine amounts.
Further, this model has less than or equal the number of
parameters to estimate when compared with other rival models.
Therefore, this residual error structure was assumed for further
analysis.
Measurement variability
Plasma concentrations for each horse were chemically analysed
in duplicate at each time point. We therefore explored whether
significant measurement error existed in the data. Measurement
variability was accounted for by including a random effect on
each individual parameter (i.e. Cli, FEi and Vi). Table 2 shows
the median deviance values for models fitted with and without
MEV. As can be seen, there was a reduction in the median
deviance when MEV was included. However, it is unclear
whether this is significant given the extra number of estimable
parameters introduced into the model. Hence, the deviance
information criterion (DIC) was used to compare the two models
(Spiegelhalter et al., 2002). The information criterion is as
follows.
DIC ¼ pD þ �D;
where pD ¼ �D� Dð�hÞ is the effective number of parameters in
the model, �D ¼ Eh½DðhÞ� and �h is the expectation of h.
From the above formula, the DIC penalizes large values of both�D and pD such that the model with the smallest DIC value is
preferred. For the model with MEV, �D ¼ 1978:26,
Dð�hÞ ¼ 1915:280, pD ¼ 62.981 and DIC ¼ 2041.240, com-
pared with �D ¼ 2048:94, Dð�hÞ ¼ 2012:48 , pD ¼ 36.457 and
DIC ¼ 2085.390 for the model without MEV. The results show
that the model with MEV gave the smallest DIC value, and
hence, this model was retained for further investigation.
Covariate modelling
Covariate relationships were initially investigated through the
use of scatter plots against individual parameters. If any of the
relationships warranted further investigation, then models with
these covariate relationships were fitted. Limited data were
available, so results from this part of the analysis should be
treated with caution.
0 5 10 15 20 25–1.5
–1
–0.5
0
0.5
1
1.5
2
Time (h)
log
Con
cent
ratio
n (µ
g/L)
HEPS plasma
0 10 20 30 40 500
1000
2000
3000
4000
Time (h)
Cum
ulat
ive
amou
nt (µ
g)
HEPS urine
Fig. 3. Observed plasma concentrations and
urine amounts for the acepromazine (ACP)
metabolite HEPS after intravenous dose of
30 mg of ACP to 12 horses.
Table 2. Median deviance and 95% credible interval of deviance
Model
Res.
plasma
Res.
urine
Median
deviance
95% CI for
deviance
Fig. 2 Add. Add. 2115 2100, 2137
Fig. 2 Prop. Prop. 2212 2197, 2233
Fig. 2 Add. Prop. 2312 2296, 2334
Fig. 2 Prop. Add. 2049 2034, 2069
Fig. 2 Mix. Mix. 2120 2104, 2148
Fig. 2 + MEV Prop. Add. 1977 1951, 2007
Fig. 2 + MEV + wt)V Prop. Add. 1977 1951, 2008
A Bayesian approach for estimating detection times in horses 5
� 2012 Blackwell Publishing Ltd
The covariate plots can be seen for HEPS in Fig. 4. There did
not appear to be a relationship between any covariates and the
estimates of individual parameters except for log Vi and median
horse weight (Fig. 4c). That is, it appeared that as the horse
weight increased, log Vi also increased. The plot also suggests
that the relationship (if it exists) could be linear. As such, the
400 500 600 7006
6.5
7
7.5(a)
Median weight (kg)
Log
Cl i
400 500 600 7000
0.05
0.1
0.15
0.2(b)
Median weight (kg)
FEi
400 500 600 7008
8.5
9
9.5(c)
Median weight (kg)
Log
Vi
4 6 8 10 12 14 166
6.5
7
7.5(d)
Age (years)
Log
Cl i
4 6 8 10 12 14 160
0.05
0.1
0.15
0.2(e)
Age (years)FE
i
4 6 8 10 12 14 168
8.5
9
9.5(f)
Age (years)
Log
Vi
1 26
6.5
7
7.5(g)
Breed
Log
Cl i
1 20
0.05
0.1
0.15
0.2(h)
Breed
FEi
1 28
8.5
9
9.5(i)
Breed
Log
Vi
Fig. 4. Plots of posterior means of individual parameter estimates vs. covariates.
0 5 10 15 20 25
0
2Horse 1
Time (h)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
log
Con
c (µ
g/L)
0 5 10 15 20 25−1
0
1
2Horse 2
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 3
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 4
Time (h)
0 5 10 15 20 25−1
0
1
2Horse 5
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 6
Time (h)0 5 10 15 20 25
−1
0
1
2Horse 7
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 8
Time (h)
0 5 10 15 20 25−2
−1
0
1
2Horse 9
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 10
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 11
Time (h)0 5 10 15 20 25
−2
−1
0
1
2Horse 12
Time (h)
Fig. 5. Individual fitted values for plasma concentrations of HEPS. The dashed and solid lines (-, - -) represent the duplicate measurements on the same
blood sample which the dots represent the observed data points.
6 J.M. McGree et al.
� 2012 Blackwell Publishing Ltd
(median) weight of each horse was included in the model as
follows:
log bi ¼ b1 þ b2
wi
548
� �;
where wi is the weight of the ith horse and 548 is the median
weight over all horses.
The results from the fit where weight was included as a
covariate on log V are given in Table 2. As can be seen, there
was no reduction in the median deviance of the model with
weight, compared with the model without weight. However, this
is not the full picture as it is possible for a covariate to be
significant without there being a significant reduction in
deviance. Therefore, we explored other indicators such as
whether the BSV on log V had been reduced and whether the
credible interval for b2 (from above) included zero or not.
The posterior distribution of b2 was examined, and it was
found that zero was a highly probable value. Further, the BSV of
log V without weight in the model was 0.0729, while with
weight, this reduced to 0.0709, reduction of about 3%. Given
this, it was concluded that weight was not significant, and
therefore, it was not included into the final model.
Model fit
The individual fits to the data can be seen for plasma
concentrations in Fig. 5 and cumulative urine amounts in
Fig. 6. The individual fits for plasma concentrations are repre-
sented by two lines corresponding to duplicate measurements
taken on each blood sample as the MEV was found to be
significant. In some horses, this is difficult to see, but is
particularly visible for horses 4 and 5. All individual specific
models seemed to fit the data well. The same was true for the
individual cumulative urine fits.
A summary of parameter estimates (posterior medians) for the
final model is given in Table 3. The large estimate of k12 shows a
rapid rate of conversion from ACP to HEPS. The clearance rate of
HEPS was 769 L/h with a volume of distribution of 6874 L (both
scaled by the fraction of parent converted to metabolite). The BSV
of Cl, FE and V is small, at around 30%. No estimate for the BSV of
k12 is given as this was considered as a fixed effect in the model.
Model checking and validation
Figure 7 shows the residual error plots for the final model. The
standardized residuals are given for both responses. For plasma
Table 3. Summary of parameter estimates (posterior medians) for final
model
Parameter
Cl/FPM
(L/h) FE
V/FPM
(L) k12 rprop,1 radd,2
Estimate 769 0.0856 6874 35.87 0.16 109
BSV (%) 0.09 (30) 0.11 (32) 0.07 (27) – – –
0 20 400
1000
2000
3000Horse 1
Time (h) Time (h) Time (h) Time (h)
Time (h) Time (h) Time (h) Time (h)
Time (h) Time (h) Time (h) Time (h)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
Cum
. am
ount
(µg)
0 20 400
2000
4000Horse 2
0 20 400
1000
2000
3000Horse 3
0 20 400
1000
2000
3000Horse 4
0 20 400
2000
4000Horse 5
0 20 400
500
1000
1500Horse 6
0 20 400
2000
4000
6000Horse 7
0 20 400
1000
2000Horse 8
0 20 400
1000
2000
3000Horse 9
0 20 400
1000
2000
3000Horse 10
0 20 400
1000
2000
3000Horse 11
0 20 400
1000
2000Horse 12
Fig. 6. Individual fitted values for cumulative urine amounts of HEPS. The solid lines represent the predicted values, and the dots represent the observed
data points.
A Bayesian approach for estimating detection times in horses 7
� 2012 Blackwell Publishing Ltd
concentrations, all plots appear reasonable given the assumption
of normally distributed residuals. The QQ plot shows discrepancy
between the standardized residuals and the fitted normal density
at the lower tail, but this was not deemed significant. Similarly,
the residual plots for cumulative urine amounts generally do not
violate the assumption for normality. Again, discrepancies
appear in the QQ plot, but at both tails.
Posterior predictive checks were used to determine whether the
model developed was consistent with the data. The checks were
performed in the following way. After fitting the final model, each
parameter has a posterior distribution. This distribution summa-
rizes the uncertainty about the estimate (given the prior and the
observed data). Ten thousand random samples were drawn from
the posterior distribution, and the concentration (lg/L) at time
t hours was simulated (based on each of the 10 000 random
samples). This produced a distribution of concentrations (lg/L) at
each time point considered. The 5th, 50th and 95th percentiles of
these distributions can be seen in Fig. 8, with the observed data as
dots. The two extreme percentiles reflect the uncertainty about the
estimated concentrations.
The plots show that some of the observations fall outside the
percentiles. This is to be expected, as in theory approximately
10% of the data points should fall outside the limits given by the
90% credible interval. For the plasma concentrations, the
uncertainty below the median seems to increase with time. This
is not surprising given that less data were available at later time
points as the concentrations of HEPS in some samples fell below
the LOD. Further, the plot is on the log-scale, so these actually
relate to very low concentrations.
For the urine data, the upper bound of the 90% credible intervals
extends noticeably higher than the observed data. This suggests
that BSV and/or MEV may be inflating estimates of FE leading to a
larger than expected amount of HEPS being excreted renally.
Thus, detection times may be longer than expected. Overall, both
posterior predictive checks seem consistent with the observed data.
Withholding and detection times
The detection time for a therapeutic substance represents the
period after administration that the laboratories can identify the
substance or its metabolite in either blood or urine. The
withholding time is usually judged by a veterinarian and
represents the period between drug administration and a race,
taking into account the published detection time and a number
of other factors.
Toutain (2010) used Monte Carlo methods to estimate
withholding times by extrapolating the detection times published
by the European Horserace Scientific Liaison Committee. The
methodology is based on the terminal phase or single exponential
model for elimination and relies on estimating a plasma-to-urine
0 10 20 30–2
–1
0
1
2(a)
Time (h)
log
Con
cent
ratio
n (µ
g/L)
0 10 20 30 40 500
1000
2000
3000
4000
5000(b)
Time (h)
Cum
ulat
ive
amou
nt (µ
g)
Fig. 8. Posterior predictive check for plasma
concentrations and urine amounts of the
acepromazine metabolite HEPS in 12 horses.
The dots represent actual observations. The
solid line represents the model, and the dashed
lines represent a 90% credible interval for the
model predictions.
0 2 4–4
–2
0
2
4Residuals vs. predicted
–2 0 2–4
–2
0
2
4QQ−plot
–4 –2 0 2 40
20
40
60Histogram of residuals
–1 0 1 2–1
0
1
2
Observed. vs. predicted
0 2000 4000–4
–2
0
2
4Residuals vs. predicted
–2 0 2–5
0
5
10QQ−plot
–5 0 50
20
40
60
80Histogram of residuals
0 2000 40000
1000
2000
3000
4000
Observed. vs. predicted
Fig. 7. Residual plots for HEPS [top row refers to plasma concentration (lg/L), and bottom row refers to cumulative urine amount (lg)]
including standardized residuals vs. predicted values, normal QQ plot of the standardized residuals, histogram of the standardized residuals and
observed vs. predicted values.
8 J.M. McGree et al.
� 2012 Blackwell Publishing Ltd
concentration ratio parameter. An important feature of this
research was the identification of the most influential parameters
through a sensitivity analysis. This revealed that the manner in
which the body processed the compound determines the
detection time rather than, for example, dose.
Given we have undertaken a full compartmental analysis
under a Bayesian framework, we provide an alternative
approach for the estimation of withholding times. This approach
will allow withholding times to be estimated from both plasma
and urine samples and allows for the practical consideration that
a horse will empty their bladder.
Probability of detection in plasma. Suppose we are interested in
determining how likely it is for a horse to have a HEPS
concentration in plasma greater than, say, the LOD of 1 lg/L at
a certain time t after being given an intravenous bolus dose of
30 mg of ACP. Given the population model developed in this
research, this can be resolved by considering the simulated data
from the posterior predictive check. In the posterior predictive
check, random samples are drawn from the posterior distribution
of parameters and data generated for some time points (t). Thus,
the probability of a horse having a HEPS concentration greater
than, say, the LOD at time t can be determined from the
simulated data. For a given time t, the number of simulated
observations that are greater than the LOD is simply counted,
then divided by the total number of simulated observations at
that particular time (t). This provides an estimate for the
probability that a concentration will be larger than the LOD. A
plot of these estimated probabilities can be seen in Fig. 9.
Figure 9 shows three plots. Figure 9a shows the estimated
probabilities of detection based on the population estimates of
parameters and therefore represents estimates for a typical horse
in the sample. Alternatively, Fig. 9b shows the estimated
probabilities of detection for the individual horses in the study.
The second plot shows how the probability of detection varies
among individual horses. This highlights the potential for
extreme detection probabilities for specific horses and the
sensitivity of the estimates to the horses studied. The last plot
is an empirical estimate of a survivor type function. The plot
shows the estimates (with 95% confidence bounds) of the
probability that a detection time is larger than some time t. For
example, it shows that at 30 h after ACP administration, while
10% of the horses would be expected to have a concentration of
HEPS in plasma greater than the LOD, in practice this percentage
may be as small as 0% or as large as 25%.
Probability of detection in urine. Assume we are interested in
determining how likely it is for a horse to have a HEPS
concentration in urine greater than, say, the LOD of 1 lg/L at a
certain time (t) after being given an intravenous bolus dose of
30 mg of ACP. The model for HEPS given in section ‘Model’
predicts the accumulated urine amount of HEPS at a given time
t. In order to determine detection times, an estimate of urine
volume or accumulated urine volume at time t needs to be made.
Figure 10 shows the observed cumulative urine volumes for
all horses against time. A strong linear relationship is evident
between cumulated volume and time, with the variability in the
response increasing with time. Therefore, a linear mixed effects
model with proportional residual error was fitted to the
cumulative urine volume with time as the independent variable.
Further details are given in Appendix A.
0 10 20 30 40 500
0.2
0.4
0.6
0.8
1
Time (h)
Pro
babi
lity
of d
etec
tion
(a)
0 10 20 30 40 500
0.2
0.4
0.6
0.8
1
Time (h)
Pro
babi
lity
of d
etec
tion
(b)
15 20 25 300
0.2
0.4
0.6
0.8
1
Time (h)
Pro
porti
on
(c)
Fig. 9. Probability plots for the detection of the acepromazine (ACP) metabolite HEPS in plasma after the intravenous dose of 30 mg of ACP; (a)
population estimate of the probability of detection, (b) individual horse estimate of the probability of detection and (c) estimated survivor type function
for detection times.
0 10 20 30 40 50 600
0.5
1
1.5
2
2.5x 104
Time
Cum
. vol
ume
Fig. 10. Observed cumulative urine volumes against time.
A Bayesian approach for estimating detection times in horses 9
� 2012 Blackwell Publishing Ltd
It is now possible to predict the cumulative amount of HEPS
and cumulative urine volume for a given time t. Given that
detection in urine is based on the concentration, these cumu-
lative predictions need to be combined and converted to
concentrations. One solution is to consider, say, 4-h windows
from the time of drug administration, and the amount of HEPS
and urine volume accumulated in these windows can be used to
predict concentration. This approach assumes the following.
• The sample is representative of the population.
• The model developed describes the population.
• A linear relationship exists between urinary excretion rate of
HEPS and the production rate of urine.
• The urinary production and excretion rate on race day (and/or
days before race day) are the same as production and
excretion rates during the study.
• The bladder is completely emptied at the beginning of each 4-h
interval.
• Urine is collected at the end of a window.
Data were simulated from the cumulative urine amount and
volume models, and the probability of detection was estimated in
the same way as plasma (described above). Results can be seen in
Fig. 11.
Again, Fig. 11 shows three plots. Figure 11a shows the
estimated probability of detection in urine given a 4-h window
0 50 100 1500
0.1
0.2
0.3
0.4
0.5
Time (h)
Pro
babi
lity
of d
etec
tion
(a)
0 50 100 1500
0.1
0.2
0.3
0.4
0.5
Time (h)
Pro
babi
lity
of d
etec
tion
(b)
45 50 55 60 65 700
0.2
0.4
0.6
0.8
1
Time (h)
Pro
porti
on
(c)
Fig. 11. Probability plots for the detection of the acepromazine (ACP) metabolite HEPS in urine after the intravenous dose of 30 mg of ACP based on
4-h windows; (a) population estimate of the probability of detection, (b) individual horse estimate of the probability of detection and (c) estimated
survivor type function for detection times.
0 10 20 30 40 500
0.5
1
1.5
2 x 104
Time
Cum
. v
Horse 1
0 10 20 30 40 500
0.5
1
1.5
2
2.5
Time
Cum
. v
Horse 2
0 10 20 30 40 500
5000
10000
15000
Time
Cum
. v
Horse 3
0 10 20 30 40 500
0.5
1
2
1.5
Time
Cum
. vHorse 4
0 10 20 30 40 500
0.5
1
1.5
2
Time
Cum
. v
Horse 5
0 10 20 30 40 500
5000
10000
15000
Time
Cum
. v
Horse 6
0 10 20 30 40 500
0.5
1
1.5
2
Time
Cum
. v
Horse 7
0 10 20 30 40 500
0.5
1
1.5
2
Time
Cum
. v
Horse 8
0 10 20 30 40 500
0.5
1
1.5
2
Time
Cum
. v
Horse 9
0 20 40 600
0.5
1
1.5
2
Time
Cum
. v
Horse 10
0 10 20 30 40 500
0.5
1
1.5
2
Time
Cum
. v
Horse 11
0 10 20 30 40 500
0.5
1
1.5
2
Time
Cum
. v
Horse 12
x 104 x 104
x 104 x 104
x 104
x 104
x 104 x 104 x 104
Fig. A1. Individual fits for cumulative urine volume model.
10 J.M. McGree et al.
� 2012 Blackwell Publishing Ltd
for a typical horse in the sample. The plot shows that there is
quite a small probability of detection past 50 h, for example. This
agrees well with the observed and predicted data shown in
Fig. 8b as it appears that very little HEPS is accumulated in urine
past this point in time. Figure 11b shows the estimated
probabilities of detection based on individual estimates of PK
parameters. These estimates relate to the individual horses in the
study. The final plot shows an empirical estimate of a survivor
type function. Detection times longer than 70 h appear rare
given our best estimate. However, when calculating appropriate
withholding times, the uncertainty around this estimate should
also be considered. Thus, while only 10% of horses might be
expected to breech the detection level at 65 h, this estimate
could be between 0% and 25% for an individual horse.
CONCLUSION
A BHM was considered for the description of the metabolism of
the parent–metabolite process for ACP. Given the half-life of ACP
is much smaller than that of the metabolite, the model was
reducible to the consideration of the metabolite only. The
structural form of this model for the description of metabolite
concentration in plasma and cumulative amount in urine was
based on mass balance with vague priors chosen for the
estimable parameters. No covariates were found to be statisti-
cally significant. This may be attributed to the small sample of
horses measured and/or relatively narrow ranges of potential
covariates, particularly for weight. Measurement variability on
plasma concentrations proved to be influential, and random
effects were included in the model to account for this.
Simulation techniques were used to show agreement between
predicted and observed data and in the estimation of detection
times for a typical horse in the sample. Detection in urine was
complicated by the need to allow the horses to empty their
bladder at various times after ACP had been administered. This
was resolved by considering urine concentration in 4-h windows
(postdose). Care should be taken when interpreting these results
as the probability profiles are likely to change if different sized
windows are considered. Nevertheless, the results shown should
provide guidance for trainers and veterinarians to estimate
appropriate withholding times to ensure that horses are not
racing with prohibited substances in their system.
All inferences in this study were based on a sample of twelve
horses. In making decisions in the model-building phase and
providing estimates of, for example, clearance and withholding
times, we assume that this sample is representative of the entire
population. Indeed, in the selection of covariates, weight seemed
somewhat linearly related to the volume of distribution. This
relationship did not prove to be statistically significant, but this
may have been due to the small sample size. Hence, care should
be taken when drawing any inferences from this work.
ACKNOWLEDGEMENT
We are very grateful to Prof. S.B. Duffull of the University of
Otago for his input into the paper. We would also like to thank
the two referees for their comments and suggestions.
REFERENCES
Ballard, S., Shults, T., Kownacki, A.A., Blake, J.W. & Tobin, T. (1982)
The pharmacokinetics, pharmacological responses and behavioral
effects of acepromazine in the horse. Journal of Veterinary Pharmacology,
5, 21–31.
0 0.5 1 1.5 2x 104
–5
0
5Residuals vs. predicted
–4 –2 0 2 4–4
–2
0
2
4
6QQ−plot
–4 –2 0 2 4 60
10
20
30
40Histogram residuals
6 7 8 9 106
7
8
9
10Obs. vs. pred
Fig. A2. Residual plots for cumulative urine volume model.
0 10 20 30 40 50 600
0.5
1
1.5
2
2.5
3x 104
Time (h)
Cum
. V. (
mL)
Fig. A3. Posterior predictive check for cumulative urine volume model
where ‘Cum. V’ represents the cumulative amount of urine collected from
the 12 horses over the duration of the study.
A Bayesian approach for estimating detection times in horses 11
� 2012 Blackwell Publishing Ltd
Benet, L.Z. (1984) Pharmacokinetics: Basic principles and its use as a tool
in drug metabolism. In Drug Metabolism and Drug Toxicity. Eds Mitchell,
J.R. & Horning, M.G., p. 199. Raven Press, New York.
Chou, C.C., Chen, C.L., Rice, B.L. & Colahan, P.T. (2002) Reduced resi-
dent time and pharmacodynamic effects of acepromazine after sub-
clinical multiple dosage in exercised thoroughbreds. Journal of
Veterinary Pharmacology, 25, 379–382.
Dansirikul, C., Morris, R.G., Tett, S.E. & Duffull, S.B. (2005) A Bayesian
approach for population pharmacokinetic modelling of sirolimus.
British Journal of Clinical Pharmacology, 62, 420–434.
Duffull, S., Chabaud, S., Nony, P., Laveille, C., Girard, P. & Aarons, L.
(2000) A pharmacokinetic simulation model for ivabradine in healthy
volunteers. European Journal of Pharmaceutical Sciences, 10, 285–294.
Duffull, S.B., Kirkpatrick, C.M.J., Green, B. & Holford, N.H.G. (2005)
Analysis of population pharmacokinetic data using NONMEM and
WinBUGS. Journal of Biopharmaceutical Statistics, 15, 53–73.
Hashem, A. & Keller, H. (1993) Disposition, bioavailability and clinical
efficacy of orally administered acepromazine in the horse. Journal of
Veterinary Pharmacology and Therapeutics, 16, 359–368.
Kass, R.E., Carlin, B.P., Gelman, A. & Neal, R.M. (1998) Markov chain
Monte Carlo in practice: a roundtable discussion. Journal of the Amer-
ican Statistical Association, 52, 93–100.
Lunn, D.J., Wakefield, J., Thomas, A., Best, N.G. & Spiegelhalter, D.J.
(1999) PKBUGS User Guide 1.1. Imperial College, London.
Lunn, D., Thomas, A., Best, N. & Spiegelhalter, D. (2000) WinBUGS – a
Bayesian modelling framework: concepts, structure, and extensibility.
Statistics and Computing, 10, 325–337.
Lunn, D.J., Best, N., Thomas, A., Wakefield, J. & Spiegelhalter, D.J. (2002)
Bayesian analysis of population PK/PD models: general concepts and
software. Journal of Pharmacokinetics and Pharmacodynamics, 29,
271–307.
Marroum, P.J., Webb, A.I., Aeschbacher, G. & Curry, H. (1994) Phar-
macokinetics and pharmacodynamics of acepromazine in horses.
American Journal of Veterinary Research, 55, 1428–1433.
Schneiders, F., Noble, G., Boston, R., Dunstan, A., Sillence, M. &
McKinney, A. (2012) Acepromazine pharmacokinetics: a forensic
perspective. The Veterinary Journal in press.
Smith, B.L., Jones, J.H., Hornof, W.J., Miles, J.A., Longworth, J.E. &
Willits, N.H. (1996) Effects of road transport on indices of stress in
horses. Equine Veterinary Journal, 28, 446–454.
Spiegelhalter, D.J., Best, N.G., Carlin, B.P. & Van der Linde, A.
(2002) Bayesian measures of model complexity and fit (with
discussion). Journal of the Royal Statistical Society Series B, 64,
583–616.
Toutain, P.L. (2010) How to extrapolate a withdrawal time from an
EHSLC published detection time: a Monte Carlo simulation appraisal.
Equine Veterinary Journal, 42, 248–254.
APPENDIX: MODEL FOR CUMULATIVE URINE VOLUME
The individual fits can be seen in Fig. A1. From the plot, the
proposed model seems to fit the observed data well for all horses.
The residual plots and posterior predictive check can be seen
in Figs A2 & A3, respectively. The posterior predictive check
also shows the 90% credible interval for the predicted response.
The QQ plot, histogram and observed vs. predicted plots show
that the residuals generally follow a normal distribution. There
is a slight pattern in the residual vs. predicted plot, but this
was not deemed strong enough to violate our assumption
about the residuals. The posterior predictive check shows
agreement between the simulated and observed data and
shows that some observations lie outside the 90% credible
interval.
12 J.M. McGree et al.
� 2012 Blackwell Publishing Ltd