A Comparison of the Prediction Accuracy of Two

IVIVC Modelling Techniques

CLARE GAYNOR,1 ADRIAN DUNNE,1 JOHN DAVIS2

1
UCD School of Mathematical Sciences, University College Dublin, Belfield, Dublin 4, Ireland
2
Clinical Pharmacology, Pfizer Global Research and Development, Sandwich, England

Received 30 April 2007; revised 20 July 2007; accepted 3 September 2007

Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/jps.21220

ABSTRACT: The goal when developing an in vitro–in vivo correlation (IVIVC) model is
the ability to accurately predict the in vivo plasma concentration profile of a drug
formulation using only its in vitro dissolution data. The prediction accuracy of any model
depends on the reliability of the method used to develop it. Some statistical concerns
regarding methods based on deconvolution have been highlighted and a convolution
based technique has been proposed as an alternative. This comparison shows, by means
of a simulation study, that the modelling approach which uses convolution produces far
more accurate results, accurately predicting the observed plasma concentration–time
profile and, therefore, comfortably meeting the FDA validation criteria. The fact that the
model developed using the deconvolution based technique fails to describe the simulated
data and thus fails the FDA validation test when it ought to pass should be of great
concern to those currently implementing this method. ß 2007 Wiley-Liss, Inc. and the
American Pharmacists Association J Pharm Sci 97:3422–3432, 2008
Keywords: in vitro–in vivo correlations (IVIVC); mathematical model; deconvolution;
convolution

INTRODUCTION importance and the methods used to establish

Establishing an in vitro–in vivo correlation
(IVIVC) model has become an integral part of
the drug development process. The ability to
accurately predict an in vivo response from
in vitro observations has far-reaching conse-
quences in several areas including quality control
and formulation development. Such a model can
also act as a surrogate for human bioequivalence
studies and be used in the setting of in vitro
dissolution specifications.1 Given that these
IVIVC models are in such widespread use, their
accuracy of prediction and reliability are of key
Correspondence to: Adrian Dunne
(E-mail: adrian.dunne@ucd.ie)
Journal of Pharmaceutical Sciences, Vol. 97, 3422–3432 (2008)
ß 2007 Wiley-Liss, Inc. and the American Pharmacists Association
them should be carefully scrutinised.
The procedure routinely employed when
attempting to establish an IVIVC for an extended
release (ER) drug product involves the study of
a number of formulations of the dosage form of
interest, each with a different release rate. In
order to develop the IVIVC model, three kinds of
data (in vitro, in vivo ER and in vivo reference) are
frequently collected. A sample of the ER dosage
units of interest from each formulation are
dissolved in vitro and the fraction which has
dissolved is recorded at a series of time points for
each dosage unit. Further ER dosage units from
each formulation are administered to a number of
human subjects, usually as part of a crossover
study, and their plasma drug concentrations are
repeatedly measured for a predetermined period.
A unit reference dose, which may take the form
of a solution, IV injection or dosage unit which

3422 JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008

 PREDICTION ACCURACY OF TWO IVIVC MODELLING TECHNIQUES
instantly dissolves, is administered to each of
the subjects and, once again, their plasma drug
concentrations are measured over time. The
resulting data are used to develop a model which
can be employed to predict the in vivo plasma drug
concentration profile using in vitro dissolution
data only.
There are a number of levels of IVIVC1 but, for
the purposes of this study, only the ‘level A
correlation’, which predicts the entire in vivo
plasma drug concentration profile from the
in vitro dissolution data, will be examined. The
less informative correlations (levels B, C and
multiple level C) which relate summary para-
meters of the in vitro and in vivo profiles are not
included in this study.
In order to establish an IVIVC model’s validity,
its accuracy of prediction must be evaluated; ‘a
correlation should predict in vivo performance
accurately and consistently’.1 The United States
Food and Drug Administration (FDA) recommend
that formulations with three or more release rates
be used to develop the model. Assessment of the
model’s ability to describe these data is referred to
as internal validation. The model must be able to
predict the area under the plasma concentration
curve (AUC) and the peak plasma concentration
(Cmax) to within acceptable limits as set out by the
FDA.1 Data collected using dosage units from a
fourth or subsequent formulation which was not
used in the model development stage may also be
used to similarly assess external predictibility.
Many methods of developing IVIVC models
have been proposed which, generally, fall into
two categories: those based on deconvolution and
those based on convolution. A number of areas of
concern regarding the deconvolution based meth-
ods have been highlighted2 and it has become
clear that, in principle, the use of these techniques
is not advisable. A non-linear mixed effects
modelling approach, which is not subject to the
same criticisms, has been developed,3 this con-
volution based procedure should, in theory,
produce more accurate results. In spite of this,
the deconvolution based methods are routinely
used. There are perhaps many reasons for this
practice, among which may be the fact that the
extent of any empirical difference between the two
approaches has never been demonstrated. Since
methods involving deconvolution are regularly
used, it is of vital importance that their theoretical
drawbacks be carefully examined and any prac-
tical significance substantiated. The aim of this
research is to investigate the performance of both
DOI 10.1002/jps
3423
a deconvolution and a convolution based method
and quantify any difference in accuracy of
prediction by means of a simulation study.
DECONVOLUTION BASED APPROACH
A fundamental element of this method involves
relating the fraction of a dosage unit which has
dissolved in vitro to the fraction dissolved in vivo
at corresponding times. The fraction which has
dissolved at each observation time in vitro is
recorded as part of the dissolution study. The
fraction dissolved in vivo, however, is not directly
observable and must be derived from the available
data using the technique of deconvolution. There
are many methods of deconvolution4,5 which rely
on the same principles and assume linearity and
time invariance of the system being studied. This
group of methods is based on the convolution
integral, which is defined as
Zt
Cp ðtÞ ¼ cd ðt  tÞRðtÞdt (1)
0
where Cp(t) is the plasma drug concentration at
time t following administration of an ER dosage
unit, cd(t) the unit impulse response and R(t) the
in vivo dissolution rate. Deconvolution involves
estimating the rate at which a drug dissolves
in vivo using the observed in vivo ER data (which
contains information on dissolution, absorption,
distribution, metabolism and excretion of the
drug) and observed in vivo reference data (which
does not involve dissolution). This process may be
thought of as a kind of ‘subtraction’ technique
used to isolate the information of interest, that is,
the rate at which the dosage unit dissolves in vivo.
This rate is used to calculate the fraction dissolved
in vivo at each of the observed time points. In
order to establish an IVIVC, the relationship
between the fraction dissolved in vivo and that
dissolved in vitro must be examined. These
fractions are plotted against one another at
common times and an appropriate line or curve
is fit to the data points. This is the core of the
IVIVC model and can be used to predict the
fraction of the dosage unit dissolved in vivo using
the in vitro data. In order to predict plasma drug
concentration, a further step is necessary. The
predicted in vivo fractions dissolved must be ‘re-
convolved’ with estimates of the pharmacokinetic
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
3424 GAYNOR, DUNNE, AND DAVIS
parameters obtained from the reference data to
produce plasma concentration profiles.
Some Statistical Concerns
There are many theoretical arguments against
using deconvolution based methods which are
described in detail elsewhere.2 Descriptions of
those weaknesses which are believed to have the
most significant effect on the model’s ability to
predict with accuracy are given below, and are
also summarised in Table 1.
Very often, the observed data are averaged
across subjects or dosage units at each time point
before analysis. This leads to a significant loss of
information particularly if the between-subject
(or dosage unit) variation is high. There may also
be a marked difference between the curve which is
being analysed (made up of data averaged at each
time point) and the original individual profiles.
Deconvolution can, of course, be carried out at the
individual level6–8 but this is not as frequently
seen as analysis based on averages. In order to
ensure comparable conditions, the methods of
analysis being compared in this study were both
carried out at the individual subject level.
When the in vivo dissolution rate has been
converted to give the fraction dissolved in vivo at a
given time, these fractions are plotted against
the in vitro fractions dissolved at the same time.
Clearly only in vitro and in vivo data collected at
common times can be used. This leads to the
discarding of data where the observation times
are not coincident. The resulting loss of informa-
Table 1. A Summary of Statistical Concerns with Deconvolution Based Methods
Statistical Concern
Only data at common times are used
A straight line can be used to model in vitro–in vivo
relationship when it is inappropriate
Assumes in vitro–in vivo relationship is
time invariant
Assumptions of ordinary least squares violated
Averaged data are used
Use of deconvolution
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
tion has a negative impact on the model’s
predictions.
Once these fractions are plotted against one
another, an appropriate line or curve is fit to the
data. There is an abundance of instances docu-
mented in the literature where straight-line
models are fit to data where a curve would be
more appropriate.9–11 This can lead to inaccurate
(either over or under) predictions. It is apparent
that care must be taken when choosing the model
which describes the relationship between the
in vitro and in vivo fractions dissolved.
The process of plotting in vivo fractions dissolved
at a particular time against their in vitro counter-
parts is based on the assumption that their
relationship is time invariant. This procedure
forces the assumption that the link between
in vitro and in vivo dissolution does not change
over time; this is a somewhat ingenuous approach.
In effect, the medium in which the dosage unit
dissolves in vitro does not change as time passes,
but it may significantly alter in vivo as the dosage
unit’s location in the gastrointestinal tract, and
consequently the pH, changes. Of course this is
not necessarily the case, there are many solid
systems that are inert and whose dissolution is not
influenced by their location in the gastrointestinal
tract. It may, however, be unwise to assume that
this is always true and that the resulting IVIVC
relationship is time-invariant.
Generally, the method of ordinary least squares
is used to fit the chosen model but two of its
main assumptions are violated in this case. The
first is that the independent variable (usually
in vitro fraction dissolved) is measured without
error, which, in terms of this application, is an
Effect
Loss of information leads to inefficiency
Produces biased predictions
Model is somewhat naı̈ve and unrealistic
Error in independent variable leads to biased predictions
Correlation exists between observations made on a single
tablet or subject which leads to reduced efficiency
Cannot identify inter-subject or inter-dosage
unit differences
Curve of averages is different to the average of the
individual curves
Deconvolution is inherently unreliable
DOI 10.1002/jps
 PREDICTION ACCURACY OF TWO IVIVC MODELLING TECHNIQUES
unattainable expectation and is known as the
Errors in Variables issue.12 Secondly, it is
required that all observations be uncorrelated
or independent while it is clear that measure-
ments made on the same subject or dosage unit
cannot be uncorrelated. Violating these two
underlying assumptions of the least squares
method results in a model which can produce
inaccurate predictions.
A fundamental shortcoming of this entire group
of methods is that the process of deconvolution
is itself unstable13 and may be thought of as a
‘roughening’ of the data. It, therefore, by nature,
introduces errors and reduces the accuracy with
which the model can predict the plasma concen-
tration profiles.
There are a number of ways of evaluating
accuracy of prediction, two of the most commonly
used measures are bias and efficiency. A biased
estimate will tend to consistently over or under
predict the quantity of interest, for example, AUC
or Cmax. A low bias is therefore usually preferable.
An inefficient estimate will produce predictions
that are widely scattered around the true value.
The combination of the issues mentioned above
is believed to have a substantial impact on the
accuracy with which the deconvolution method
can predict the plasma concentration curve, and it
is expected that this method will produce biased
and inefficient predictions.
CONVOLUTION BASED APPROACH
An alternative modelling approach which incor-
porates the convolution integral (Eq. 1) has been
developed. This method, like the deconvolution
based method described above, assumes the
linearity and time invariance of the system being
studied.4 The version of this convolution based
method proposed by O’Hara et al.3 first fits a
standard compartmental model to the unit
impulse response data in order to estimate each
subject’s pharmacokinetic parameters separately.
The in vivo plasma concentrations resulting from
administration of the ER dosage unit to each
subject, and the in vitro fractions dissolved for
each dosage unit are modelled directly (i.e.
without averaging) in a one-step process. The
resulting IVIVC model does not suffer from the
same theoretical flaws as the deconvolution based
method previously described. It allows the rela-
tionship between the fraction dissolved in vitro
and that dissolved in vivo to be described as a
DOI 10.1002/jps
3425
function of time and incorporates random effects
which can account for some of the correlation
between repeated measurements on the same
subject or dosage unit. Observations in vitro and
in vivo do not have to be made at corresponding
times, eliminating the loss of information caused
by discarding some of the data. The use of this
method removes the need for a deconvolution step
and predicts plasma concentrations rather than
their deconvolved equivalents.
METHOD—SIMULATION STUDY
A simulation study to compare these two appro-
aches to establishing an IVIVC was undertaken.
Four batches of a drug, each with a different
release rate, were simulated. The first three
batches will be referred to as the fast, medium
and slow formulations while the fourth was
included for the purposes of external validation.
The data were simulated such that a marked
relationship between in vitro and in vivo dissolu-
tion would exist. The three kinds of data: in vitro,
in vivo ER and reference were produced for 12 ER
dosage units and 12 subjects with 22 observations
made from time of administration (0 h) up to 34 h
in order to simulate a study with a crossover
design.
The reference data were simulated according to
a one compartment pharmacokinetic model with
first order absorption as follows:
 
l3k
cdk ðtÞ ¼ ðel2k t  el3k t Þ (2)
Vk ðl3k  l2k Þ
where Vk, l2k and l3k represent the kth subject’s
volume of distribution and rate constants of drug
elimination and absorption, respectively. The
plasma drug concentration for the kth subject at
time t following administration of a unit reference
dose is given by
Y3k ðtÞ ¼ cdk ðtÞ þ "3k ðtÞ (3)
where e3k is approximately normally distributed
ðNð0; s 22 ÞÞ with s 22 accounting for in vivo measure-
ment error.
The observed in vitro fraction of drug dissolved
from the ith dosage unit at time t is described as
follows:
Yi1 ðtÞ ¼ f1 Fi1 ðtÞ þ "i1 ðtÞ; "i1 ðtÞ  Nð0; s 1 2 Þ (4)
where Fi1(t) is the true fraction dissolved at time t,
s 21 represents the in vitro measurement error and
the parameter f1 accounts for any difference
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
3426 GAYNOR, DUNNE, AND DAVIS

Table 2. Values of Parameters Used to Simulate Dataset

Description of Parameter Value

Tablet to tablet standard deviation (v1) 0.034
Number of tablets in vitro 12
Number of batches 4
Standard deviation of in vitro error term (s1) 0.0005
Number of time points in vitro 22
Dissolution rate constant (l1) 0.799 (fast formulation)
0.299 (medium formulation)
0.100 (slow formulation)
0.200 (external formulation)
Scale parameter in vitro (f1) 0.996
Weibull parameter (a) 1
Subject to subject standard deviation (v2) 0.61
Number of subjects 12
Standard deviation of in vivo error term (s2) 0.62
Number of time points in vivo 22
Scale parameter in vivo (f2) 0.996
Elimination rate constant for typical subject (l2) 0.2
Absorption rate constant for typical subject (l3) 0.9
Standard deviation of l2k 0.04
Standard deviation of l3k 0.18
Theta1 (u1) 1.29
Theta2 (u2) 0.900
ER dose 100
IR dose 50
Volume of distribution of typical subject 1
Standard deviation of V 0.2

between actual dose and label claim. Variation in
Fi1(t) between dosage units was modelled as:
logitðFi1 ðtÞÞ ¼ logitðF1 ðtÞÞ þ ui ; ui  Nð0; v1 2 Þ (5)
The term ui is a random effect which allows
for differences between dosage units while also
accounting for the correlation between repeated
observations made on the same dosage unit. The
inter dosage unit variation is given by v21 . F1(t) is
the fraction of the typical dosage unit (where ui is
zero) dissolved at time t and is modelled according
to the Weibull function as follows:
F1 ðtÞ ¼ 1  expðl1 ta Þ
where the rate of dissolution of the drug is
determined by l1 and a. The use of the logit
transformation logit(x) ¼ log(x/(1x)) in equation
five ensures that the simulated fraction, even with
the addition of a random effect (ui), remains
within the interval [0,1].
The in vivo plasma concentration measured
from the kth subject at time t following admin-
istration of the ith dosage unit is described as
Z
0
Yi2k ðtÞ ¼ Dose f2 cdk ðt  tÞFi2k ðtÞdt þ ei2k ðtÞ;
ei2k ðtÞ  Nð0; s 2 2 Þ ð7Þ
where Fi2k0 (t) is the in vivo dissolution rate at time
t, f2 allows for a difference in bioavailability
between the reference dose and the ER dose and
s 22 represents in vivo measurement error as
before.
The error terms ei1, ei2k and e3k are all modelled
according to truncated normal distributions to
(6)
ensure that the values of Yi1, Yi2k and Y3k remain
positive.
The relationship between in vitro and in vivo
dissolution is defined as follows:
logitðFi2k ðtÞÞ ¼ logitðF1 ðtÞÞ þ u1 þ u2 t þ ui þ sik ;
with ui  Nð0; v1 2 Þ; and sik  Nð0; v2 2 Þ ð8Þ

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008 DOI 10.1002/jps
 PREDICTION ACCURACY OF TWO IVIVC MODELLING TECHNIQUES 3427

where sik is a subject specific random effect which RESULTS

allows correlation between observations made on
a particular subject and accounts for differences While all analysis was carried out at the
between subjects so v22 gives the subject-to-subject individual subject (or dosage unit) level, for ease
variation and v21 represents inter dosage unit of presentation the in vitro dissolution data and
variability as before. The use of the logit link as in vivo ER plasma concentrations were averaged
seen above ensures that both Fi2k(t) and Fi1(t) are at each time point for each formulation and are
constrained to be between zero and one. The u2t shown in Figure 1a and b, respectively.
term allows for the possibility that the relation- The in vivo reference data were first analysed.
ship between in vitro and in vivo dissolution Each subject’s elimination and absorption rate
may vary with time as might be expected in some constants and volume of distribution were esti-
circumstances. mated separately. These parameter estimates
In order to ensure that the results of this were used to predict the immediate release
simulation could be meaningfully interpreted, the profiles which are shown in Figure 2 and were
values of all the above parameters were based on used as part of both methods of analysis.
those estimated using real datasets and are listed
in Table 2.
A dataset was simulated according to the
model described above and analysed twice: first
using a deconvolution based method and then an
alternative, convolution based technique. The
fast, medium and slow formulations were used
to develop the model in both cases, while the
fourth batch was excluded to be used later for
the purposes of external validation. First, each
subject’s reference data was analysed separately
using the one compartment pharmacokinetic
model provided in the ADVAN2 subroutine of
the NONMEM software developed by Beal and
Sheiner.14 The resulting PK parameters were
then used in both the deconvolution and con-
volution methods as required. The convolution
method used was a modified version of that
reported by O’Hara et al.,3 which fit the model
described by Eqs. (4–8) above by means of a
custom written PRED subroutine for NONMEM.
The method of deconvolution used was Con-
strained Deconvolution (CoDe) as proposed by
Hovorka et al.5,15 The required software was
kindly provided by Dr. Hovorka. The two IVIVC
models established were used to predict in vivo
plasma concentration profiles for each subject
and all four formulations as both methods of
analysis were carried out at the individual
subject level. The Cmax and AUC were computed
for each of the four batches using the in vivo ER
data and the profiles predicted by each method of
analysis. The AUC was calculated using the
trapezoidal rule, while the Cmax for each subject
was located by inspection and results were Figure 1. (a) Simulated in vitro dissolution data
averaged over all 12 subjects. The average averaged across dosage units for each of the four for-
values of AUC and Cmax obtained using each mulations. (b) Simulated in vivo plasma concentration
method were compared to the true values to data averaged across subjects for each of the four for-
calculate percentage prediction error (%PE). mulations.

DOI 10.1002/jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
3428 GAYNOR, DUNNE, AND DAVIS

Figure 2. In vivo reference data and fitted curves for each subject.

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008 DOI 10.1002/jps
 PREDICTION ACCURACY OF TWO IVIVC MODELLING TECHNIQUES
The deconvolution based method was first to
be implemented. The fraction dissolved in vivo
for each subject at each observation time was
calculated and averaged across subjects to obtain
an in vivo dissolution profile for each formulation
as shown in Figure 3. These in vivo fractions were
plotted against the corresponding in vitro value
and the method of Ordinary Least Squares was
used to fit a polynomial curve of degree four to the
data as illustrated in Figure 4. This fitted curve
was used to predict the fraction dissolved in vivo
which corresponded with the fraction dissolved
in vitro at each time point. These in vivo fractions
were then reconvolved with each subject’s PK
parameters to produce predicted plasma concen-
trations for each subject and each formulation.
The convolution-based method was used to
develop the IVIVC model in a one-step process,
analysing the in vitro and in vivo data simulta-
neously. Predicted plasma concentration curves
were produced for each subject directly.
To illustrate the differences between the two
modelling techniques, Figure 5 shows simulated
and predicted plasma concentration curves for each
of the four batches averaged over all 12 subjects.
The solid line represents predictions made when
using the convolution method, while the dashed
line shows those of the deconvolution method.
The corresponding percentage prediction errors
for each formulation resulting from analysis using
each method are shown in Tables 3a and 3b.
Figure 3. Average in vivo fractions dissolved esti-
mated using the deconvolution method for each of the
three formulations included in developing the model.
The fast formulation is represented by the þ sign, the
medium by  and the slow formulation by the * symbol.
DOI 10.1002/jps
3429
Figure 4. Average in vivo fraction dissolved against
average in vitro fraction dissolved at common times.
The fast formulation is represented by the þ sign, the
medium by  and the slow formulation by the symbol.
The fitted polynomial curve is given by the solid line.
CONCLUSIONS AND DISCUSSION
The summary in Tables 3a and 3b shows that
analysis using the deconvolution based method
yields inaccurate predictions, while those of the
convolution based technique are very precise.
The United States FDA require, that for an
IVIVC model to be validated, average absolute
percentage prediction error of both AUC and Cmax
must be 15% or less for each batch (and less
than 10% on average over the three batches) to
establish internal predictibability and 10% or less
for external predictability.1 For this sample
dataset, where an IVIVC relationship does exist,
the convolution method produces an IVIVC
model that comfortably meets the FDA criteria
while the deconvolution based method fails to
establish the in vitro–in vivo relationship.
There is a noticeable difference in the accuracy
with which the deconvolution-based method pre-
dicts the four different formulations. The reasons
for this can easily be seen by examining the core of
the IVIVC model which is illustrated in Figure 4.
The majority of points representing the fastest
formulation are in the upper right hand corner
of the plot and are quite close to the fitted curve.
The early time points, however, lie below the fitted
curve. This leads to overprediction of plasma
concentrations at earlier times and quite a good fit
at later times as can be seen in Figure 5a. The
batch with a medium rate of dissolution has points
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
3430 GAYNOR, DUNNE, AND DAVIS

Figure 5. Observed and predicted plasma concentration profiles averaged across

subjects for each formulation. Panels (a–d) represent the fast, medium, slow and
external formulations, respectively. In each panel, the symbols denote the simulated
data, the solid line shows the convolution method’s predictions and the dashed line
represents the deconvolution method’s results.

which lie quite close to the fitted curve in Figure 4, can see the resulting underprediction of in vivo
resulting in good predictions in Figure 5b. Most of plasma concentration values in Figure 5c at early
the points in Figure 4 which belong to the slowest time points and, similarly, concentrations at later
formulation are above the fitted curve and one times are overpredicted. This lack of accuracy in

Table 3a. Percentage Prediction Errors (Used for Internal Validation)

Batch Method of Analysis %PE AUC %PE Cmax

Fast Deconvolution 0.4692 6.1409
Convolution 0.2427 0.0320
Medium Deconvolution 0.1988 2.2250
Convolution 0.0380 0.3745
Slow Deconvolution 0.5670 28.0391
Convolution 0.0791 0.0324
Average j%PEj Deconvolution 0.4116 12.135
(first three batches) Convolution 0.1199 0.1463

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008 DOI 10.1002/jps
 PREDICTION ACCURACY OF TWO IVIVC MODELLING TECHNIQUES
Table 3b. Percentage Prediction Errors (Used for
External Validation)
Method of
Batch Analysis %PE AUC %PE Cmax
External Deconvolution 0.2730 11.0301
Convolution 0.0025 0.0232
prediction leads to a very high %PE value for
the Cmax of this formulation, resulting in the
failure to satisfy the internal validation criteria.
The formulation used for external validation
has a release rate which lies between the medium
and slow formulations as seen in Figure 1a,
predictions made using deconvolution for this
formulation are not sufficiently accurate to meet
the FDA external validation criteria.
The fact that the convolution method is, in
effect, fitting the correct model leads to results
that are, perhaps, over optimistic. The degree to
which the predictions would deteriorate when
using real data would depend on the extent to
which the model used is mis-specified.
The deconvolution method not only fails to
produce a model which meets the FDA validation
criteria but, as can be seen in Figure 5 (particu-
larly in panel c), the predicted plasma concentra-
tion curve is entirely the wrong shape. This is
potentially very misleading if the IVIVC model is
to be used during product development.
The deconvolution method and software used in
this study (CoDe15) has been chosen from a large
number of such techniques. The CoDe method was
selected on the basis of its good performance in
testing such as that conducted by Madden et al.5
There are also many variations possible at the
implementation stage of both deconvolution and
the convolution-based methods so a number of
further decisions were necessary. The choice of
whether or not to average the data before analysis
is believed to be a crucial one. Current advice on
best practise16 suggests that deconvolution should
take place on an individual subject (or dosage
unit) level. Consequently, all of the analysis
conducted as part of this study was carried out
entirely at the individual subject level for both the
deconvolution and convolution-based methods.
Averaging of the raw data is believed to be so
fundamental an issue that it warrants further
exploration and, as such, this point will form the
basis of a separate study. The use of a one-stage
rather than a two-stage approach to fitting the
in vitro and in vivo ER data in the convolution
DOI 10.1002/jps
3431
based method may also have an impact on the
results. This is a question which has previously
been studied, albeit not in this context. Zhang
et al.17 point out that ‘conditioning on some of the
data or on prior fits to them may yield inaccurate
standard errors as there is not a proper account-
ing for the uncertainty in conditioning quantities’.
In general, it is agreed that simultaneous analysis
of data is, statistically, preferable to first estimat-
ing parameters, which are then fixed and used in
the second stage of the analysis.17,18
The findings of this study confirm the expecta-
tion that the convolution based method would
outperform its counterpart. The concerns about
the deconvolution method are supported by the
results and its lack of accuracy in prediction is a
result of the combined influence of its theoretical
flaws. While these may have seemed like purely
statistical considerations, they have a very strik-
ing practical effect in terms of quality of predic-
tions and, therefore, ability to meet the FDA
validation criteria. These results substantiate the
statistical theory, illustrating and quantifying the
superiority of the convolution based approach and
providing support for its use in preference to the
deconvolution-based group of methods.
REFERENCES
1. Food and Drug Administration. 1997. Guidance for
industry: Extended release oral dosage forms:
Development, evaluation, and application of in
vitro/in vivo correlations.
2. Dunne A, Gaynor C, Davis J. 2005. Deconvolution
based approach for level A in vivo-in vitro correla-
tion modelling: Statistical considerations. Clin Res
Regul Aff 22:1–14.
3. O’Hara T, Hayes S, Davis J, Devane J, Smart T,
Dunne A. 2001. In vivo–in vitro correlation (IVIVC)
modeling incorporating a convolution step. J Phar-
macokinet Pharmacodyn 28:277–298.
4. Dunne A. 2007. Approaches to developing in vitro-
in vivo correlation models. In: Chilukuri DM,
Sunkara G, Young D, editors. Pharmaceutical pro-
duct development in vitro-in vivo correlation.
pp 47–70.
5. Madden FN, Godfrey KR, Chappell MJ, Hovorka R,
Bates RA. 1996. A comparison of six deconvolution
techniques. J Pharmacokinet Biopharm 24:283–
299.
6. Abuzarur-Aloul R, Gjellan K, Sjölund M, Löfqvist
M, Graffner C. 1997. Critical dissolution tests
of oral systems based on statistically designed
experiments. I. Screening of critical fluids and
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
3432 GAYNOR, DUNNE, AND DAVIS
in vitro/in vivo modelling of extended release coated
spheres. Drug Dev Ind Pharm 23:749–760.
7. Abuzarur-Aloul R, Gjellan K, Sjolund M, Graffner
C. 1998. Critical dissolution tests of oral systems
based on statistically designed experiments. II. In
vitro optimization of screened variables on ER-
coated spheres for the establishment of an in
vitro/in vivo correlation. Drug Dev Ind Pharm
24:203–212.
8. Abuzarur-Aloul R, Gjellan K, Sjolund M, Graffner
C. 1998. Critical dissolution tests of oral systems
based on statistically designed experiments. III. In
vitro/in vivo correlation for multiple-unit capsules
of paracetamol based on PLS modeling. Drug Dev
Ind Pharm 24:371–383.
9. Liu Y, Schwartz JB, Schnaare RL, Sugita ET. 2003.
A multi-mechanistic drug release approach in a
bead dosage form and in vitro/in vivo correlations.
Pharm Dev Technol 8:409–417.
10. Cardot JM, Beyssac E. 1993. In vitro/in vivo corre-
lations: Scientific implications and standardisation.
Eur J Drug Metab Pharmacokinet 18:113–120.
11. Kortejarvi H, Mikkola J, Bacjman M, Antila S,
Marvola M. 2002. Development of level A, B
and C in vitro-in vivo correlations for modified-
release levosimendan capsules. Int J Pharm 241:
87–95.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008
12. Snedecor GW, Cochran WG. 1980. Statistical meth-
ods, 7th edition. Ames, Iowa: Iowa State University
Press.
13. Wing GM. 1991. A primer on integral equations of
the first kind—The problem of deconvolution and
unfolding. Philadelphia: Society for Industrial and
Applied Mathematics.
14. Beal SL, Sheiner LB. 1992. NONMEM user’s
guides, NONMEM Project Group. San Francisco:
University of California.
15. Hovorka R, Chappell MJ, Godfrey KR, Madden FN,
Rouse MK, Soons PA. 1998. CODE: A deconvolution
program implementing a regularization method of
deconvolution constrained to non-negative values.
Description and pilot evaluation. Biopharm Drug
Dispos 19:39–53.
16. Farrell C, Hayes S. 2007. IVIVC for oral drug
delivery: Immediate and extended release dosage
forms. In: Chilukuri DM, Sunkara G, Young D,
editors. Pharmaceutical product development in
vitro-in vivo correlation. pp 47–70.
17. Zhang L, Beal S, Sheiner L. 2003. Simultaneous vs.
sequential analysis for population PK/PD data I:
Best-case performance. Journal of Pharmacoki-
netics and Pharmacodynamics 30:387–403.
18. Verbeke G, Molenberghs G. 2000. Linear mixed
models for longitudinal data. New York: Springer.
DOI 10.1002/jps