Homogeneity in Pharmaceutical Mixing Processes
Homogeneity in Pharmaceutical Mixing Processes
Homogeneity in Pharmaceutical Mixing Processes
Biosystems Data Analysis, Swammerdam Institute for Life Sciences, University of Amsterdam,
Nieuwe Achtergracht 166, 1018 WV Amsterdam, The Netherlands
b Pharmaceutical Site Maaloev, Novo Nordisk A/S, Novo Nordisk Park, 2760 Maaloev, Denmark
c N.V. Organon, Molenstraat 10, P.O. Box 20, 5340 BH Oss, The Netherlands
Received 16 June 2005; received in revised form 6 October 2005; accepted 6 October 2005
Abstract
Determination of homogeneous mixing of the active pharmaceutical ingredient (API) is an important in-process control within the manufacturing
of solid dosage forms. In this paper two new near-infrared (NIR) based methods were presented; a qualitative and a quantitative method. Both
methods are based on the calculation of net analyte signal (NAS) models which were very easy to develop, specific with respect to the API
and required no additional reference analysis. Using a well-mixed batch as a golden standard batch, control charts were developed and used
for monitoring the homogeneity of other batches with NIR. The methods were fast, easy to use, non-destructive and provided statistical tests of
homogeneity. A mixing study was characterized with the two methods and the methods were validated by comparison with traditional HPLC
analysis.
2005 Elsevier B.V. All rights reserved.
Keywords: Net analyte signal; Process analytical technology; Powder homogeneity; Near-infrared spectroscopy; Process control charts
1. Introduction
An important unit operation in the production of a pharmaceutical solid dosage form is the mixing of the powder constituents. One of the primary objectives of the mixing process
is to obtain a homogeneous mixing of the active pharmaceutical ingredient (API) in the formulation. The homogeneity is
normally determined by removing a small number of powder
samples from the powder mixer with a probe thief and sending the samples to a quality control laboratory distant from
the process line for, e.g. high pressure liquid chromatography (HPLC) analysis. The API content in the samples and
the relative standard deviation (R.S.D.) of the API content
between the samples is then obtained and compared to specified values in order to decide wherever the batch is sufficiently
homogeneous [1]. This procedure is slow and is not suited
for process control of the mixing process to obtain the highest possible homogeneity. Also a limited number of samples
are used which makes a full characterization of homogeneity
difficult.
Therefore, near-infrared (NIR) based spectroscopic methods
which are non-destructive, fast and require little resources have
been proposed [210]. NIR has been applied in many different
manners. Samples were removed from the system and measured
with NIR (off-line) [2,4]. This method faces some of the classical
powder sampling problems where large variance is introduced
when inserting a sample thief into the powder bed and removing
a sample. In other papers, diffuse reflectance probes have been
inserted directly into the powder bed at a fixed position (in-line)
[5,6] or the mixer was fitted with quartz windows through which
monitoring could be applied from outside the mixer (semi1 noninvasive) [7]. These types of methods only monitor the powder
1
Corresponding author.
E-mail address: westerhuis@science.uva.nl (J.A. Westerhuis).
The application was not truly in-line because the mixer had to be stopped
to perform monitoring with a diffuse reflectance probe that was moved around
from window to window.
0731-7085/$ see front matter 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.jpba.2005.10.009
PBA-5540;
No. of Pages 10
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
2. Theory
In this theory section we will first discuss the mixing of pharmaceutical powders and the idea of a random and homogeneous
mixture. Then the different sources of variation in analytical
measurements of homogeneity are discussed. The net analyte
signal approach used to relate a multivariate NIR measurement
to the API content is discussed shortly. Afterwards we first
introduce the qualitative approach for monitoring homogeneity. Finally we will also show how a quantitative model can be
developed in which actual concentrations of the API are monitored.
Mixing is the treatment of two or more components in such a way, that the individual particles
of the different components in the mixture are
evenly distributed and lie adjacent to each other
within the highest possible probability [11].
(1)
2
analysis
,
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
2.4. Notation
The following notation is used during the presentation of
equations. Boldface capital characters denote matrices, boldface lower-case characters denote vectors and lower-case italic
characters denote scalars, superscript T denotes the transposed matrix or vector and the superscript + denotes the
MoorePenrose generalised inverse of a matrix. The matrix IJ
is the J J identity matrix.
2.5. Calculation of the NAS value of a spectrum
Using a set of NAS vectors, a NAS regression vector can be
calculated. With the NAS regression vector the net analyte signal
value or simply the net analyte signal of a sample spectrum can
be computed. The NAS is directly proportional to the analyte
concentration [13] which makes this value a potential candidate
to be used to evaluate the distribution of analyte in a set of
samples.
To compute the NAS regression vector, first the interference space needs to be defined. Pure component spectra of the
interferents can be used to span the interferent space. In our
experience [1416] it is better to use spectra of blank samples,
i.e. mixture samples of the interfering constituents only.
First the anti-projector (AR ) is constructed using a set of blank
spectra (Rk )
AR = (IJ Rk R+
k )
(2)
(3)
(5)
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
F(,Nnew 1,Ngb 1)
(6)
(7)
NASnew NASgb
1
s Nnew
+ N1gb
(8)
2
2
(Nnew 1)stotal
NAS,new + (Ngb 1)stotal NAS,gb
Nnew + Ngb 2
be used for the t-test to detect whether the operator put too much
or too little API in the mixer (compared to the golden batch).
The control limits for the mean chart are then:
, Nnew + Ngb 2
(9)
control limits = 0 t
2
The critical step in this method is in Step 2. It is stated that a
batch is mixed to homogeneity (golden batch) and then measured extensively with NIR. One approach to identify the point
of homogeneity could be to make a temporary NAS model. First
the anti-projector is calculated from the placebo batch spectra.
Carry out a few minutes of mixing of the calibration batch and
measure a few NIR spectra. These spectra are used to calculate a
temporary NAS regression vector. Then continue mixing of the
calibration batch. At different time points the mixer is stopped
and high number (e.g. 50) NIR spectra are recorded. The NAS
values of the spectra are calculated with the temporary NAS
regression vector. By simply plotting the variance over time the
point of homogeneity can be identified as time point with lowest
variance.
2.8. Development of a quantitative model based on NAS
values
When a final formulation has been chosen during the R&D
development phase more calibration work can be added and a
quantitative model between the NIR spectra and the API concentration developed. A major benefit of a quantitative model
is that numbers are now expressed in concentration of the API
which is comparable to the standard regulatory methodologies.
Also now the mean can be compared to the target content of API
instead of mean of golden batch.
The development of the quantitative method is described by
the following steps:
Step 1: First a placebo batch is prepared and measured with the
NIR probe. The spectra are used to calculate the anti-projector
(Eq. (2)).
Step 2: A set of calibration batches spanning a fair API concentration range are prepared, e.g. 70, 85, 100, 115 and 130%
of the concentration in the final drug product. The batches are
mixed until they are homogeneous and measured extensively
with the NIR probe. The spectra are split into a calibration set
and a test set.
Step 3: The calibration spectra from each batch are averaged
to one calibration spectrum for each calibration batch. Each
calibration spectrum is assigned a reference value which is the
average API concentration in the particular calibration batch.
This value is known from the preparation of the calibration
batch [8].
Step 4: A few spectra from one of the calibration batches are
used as model spectra (Rmod ) and the NAS regression vector
is calculated (Eqs. (3) and (4)).
Step 5: Now the NAS value of the average calibration spectra
and some extra placebo spectra is calculated (Eq. (5)). These
extra placebo spectra should not have been used to construct
the anti-projector in Step 1.
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
(10)
where c 1 is the average concentration of N1 predictions from
calibration batch 1.
Step 9: A variance and a mean chart is developed and used
for monitoring of future batches.
Table 1
Mixture composition
Component
API
Microcrystalline cellulose
Tablettose
Magnesium stearate
Talc
Sum
g
100
200
677.5
7.5
15
1000
% (w/w)
10
20
67.75
0.75
1.50
100
F(,Nnew 1,Ngb 1)
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
Table 2
Batch overview
Name
Use
Placebo batch
70, 85, 100, 115 and 130%
calibration batches
Mixing batch
No API is added
Varying amount of API from 7 to 13% (w/w). Fifty NIR spectra/batch.
Thirty samples for HPLC analysis/batch
A batch with 10% (w/w) API. Fifty spectra/time point. Several time
points measured during mixing
Each was with 10% (w/w) API concentration. The batch differences
were mixing time and mixing order. Thirty NIR spectra/batch. Twelve
samples for HPLC analysis/batch
Fig. 1. Mixing experiment data plotted in variance and mean chart using qualitative method. The mean values symbolized with () indicate that the assumption
for the t-test is not fulfilled because the variance is different from the golden
batch.
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
Fig. 2. The five average calibration spectra and the NAS regression vector (bolded line) (left). The API had a characteristic peak near 6000 cm1 . This feature was
easily recognized in the calibration spectra and the NAS regression vector (right).
very easy, fast and precise method for making the quantitative
calibration models.
The NAS values of the average calibration spectra and 20
extra placebo spectra were plotted against their reference values
(symbolized with circles in Fig. 3) and a straight line was fitted to
the points in a least square sense (line in Fig. 3). The correlation
coefficient was 0.9998. The NAS values of test set spectra were
also plotted in order to demonstrate the variance of the API
within each batch (symbolized with dots in Fig. 3).
Using the calibration model concentration predictions were
obtained from the test set spectra. The pooled variance was calculated and control charts made.
Fig. 3. NAS values vs. concentration plot. The open circle symbolizes the average calibration spectra and the dots symbolize the test set spectra.
Fig. 4. Mixing experiment data in variance chart and mean chart using quantitative model. The mean values symbolized with () indicate that the assumption
for the t-test is not fulfilled because the variance is different from the golden
batches.
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
The means at 16, 20 and 25 min were also within the target
limits. At 25 min the batch was on target with a variance comparable to the golden batch. This optimal mixing time was also
identified using the qualitative model.
The results showed that with the qualitative NAS method
it was possible to monitor API homogeneity though no calibration model with reference values was developed. The qualitative
method required fewer batches and is therefore a strong candidate as method in the early development stages. The advantage
of the quantitative model over the qualitative approach was that
tighter limits were obtained due to the more degrees of freedom and calculations were in concentration values. Also the
mean could be compared to the declared target content of the
pharmaceutical product. These differences make adaptation into
a highly regulated manufacturing environment easier and the
interpretation is straightforward.
Secondly, by making a quantitative calibration line, nonlinear behaviour can be detected and more confidence can be put
into the pre-processing method and wavelength selection, this is
not possible in the qualitative model. With the current formulation no such problems were observed but this is not always the
case.
4.3. Validation of proposed methods
To validate the NIR methods comparisons were made
between traditional HPLC and NIR using four different validation batches. All four validation batches were prepared with
10% (w/w) API. The differences between the batches were the
order of mixing the constituents and the mixing time. When mixing was finalized, 30 NIR spectra were recorded and 12 powder
samples were removed for HPLC analysis from each batch. First
results from HPLC will be demonstrated and then compared to
NIR results.
4.3.1. Validation batch evaluation using HPLC
From the 100% calibration batch (golden batch) 30 powder samples of 300 mg were removed and subject to HPLC
analysis. A variance chart was developed. The limits were cal2
culated using Eqs. (6) and (7), substituting stotal
NAS,gb with
2
stotal HPLC,gb and a mean chart was developed using 10% (w/w)
as target and Eq. (9) to calculate the upper and lower control
2
2
limits, substituting stotal
c,gb with stotal HPLC,gb .
Now for each of the four validation batches the following was done; the variance and mean of the 12 HPLC values
were calculated and plotted in the variance and the mean chart
(Fig. 5). Outliers were identified using a Grubbs test ( = 0.05)
and removed prior to the calculation of the variance and mean.
The variances of batches 2 and 4 were within the limits and
thereby equal to the golden batch. The variance of batches 1
and 3 was below the warning limit signalling model incompetence. The means of the validation batches were all lower then
the target value. The means of batches 1 and 3 (symbolized with
) could not be compared with the golden batch using the tstatistic because their variances were different from the golden
batch.
Fig. 5. Variances and means of HPLC values from validation batches in variance
and mean charts.
Fig. 6. Variances (upper figure) and means (lower figure) of concentration NIR
predictions for the validation batches.
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
Steps 14 were performed using 5, 10, 15, 20, 25, 30, 35, 40,
45 and 50 values using the power calculation method described
above. The results are displayed in Fig. 7.
It was observed that when using a small number of samples,
e.g. 10 a large proportion of the variances and means were within
the control limits (Fig. 7) though they supposed to be above the
upper control limits in both charts. When the number of samples
was increased to 20, only a few results were below the upper
warning limits.
The results showed how the risk of committing a Type II error
(false positive) increased if Nnew was low. One has to balance
the risk of committing Type II errors with a practical number of
spectra in order to choose Nnew . In the present case 20 spectra
seemed sufficient to identify de-mixing.
5. Conclusions
Two NIR models were developed for monitoring blend homogeneity using a NIR probe. First a qualitative model using NAS
values from a well-mixed batch were used to generate control
charts for future observations. Second a quantitative regression
model between NAS values and reference concentration was
developed for a set of calibration batches. With the regression
model and well mixed batches, control charts were developed
for the quantitative model. The qualitative method required less
batches to develop and the performance was similar to the quantitative model but some other advantages of the quantitative
method were presented. The qualitative model is suitable for
initial R&D studies when developing a new solid dosage form
drug product. When further development activities have been
conducted or prior to implementation in manufacturing supplementary calibration batches can be measured and a quantitative
model developed. Generally the models are easily developed
and the method and control charts mimics to some degree the
current methodologies for determination of blend uniformity
which makes implementation or replacement of current methods
easier.
In our opinion the proposed methods are better to compare
mixing processes and formulation performance then current
methods [1]. In current methods two statistics are usually calculated and used for homogeneity evaluation and batch-to-batch
comparison, i.e. the mean value (c) and the relative standard
deviation (R.S.D.%). The R.S.D.% is calculated by dividing
standard deviation with the mean value (R.S.D.% = 100 s/c).
The R.S.D.% is then dependent on two numbers and therefore not a good statistic to do batch-to-batch comparison of
10
E.T.S. Skibsted et al. / Journal of Pharmaceutical and Biomedical Analysis xxx (2005) xxxxxx
homogeneity, i.e. different information is mixed up in the statistic. With our proposed NIR methods the comparison is performed with two statistics, i.e. mean concentration and variance
which are independent and both can be directly compared to
the target concentration of the drug product and the variance of
a well-mixed batch. Another aspect of the methods is that the
control limits are based on actual process measurements and is
therefore related to the process capability. In the standard regulatory approaches limits are generic limits and do not relate to
the specific formulation or process.
The proposed methods were validated by comparing results
obtained with NIR and with traditional HPLC analysis. The comparisons showed that both the qualitative and quantitative NIR
models showed similar results as HPLC.
It was shown how the width of the control limits decreased
when more samples were used and how the risk of committing
a Type II observation error also dropped when the number of
measurements were increased. The result was used to provide
guidance of picking a practical number of measurements.
The influence of the effective sample size on the variance was
demonstrated by comparing NIR variance with HPLC variance.
The NIR variance was approximately 10 times higher then the
HPLC variance because of the small sample size in NIR.
The NIR models and control charts were used for monitoring of a mixing study and evidence was found of de-mixing
after certain duration of mixing. The result showed the necessity and importance of monitoring tools like the proposed NIR
methods.
Acknowledgements
Novo Nordisk, Corporate Research Affairs (CORA) who
sponsored this work as a part of E.T.S. Skibsteds Ph.D. project.
Technician Inga Fuchs who performed most of the mixing experiments and the staff at Analytical Development, CMC Development, Novo Nordisk A/S who performed the HPLC analysis.
References
[1] Uniformity of dosage units, in: U.S. Pharmacopeia (USP), 2003. pp.
22272229 (Chapter 905).
[2] F.C. Sanchz, J. Toft, B. Bogaert, D.L. Massart, S.S. Dive, P. Hailey,
Fresenius J. Anal. Chem. 352 (1995) 771778.
[3] S.S. Sekulic, H.W. Ward, D.R. Brannegan, E.D. Stanley, C.L. Evans, S.T.
Sciavolino, P.A. Hailey, P.K. Aldridge, Anal. Chem. 68 (1996) 509513.
[4] D.J. Wargo, J.K. Drennen, J. Pharm. Biomed. Anal. 14 (1996)
14151423.
[5] P.A. Hailey, P. Doherty, P. Tapsell, T. Oliver, P.K. Aldridge, J. Pharm.
Biomed. Anal. 14 (1996) 551559.
[6] S.S. Sekulic, J. Wakeman, P. Doherty, P. Hailey, J. Pharm. Biomed.
Anal. 17 (1998) 12851309.
[7] R.D. Maesschalck, F.C. Sanchz, D.L. Massart, P. Doherty, Appl. Spectrosc. 52 (1998) 725731.
[8] O. Berntsson, L.-G. Danielsson, M.O. Johansson, S. Folestad, Anal.
Chim. Acta 419 (2000) 4554.
[9] O. Berntsson, Characterization and application of near infrared spectroscopy for quantitative process analysis of powder mixtures, Doctoral
Thesis, Kungl Tekniska Hogskolan, Department of Chemistry, Stockholm, 2001.
[10] A.S. El-Hagrasy, H.R. Morris, F.D. Amico, R.A. Lodder, J.K. Drennen,
J. Pharm. Sci. 90 (2001) 12981307.
[11] O. Vaizoglu, Tr. J. Phys. 23 (1999) 97104.
[12] F.J. Muzzio, P. Robinson, C. Wightman, D. Brone, Int. J. Pharm. 155
(1997) 153178.
[13] A. Lorber, Anal. Chem. 58 (1986) 11671172.
[14] E.T.S. Skibsted, J.A. Westerhuis, H.F. Boelens, A.K. Smilde, D.T. Witte,
Appl. Spectrosc. 58 (2004) 264271.
[15] P.W. Hansen, J. Chemometr. 15 (2001) 123131.
[16] H.F. Boelens, W.T. Kok, O.E. de Noord, A.K. Smilde, Anal. Chem. 76
(2004) 26562663.
[17] N.M. Faber, Chemometr. Intell. Lab. Syst. 50 (2000) 107114.
[18] J.N. Miller, J.C. Miller, Statistics and Chemometrics for Analytical
Chemistry, 4th ed., Pearson Prentice Hall, Harlow, 2000.
[19] http://www.brukeroptics.com.
[20] Brukeroptics, OPUS, Version 4.2.
[21] MathWorks Inc., MatLab, Version 6.5 R13.
[22] O. Berntsson, L.G. Danielsson, S. Folestad, Anal. Chim. Acta 364 (1998)
243251.
[23] B. Efron, R.J. Tibshiran, An Introduction to the Bootstrap, Chapman &
Hall, New York, 1993.