Measurement of the percentage volume particle ... - Pharmacy Eprints

Measurement of the percentage volume particle size
distribution of powdered microcrystalline cellulose using
reflectance near-infrared spectroscopy†
Andrew J. O’Neil, Roger D. Jee and Anthony C. Moffat*
Centre for Pharmaceutical Analysis, The School of Pharmacy, University of London, 29/39
Brunswick Square, London, UK WC1N 1AX
Received 25th June 2003, Accepted 29th September 2003
First published as an Advance Article on the web 13th October 2003
This is the first reported method for determining the percentage volume particle size distribution of a powder
(microcrystalline cellulose) by near-infrared (NIR) reflectance spectroscopy. A total of 113 samples of powdered
microcrystalline cellulose were used from six different commercially available grades, with different moisture
contents (range: 0.9–4.8% m/m). NIR reflectance measurements of these samples were made in narrow soda glass
vials. Reference particle size data for the samples were acquired by laser diffraction. The NIR data were then
calibrated to measure particle size by partial least squares regression. The effects of a range of different NIR data
pre-treatments on calibration and prediction precision were investigated. Overall, simple absorbance data were
found to produce regression models with the best predictive ability (root mean square error of prediction =
0.90%). The method was also found to be insensitive to moisture content.
Introduction
Determination of powdered pharmaceutical material particle
size is important 1,2 since a material’s particle size distribution
exerts a significant effect on the physical properties of the bulk
material. 3 As a consequence of this, particle size measurements
are routinely performed on raw materials prior to commencement
of pharmaceutical manufacture. 4 The techniques typically
used for measurement include forward angle laser light
scattering (FALLS) and electrical zone sensing. 5 Though
accurate, these methods suffer from the disadvantage of being
time consuming and may therefore delay manufacture. 6
The potential application of near-infrared (NIR) spectroscopy
with its ability to analyse rapidly powdered materials and
with minimal sample preparation, has long been suggested for
particle size determination of powdered pharmaceuticals. 7–9
Pasikatan et al. 10 has reviewed the near-infrared literature on
particle size measurements. Despite these suggestions, NIR has
remained largely a technique used for chemical analysis. 11
Powdered pharmaceutical materials may be suited for NIR
particle size determinations since they are diffusely reflecting
materials. 7 In the NIR region (1000 nm to 2500 nm) these
materials both absorb and scatter light, 7 resulting in spectra with
overlapping absorption bands, non-uniform baselines and
varying offsets. The effects of scatter have been shown to vary
with the particle size, 9 sample porosity 6 (and hence compaction
pressure) and with the wavelength 12 and can be described using
Rayleigh and Mie theory, 13 or alternatively using the Kubelka–
Munk theory of diffuse reflectance. 13
Previous studies that have examined the effects of particle
size on NIR spectra have demonstrated that reflectance varies
non-linearly with particle size. 7,13–15 Ciurczak et al. 7 found that
reflectance exhibited an inverse relationship with mean particle
size in agreement with Mie theory. 13 However, this relationship
does not necessarily apply in all cases and is dependent on the
shape of the particle size distribution of the sample, the particle
shape and the material’s refractive index. 13 The presence of
very small particles will further complicate the relationship as
† Electronic supplementary information (ESI) available: Particle size and
spectral data. See http://www.rsc.org/suppdata/an/b3/b307263k/
THE
ANALYST
FULL PAPER
www.rsc.org/analyst
these may exhibit Rayleigh scatter, which is proportional to the
third power of the particle size. 13 Although NIR spectroscopy
has been used for the measurement of nanoparticles. 16
The complicated relationship between reflectance at a
spectral wavelength and particle size has resulted in most work
in this area focusing on chemometric calibration methods,
rather than theoretical models. 9,15,17 A novel method for
classifying pharmaceutical powders involved transforming
second derivative NIR spectra to polar co-ordinates. 9 Each NIR
spectrum was reduced to a single quality point in a plane, that
could be plotted in Cartesian co-ordinates. Linear plots of the
logarithm of the particle size versus the x or y co-ordinate were
found to show significant correlation.
Multivariate calibration methods have proven most successful
in calibrating NIR data to measure particle size. One recent
application 18 determined the median particle size of drugs by
multiple linear regression whilst Frake et al. 19,20 used neural
networks.
A significant advance was the production of calibrations to
measure the cumulative particle size distribution, at interpolated
cumulative percentage intervals (range: 5% to 95%, in 11
discrete steps) using multiple linear least squares regression
(MLR) and principal components regression. 21 However,
because the method involved modelling each discrete interval,
and the method used fewer particle size intervals than the
original laser diffraction data for modelling, it consequently
produced predicted distributions with less detail than in the
original reference method’s distributions.
The aim of this work was to produce a procedure for NIR
calibrations to measure the percentage volume particle size
distributions of powdered microcrystalline cellulose using all
the data from the reference (laser diffraction) method in one
mathematical step using partial least squares regression.
Experimental
Instrumentation
NIR reflectance measurements of the powdered samples were
made using a FOSS instruments grating spectrometer (Model
1326 Analyst, 2003, 128, 1326–1330
This journal is © The Royal Society of Chemistry 2003
DOI: 10.1039/b307263k

6500, FOSS NIRSystems, Silver Springs, MD, USA) equipped
with a rapid content analyser module. Particle size distribution
data for the microcrystalline cellulose samples were acquired by
laser diffraction using a Malvern Mastersizer X Laser Diffraction
instrument (Malvern Instruments, Malvern, UK) equipped
with a dry powder feeder.
Materials
The microcrystalline cellulose samples used were obtained
from one supplier (FMC International, Wallingstown, Little
Island, Co Cork, Ireland). These samples (n = 113 from 109
different batches with 2 samples from each of 4 of the batches)
were from six different grades with different median particle
sizes (range of percentage volume median particle sizes:
14.96–198.6 mm; mean of percentage volume median particle
sizes = 95.47 mm) and moisture contents which ranged from
0.9 to 4.8% m/m.
Sample preparation and presentation
A single narrow soda glass vial was filled with material for each
powdered sample. A NIR diffuse reflectance spectrum was then
recorded as Log10(1/R), where R is the reflectance, for each
sample as the average of 32 scans, over the wavelength range
1100 to 2498 nm, at 2 nm increments (700 data points).
Reference particle size data for each sample were acquired by
laser diffraction with a dry powder feeder.
Data analysis
Data were processed using programmes written in Matlab 5
Scientific and Technical Programming language (The Mathworks
Inc., Natick, MA, USA). The near-infrared spectra and
particle size data are available as Electronic Supplementary
Information (ESI)†.
Results and discussion
Particle size and spectral characteristics
The six different grades of microcrystalline cellulose were
found to broadly fit into seven distinct particle size distributions
(Fig. 1). These distributions varied from narrow and very
positively skewed to broad slightly negative skewed distributions
(d90–d10 ranged from 34.4 to 266.2 mm). While a more
evenly distributed set of values in each particle size channel
would have been desirable, the data was adequate for demonstrating
the feasability of measuring particle size distributions
by NIR spectroscopy.
All absorbance spectra exhibited non-uniform baselines and
overlapped combinations and overtones from the mid-infrared
region. The non-uniform baselines observed were the result of
multiple scatter which has been shown to be due to both particle
size and surface moisture content (Fig. 2a).
Spectral data pre-treatments
A range of data pre-treatments commonly employed in NIR
spectrometry were applied to the spectral data and their effects
on calibration and prediction precision were compared against
results for absorbance (Log10(1/R)) data. The data pretreatments
tested were: standard normal variate (SNV), 22
quadratic baseline correction (‘detrend’), 22 SNV coupled with
detrend 22 and Savitzky-Golay 11 point quadratic second
derivative 23 (Fig. 2).
Owing to the increase in noise in the Savitzky-Golay
smoothed second derivative spectra beyond 2200 nm, the
wavelength range used with this pre-treated data was truncated
to 1110 to 2200 nm (546 data points).
Preliminary data analysis
The Malvern Mastersizer measured both the percentage volume
and the percentage cumulative volume particle size distributions
in 32 different channels. The channels’ width followed a
geometric progression so that they had the same resolution
(channel width divided by the diameter in the centre of the
channel). For calibration purposes, the mean value between the
low and high particle sizes for a given channel were used (range:
0.9–448.3 mm). Preliminary investigation of the total particle
size data set revealed that the largest particle size channel did
not contain any useful information, hence this channel was
removed from the data set giving 31 channels for calibration.
With the data sets and chemometric techniques used previously,
18,21 it was not found possible to produce accurate
calibrations for the percentage volume particle size distribution.
The requirement was to model all 31 channels from the laser
diffraction instrument (sets of particle sizes) and preliminary
investigation showed that this could be calibrated by a PLS2
model. The PLS2 model allowed all the particle size channels to
Fig. 1 Percentage (a) volume and (b) cumulative volume particle size
distributions for the microcrystalline cellulose samples (n = 113). Note,
lines drawn between the mid points of each channel in (a). Channel widths
follow a geometric progression.
Analyst, 2003, 128, 1326–1330 1327

e modelled at the same time, rather than one at a time as with
conventional PLS1 models. 24
Model generation
Regression models for particle size distribution data were
produced with original and pre-treated NIR spectral data by
PLS2. 24 This summarises the important variability in both the
NIR spectral data (X) and the particle size data (Y). This
procedure projected the information in the high-dimensional
data spaces (X, Y) down onto low-dimensional spaces defined
by a small number of latent variables. The X and Y data sets
were first mean-centred and scaled to unit variance.
The number of PLS2 dimensions, a, for each of the pretreated
data sets (X, Y) were estimated by cross-validation. For
this, the first 110 of 113 observations (X, Y) were divided into
11 subsets of 10 observations. Next, calculation of PLS2 models
of rank a (where a = 1, 2, 3, …) was performed for all
combinations of 10 of the 11 subsets of data. With each PLS2
model, the remaining subset of NIR spectral and particle size
data was used to test the goodness of fit of the model (a PLS2
dimensions) by measuring the predicted residual error sum of
squares (PRESS) between model predicted and reference
particle size data. This approach of dividing the data into subsets
was preferable to ‘leave-one-out-cross-validation’, requiring
considerably less computation time. The number of PLS2
components required to extract the information from X and Y
was that number a with the lowest PRESS value.
Rank of PLS2 models
With the exception of the model produced with Savitzky–Golay
second derivative data, 6 PLS2 components were required to fit
the data and give the lowest PRESS value (Table 1). Clearly, the
Savitzky–Golay smoothed second derivative did remove more
scatter and baseline drift information from the NIR data than the
other pre-treatments, requiring only 4 components to model the
Table 1 PLS2 results of cross validation, calibration and prediction for the
NIR data sets tested
Fig. 2 NIR spectra of microcrystalline cellulose of different volume median sizes (from bottom to top) 14.96, 53.34, 86.67, 99.03, 119.91 and 171.96 mm
and spectral pre-treatments: (a) absorbance, (b) SNV of absorbance, (c) detrend of absorbance, (d) SNV of absorbance and (e) Savitzky-Golay 11 point
quadratic second derivative of absorbance.
1328 Analyst, 2003, 128, 1326–1330
Data
PLS
components PRESS a
RMSEP
(%) b
Absorbance 6 0.220 0.90
SNV 6 0.165 1.03
SNV detrend 6 0.167 1.08
Detrend 6 0.193 1.07
Savitzky–Golay 2nd derivative 4 0.201 1.69
a Predicted residual error sum of squares using mean centred and
standardised particle size data. b Root mean square error of prediction for
predicted particle size data.
Fig. 3 Predicted residual error sum of squares (PRESS) for successive
PLS2 components extracted using NIR absorbance and laser diffraction
data.

data. With all data pre-treatments tested, typical idealised plots
of PRESS versus number of PLS components were obtained,
with clear minima. This is shown for the absorbance data set in
Fig. 3. The SNV transformation provided a model with the
lowest PRESS value. All other pre-treated NIR data sets
produced models with low PRESS values except for absorbance
data which had the highest PRESS value—a third greater than
for the model derived from SNV data (Table 1).
Calibration and validation precision
In order to test the predictive capacity of the method, the NIR
particle size data set was split into a calibration and a validation
set. For PLS2 modelling, 90 spectra and particle size results
were randomly selected for calibration (range of percentage
volume median particle sizes: 14.96–198.6 mm; mean of
percentage volume median particle sizes = 102.3 mm). The
Fig. 4 Plots of laser diffraction measured percentage volume distributions for the validation samples (n = 23) with NIR predicted values (+) overlaid.
Analyst, 2003, 128, 1326–1330 1329

emaining 23 samples were used to test the predictive abilities
of the models (range of percentage volume median particle
sizes: 15.15–126.3 mm; mean of percentage volume median
particle sizes: 68.74 mm). PLS2 models for original and pretreated
NIR-particle size data were created from the calibration
data sets, each having the number of components determined by
cross-validation. The predictive ability of the models were
determined using the validation data sets by calculation of the
root mean square error of prediction (RMSEP, eqn. 1) between
PLS2 model predicted and reference percentage frequency, y,
over all channels, m, and samples, n:
The best predictive model was obtained using absorbance
data which produced the lowest prediction errors: RMSEP =
0.90% (Table 1). The SNV, detrend and SNV-detrend pretreatments
also produced models which showed low prediction
errors (range: 1.03 to 1.08%), however these were higher than
those obtained with absorbance data (Table 1). The Savitzky–
Golay smoothed second derivative transformation produced a
model with far higher prediction errors, with an RMSEP nearly
double than that obtained with absorbance data (Table 1). Plots
of percentage volume particle size distributions for the 23
validation samples are shown in Fig. 4 and show the very good
agreement of the NIR predicted results with those obtained by
laser diffraction. For example, Fig. 3a shows excellent agreement
over the whole range of particle sizes studied (0.9–448.3
mm). Bimodal distributions were also reasonable well modelled
(Fig. 4f and m) and even skewed distributions were modelled
well (Fig. 4d and t).
The linear association between NIR predicted and laser
diffraction analysis measured percentage volume for the
validation set (all 23 samples and 31 channels) was found to
have a highly significant correlation, r, of 0.97 (n = 713, p =
0.005), with a slope of 1.008 and intercept of 20.024%.
The method was also insensitive to the moisture content of
the samples. No significant correlation between the d10, d50 and
d90 reference particle size values and moisture content of the
samples was found (probability for no correlation: 0.115, 0.212
and 0.274 respectively).
Conclusion
The measurement of the percentage volume particle size
distribution of powdered microcrystalline cellulose may be
achieved by NIR spectrometry. PLS2 models of absorbance and
reference particle size data produced the best calibrations,
robust to variation in moisture content, with low prediction
errors. In all cases, pre-treatment of the NIR spectral data with
commonly used scatter correcting transformations was found to
reduce the particle size information in the spectral data set,
especially with the Savitzky-Golay smoothed second derivative
pre-treatment. However, sufficient particle size information still
remains in the data after pre-treatment for calibration. The rank
of PLS2 models can be effectively determined by crossvalidation
using the PRESS statistic; however as reported by
other workers, this statistic did not act as a good estimate of the
robustness of the model as the model derived from absorbance
1330 Analyst, 2003, 128, 1326–1330
(1)
data had the highest PRESS value. An independent validation
set provided a better estimate of this. Measurement of particle
size distribution by NIRS therefore offers a great saving in time
and has the potential to be used on-line in particle size
analysis.
Acknowledgements
The authors thank FMC International for provision of microcrystalline
cellulose samples and reference particle size analyses.
FOSS NIRSystems are thanked for the loan of a NIR
spectrometer and The Mathworks Inc. are thanked for supplying
Matlab software. A J O’Neil thanks Pfizer for a research
grant.
References
1 M. E. Aulton, Pharmaceutics: The Science of Dosage Form Design,
Churchill Livingstone, Edinburgh, 1988, p. 564.
2 H. G. Barth, S. Sun and R. M. Nickol, Anal. Chem., 1987, 59,
142–162.
3 C. Washington, Particle Size Analysis in Pharmaceutics and Other
Industries, Ellis Horwood, New York, 1992, pp. 9–10.
4 British Pharmacopoeia 2002, HM Stationery Office, London, 2002,
vol. 2, Appendix XVII A & B, pp. A325–A326.
5 A. Simmons, Int. LABMATE, 1993, 17, 23–24.
6 P. A. Hailey, A. C. E. Oakley, P. Doherty, A. J. Pettman, D. C. A.
Sharp and D. M. H. Barnes, NIR News, 1994, 5, 10–12.
7 E. Ciurczak, P. Torlini and P. Demkowicz, Spectroscopy, 1986, 1,
36–39.
8 W. Plugge and C. van der Vlies, J. Pharm. Biomed. Anal., 1993, 11,
435–442.
9 C. van der Vlies, W. Plugge and K. J. Kaffka, Spectroscopy, 1995, 10,
46–49.
10 M. C. Pasikatan, J. L. Steele, C. K. Spillman and E. Haque, J. Near
Infrared Spectrosc., 2001, 9, 153–164.
11 M. Blanco, J. Coello, H. Itturriaga, S. Maspoch and C. de la Pezuela,
Analyst, 1998, 123, 135R–150R.
12 C. R. Bull, Analyst, 1991, 116, 781–786.
13 G. Kortum, Reflectance Spectroscopy, Principles, Methods, Applications,
Springer-Verlag, Berlin, 1969.
14 K. H. Norris and P. C. Williams, Cereal Chem., 1984, 61,
158–165.
15 J. L. Ilari, H. Martens and T. Isaksson, Appl. Spectrosc., 1988, 42,
722–728.
16 J. P. Higgins, S. M. Arrivo, G. Thurau, R. L. Green, W. Bowen, A.
Lange, A. C. Templeton, D. L. Thomas and R. A. Reed, Anal. Chem.,
2003, 75, 1777–1785.
17 W. Plugge and C. van der Vlies, J. Pharm. Biomed. Anal., 1996, 14,
891–898.
18 A. J. O’Neil, R. D. Jee and A. C. Moffat, Analyst, 1998, 123,
2297–2302.
19 P. Frake, C. N. Luscombe, D. R. Rudd, J. Waterhouse and U. A.
Jayasooriya, Anal. Commun., 1998, 35, 133–134.
20 P. Frake, I. Gill, C. N. Luscombe, D. R. Rudd, J. Waterhouse and U.
A. Jayasooriya, Analyst, 1998, 123, 2043–2046.
21 A. J. O’Neil, R. D. Jee and A. C. Moffat, Analyst, 1999, 124,
33–36.
22 I. J. Barnes, M. S. Dhanoa and S. J. Lister, Appl. Spectrosc., 1989, 43,
772–777.
23 Advances in Near-Infrared Measurements, ed. G. Patonay, JAI Press
Inc., Greenwich, 1993, p. 57.
24 R. G. Brereton, Analyst, 2000, 125, 2125–2154.