Semelhante a Spectroscopic-Based Chemometric Models for Quantifying Low Levels of Solid-State Transitions in Extended Release Theophylline Formulations
Semelhante a Spectroscopic-Based Chemometric Models for Quantifying Low Levels of Solid-State Transitions in Extended Release Theophylline Formulations (20)
2. and isothermal micro-calorimetry, and vibrational techniques (i.e.,
near-infrared [NIR] and Raman spectroscopy) in quantifying solid-
state transitions in pharmaceutical solids.5-8
Of these techniques,
vibrational spectroscopic techniques such as NIR and Raman have
garnered much attention because of their rapid data acquisition
rate, low cost, ease of use, nondestructive nature, and the
requirement of little or no sample preparation. Moreover, these
techniques are sensitive to structural, conformational, and envi-
ronmental changes at the molecular level.9
Although both tech-
niques measure molecular vibrations, NIR spectra is due to
overtone and combination bands arising from light absorption,
whereas Raman spectra arises from inelastic light scattering asso-
ciated with loss of vibrational energy. The development of fiber
optic probes and instrumental advancements has allowed NIR and
Raman to be used as a process analytical technique (PAT) tool for in-
line, at-line, and off-line process monitoring. In addition, combi-
nation of vibrational spectroscopy and data analytical tools such as
chemometrics have significantly enhanced the sensitivity of the
techniques for quantitative analysis and improved the wealth of
information that can be obtained from the spectra.10
Theophylline is bronchodilator used in the treatment of asthma
and chronic obstructive pulmonary disease. More than 7 decades
since its discovery, theophylline still remains the most widely used
bronchodilator worldwide although its use has been limited to pa-
tients with poorly controlled disease conditions.11,12
Currently, 4
different polymorphs of theophylline, 3 anhydrous polymorphs and
theophylline monohydrate (TMO), have been identified. Theophyl-
line anhydrous form II (THA) is the most stable form at room tem-
perature and used in pharmaceutical formulations. The most
commonly encountered alterations in the solid state of THA during
storage and in use are the transitions to and from THA to TMO.13,14
Transition of THA to TMO is associated with a significant decrease
in dissolution and bioavailability.13,15,16
Theophylline is a narrow
therapeutic index drug; in addition, its patient population consists
mainly of individuals with uncontrolled disease conditions, and
hence, any minor variations in bioavailability may have a very sig-
nificant impact on the clinical efficacy and incidence of side effects.
Several authors have reported the use of NIR and Raman as PAT
tools for monitoring transitions of THA during manufacturing.17-20
NIR has also been used in differentiating between the unbound
and bound water content during wet granulation.18
Otsuka et al.
also reported quantitative chemometric PXRD models for predict-
ing the hydrate content of powders containing both THA and
TMO.21
The present study extends the use of NIR and Raman
spectroscopy beyond their present role as PAT tools to quantifica-
tion of low-level pseudopolymorphic transition (less than 5% of the
total solid content) of THA in controlled release formulations of
theophylline. The authors also report on the effectiveness of
different data preprocessing techniques and commonly used
regression models in chemometric method development.
Materials and Methods
Material
THA, magnesium stearate (MgS), and lactose monohydrate (LM)
were purchased from Sigma Aldrich (St. Louis, MO). Hydroxy-
propylmethylcellulose K100M was obtained from Colorcon (Har-
leysville, PA). Colloidal silicon dioxide (Aerosil 200) was obtained
from (Evonik, Parsippany, NJ). TMO was prepared by recrystalliza-
tion of saturated THA solution in deionized water at 70C. The
monohydrate crystals obtained were filtered, dried overnight at
ambient temperature, and stored in a chamber maintained at 95%
relative humidity. The TMO crystals were characterized by PXRD,
ssNMR, and DSC before use.
Figure 1. (a) DSC, (b) thermogravimetric analysis, (c) ssNMR, and (d) XRD characterization of THA and TMO.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e10598
3. Preparation of Calibration Samples
Theophylline formulations were prepared in-house using ex-
cipients commonly found in commercial theophylline products.
The formulation consisted of THA 53%, K100M 33%, LM 11%, aerosil
0.1%, and MgS 2.5%. A similar formulation was prepared using TMO
in place of THA. The calibration samples were prepared by mixing
the THA and TMO formulations to obtain sample matrices with 0%-
25% of the active pharmaceutical ingredient (API) being TMO.
NIR Spectroscopy
NIR spectra of THA, TMO, calibration, and the independent test
samples were collected over the wavelength range of 1100-2500
nm in 2-nm increments with FOSS NIR Spectrophotometer 6500
(FOSS NIR Systems, Inc., Laurel, MD) equipped with rapid content
analyzer. The spectra were acquired with the Vision software,
version 3.2 (FOSS NIR System Inc.). Each spectrum was an average
of 60 scans. The spectra of polytetrafluorethylene were used as the
reference spectra.
Raman Spectroscopy
Raman spectra were measured with a noncontact Raman probe
(Raman PhAT-RXN1 Analyzer; Kaiser Optical System Inc., Ann Ar-
bor, MI) with a spot size of 6 mm and laser wavelength and strength
of 785 nm and 350 W, respectively. The spectra were recorded in
triplicates over the wavelength range of 175-1875 cmÀ1
at a
resolution of 1 cmÀ1
and measurement time of 1 min (exposure
time of 15 s and 4 scans). Data acquisition was performed with iC®
Raman software, version 4.1 (Kaiser Optical System Inc.).
Data Analysis
Data analysis and chemometric models were developed with the
Unscrambler X software (version 10.1; CAMO ASA, Norway) and
SIMCA, version 14 (Umetrics AB, Umea, Sweden). Principal compo-
nent regression (PCR) analyses were performed with Unscrambler,
whereas all partial least squares regression models (PLSR) were
developed with the SIMCA software. Model development was based
on the current US Food and Drug Administration guidelines on
development and submission of NIR analytical procedures.22
Results and Discussion
Characterization of TMO and THA Crystals
The ssNMR spectra, PXRD spectra, and DSC thermogram of the
THA and TMO crystals used are shown in Figure 1. The DSC scan of
THA had a sharp melting endotherm at 273C. An additional broad
endotherm was observed between 50C to 102C in the DSC ther-
mogram of TMO. This was due to the loss of water from the mon-
ohydrate crystals. This was confirmed by the thermogravimetric
analysis thermogram, which showed a weight loss of about 8.9%.
The extent of weight loss matches the theoretical value of 9%, which
further confirms the transformation of THA to TMO. Also, the PXRD
Figure 2. Raw (a, b) and MSC-SG second derivative, (c, d) NIR spectra of THA, TMO, THA- and TMO-formulated drug products, and of sample matrices.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105 99
4. pattern of the prepared TMO crystals showed the absence of
characteristic THA peaks at 7.2 and 12.5 2q, and the presence of
distinctive TMO peaks at peaks at 8.8, 11.5, and 27. In addition,
the conversion of THA to TMO resulted in a change in the chemical
shift position of the carbonyl carbon in the ssNMR spectra from
150.9 ppm to 148.4 ppm. However, the use of PXRD and ssNMR for
quantification of low levels of solid-state transition required a run
time of more than 2 h and 8 h, respectively, to obtain an optimum
signal-to-noise ratio. For these reasons, quantitative models based
on PXRD and ssNMR were not pursued further.
NIR Spectroscopy
The NIR spectra of THA, TMO, and the sample matrices are
shown in Figure 2. The NIR spectra of THA were mainly due to CH
stretching and deformation bands of the methyl carbon (1170 nm
and 1368 nm), CH stretching bands of the methine carbons (1660
nm) and combination bands due to NH stretching vibration (2258
nm). Pseudopolymorphic transition of THA to TMO results in
changes in the NIR spectra along the entire wavelength range
because of the high susceptibility of this technique to moisture. The
most notable difference between the NIR spectra of THA and TMO
was the appearance of OH peaks from adsorbed and crystalline
water (1937 nm and 1970 nm) and an OH stretch first overtone
peak (1476 nm). The presence of excipients did not significantly
interfere with peaks of THA and TMO other than the appearance
of crystalline water peaks of LM (1934 nm). The differences be-
tween THA and TMO sample matrices were further enhanced by
the application of mathematical algorithms, which unraveled
overlapping bands and increased the signal-to-noise ratio. The
mathematical algorithms used are discussed further in the later
section.
Raman Spectroscopy
As water is a weak Raman scatter, any variations in the Raman
spectra of THA and TMO were mainly because of the changes in
molecular vibrations due to hydrogen bond interactions between
theophylline and water molecules. The Raman spectra of THA,
TMO, and the sample matrices can be seen in Figure 3. The most
Figure 3. Raman spectra of (a) THA, TMO, THA- and TMO-formulated drug products. (b) Calibration sample matrices.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105100
5. notable difference between the spectra of THA and TMO was the
replacement of the double carbonyl peaks at 1662 cmÀ1
and 1704
cmÀ1
with a single sharp peak at 1686 cmÀ1
. THA Raman peaks at
1612 cmÀ1
(C¼C), 1572 cmÀ1
(C¼N), 1427 cmÀ1
(CH3 deformation),
1190 cmÀ1
(C-C), and 928 cmÀ1
(CH3 rocking) shifted to lower wave
numbers in TMO. On the other hand, THA peaks at 1316 cmÀ1
, 1286
cmÀ1
(C-N stretching), and 558 cmÀ1
(O¼C-N bend) shifted to
higher wave numbers. LM and MgS also had Raman peaks in be-
tween 1500 cmÀ1
and 500 cmÀ1
. However, these peaks did alter the
differences between varying THA and/or TMO contents in the
sample matrices as APIs have higher Raman activity than the ex-
cipients. The Raman spectra of the sample matrix showed
Table 1
Statistical Figures of Merit and Extent of Variance Explained by PLSR and PCR Latent Variables
Method Pretreatment Model Latent
Variables
Variance Explained Statistical Figures of Merit
X Block Y Block
Factor Accumulated Factor Accumulated Slope Offset R2
Correlation RMSEE/C RMSECV
NIR MSC-SG PLSR 1 0.85 0.85 0.98 0.98 1 8.8E-07 0.989 0.994 0.831 0.846
2 0.10 0.95 0.01 0.99
PCR 1 0.85 0.85 0.98 0.98 0.985 0.240 0.986 0.993 0.881 0.957
2 0.08 0.93 0.01 0.99
SNV-SG PLSR 1 0.85 0.85 0.98 0.98 1 1.6E-07 0.999 0.995 0.851 0.855
2 0.10 0.95 0.01 0.99
PCR 1 0.86 0.86 0.98 0.98 0.982 0.221 0.985 0.992 0.992 1.077
2 0.08 0.99 0.01 0.99
Raman MSC-SG PLSR 1 0.86 0.86 0.99 0.99 0.999 1.6E-07 0.998 0.999 0.405 0.383
2 0.13 0.99 0.01 1.00
PCR 1 0.86 0.86 0.98 0.98 0.998 0.019 0.998 0.999 0.393 0.434
2 0.13 0.99 0.02 1.00
SNV-SG PLSR 1 0.81 0.81 0.99 0.99 1 2.1E-07 0.998 0.999 0.392 0.315
2 0.18 0.99 0.01 1.00
PCR 1 0.86 0.86 0.98 0.98 0.998 0.037 0.999 0.999 0.387 0.437
2 0.13 0.99 0.02 1.00
RMSEE, root mean error of estimate; RMSEC, root mean square error of calibration.
Figure 4. (a) Weighted residual, (b) Hotelling T2
plots of PLSR Raman models and PLSR score plots for (c) MSC-SGetreated and (d) SNV-SGetreated NIR models.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105 101
6. discernable differences in spectra at the concentration range
studied (0%-25% wt/wt TMO).
Chemometric Analysis
Spectra Pretreatment
Although proper data collection is essential in developing
quantitative NIR models, NIR spectra is inherently confounded by
systemic variations due to light scattering from smaller particles
with size comparable to the NIR wavelength, surface roughness and
shape, crystalline defects, and density fluctuations.23,24
Another
source of variation is the differences in effective path length. These
factors may lead to baseline shift (multiplicative scatter effect) and
nonlinearity in the spectra data. On the other hand, Raman spectra
are less affected by physical variations. However, sample fluores-
cence, subsampling, and sample inhomogeneity can alter the data
of the Raman spectra. These variations were removed by mathe-
matical treatment of the spectra data.
Scatter removal techniques (multiplicative signal correction
[MSC] and standard normal variate [SNV]) were used to minimize
unwanted variations. The effectiveness of these techniques was
compared using the root mean square error of prediction (RMSEP)
and standard error of prediction (SEP) of their respective models.
Both MSC and SNV remove physical light scattering effect and
correct baseline shifts.23,25,26
MSC removes artifacts and imper-
fections by estimating the correction coefficients by the least
squares method using the average spectrum of the calibration set as
the reference spectrum and correcting the recorded spectrum using
the slope and intercept of the linear regression model. The offset
correction concept for SNV is similar to MSC; however, each spec-
trum is processed on its own without the need for a reference
spectrum or linear regression.23,24
Application of MSC and SNV
alone did not significantly improve the RMSEP and SEP when
compared to the raw data therefore the spectral data was further
subjected to secondary derivative treatment based on the
SavitzkyeGolay (SG) with a third-order polynomial and 15-point
smoothing. This technique further removed any additive and
multiplicative effects in the data without decreasing the signal
strength or signal-to-noise ratio. Further data treatment by second
derivative SG led to a significant improvement in RMSEP value
when compared to the untreated data and SG only treated data. The
RMSEP values for MSC-SG and SNV-SG treated data were not sta-
tistically different. (Table 1) Spectral truncation did not improve the
models, hence, was not pursued further.
Figure 5. PLSR and PCR loading plots for (a) NIR and (b) Raman sample matrices.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105102
7. Outlier Detection
The normalized weighted residual plot (DModX norm) and the
Hotelling T2
plots were used in identifying outliers and extreme
samples that may alter the prediction capability of the models
(Fig. 4). The residual plot is a measure of the normalized distance of
an observation in the training set from the model at a critical value
of 0.05. An observation is considered a moderate outlier if the
weighted residual is more than twice the critical value computed
from the F-distribution. All the observations used in model devel-
opment were below twice the critical values (Dcrit of 1.235 and
1.243 for Raman and NIR models, respectively). In addition, the
Hotelling T2
and the score plots were used as complementary
techniques in the detection of sample outliers and the influence of a
sample on the model. A T2
value greater than the critical value at
95% and 99% confidence limit indicates the observation is further
away from similar observations in the score space. These values are
proportional to the sample leverage which is a measure of the in-
fluence of a sample on the model. All the sample matrices used in
developing the Raman models had T2
values lower than the critical
values of 6.853 and 11.14 for 95% and 99% confidence limit,
respectively. A similar observation was obtained for all models
developed from the NIR spectra.
PLSR and PCR Models
The 2 most used regression models in multivariate analysis,
PCR and PLSR, were used for quantitative model development for
both NIR and Raman spectra. PCR models are based on spectral
decomposition of the X matrix into principal components (prin-
cipal component [PC] or X-scores), which explain the maximum
variation in the data. However, the predictive variables or PCs in
PCR models may not necessarily correlate with the predicted
response (Y matrix or concentration). On the other hand, in PLSR
models, the spectra are decomposed into X and Y scores such that
there exist a strong correlation between the predictive variables
and the predicted response.27-29
All models were developed and
validated taking into account the European Medicine Agency and
United States Food and Drug Administration guidelines on devel-
opment and submission of NIR analytical procedures guidance for
industry.22,30
The models were validated using the cross validation
approach in which the same data set was used for model cali-
bration and validation. Two latent variables (PLSR factors or PCR
PCs) were used in developing all the models. This was because it
gave better statistical values for predicted residual error sum of
squares and root mean square error of cross validation (RMSECV).
Choosing the right number of latent variables is essential to avoid
an over fitted or under fitted model. The 2 latent variables of the
PLSR and PCR models accounted for !93% and !99% of the vari-
ance in the X and Y blocks in both the MSC-SG and SNV-SG NIR
models (Table 1). Similarly, more than 99% of the variance in the X
and Y blocks were explained in all models based on the Raman
spectra. In addition, the first latent variables (PC-1 and factor-1 for
PCR and PLSR, respectively) for all the models accounted for most
of the X and Y variance in the data set. This suggests a possible
correlation between the latent variables and increasing TMO
content.
Furthermore, the score plots reinforced the possible correlation
between the latent variables and increasing TMO content as
demonstrated by an increase in score number with increasing
amounts of TMO (Fig. 4). The loading and coefficient plots were also
used as complimentary parameters in assessing the relationship
Figure 6. Permutation plots for (a) Raman and (b) NIR PLSR models and predicted versus actual % TMO plots for (c) MSC-SGetreated and (d) SNV-SGetreated Raman models.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105 103
8. between the latent variables and changes in the concentration of
TMO (Y variable). The coefficient plot represents changes in the Y
variable due to variations in an X variable, whereas all the other X
variables are kept at the average value. The NIR loading plots for
both PLSR and PCR models showed peaks and very high coefficient
values at 1970 nm and 2340 nm. Both peaks are common to TMO
and also increased in intensity as the amount of TMO in the sample
matrix increased (Figs. 2 and 5). Similarly, the most prominent
peaks of the latent variables (PC-1 and factor-1) for PCR and PLSR
loading plots of the Raman spectra were stretching and bending
vibrations of TMO at 1686 cmÀ1
(C¼O), 1322 cmÀ1
(C-N), 1250
cmÀ1
(H-N¼C), and 674 cmÀ1
(O¼C-N).
Model Fit and Performance Assessment
The fitness of the models was assessed using statistical param-
eters: R2
, correlation, bias, residuals, root mean square error of
calibration and RMSECV. The R2
and correlation coefficients for all
the models were !0.985. The slopes were all very close to 1 (!0.98)
which indicated the low levels of systematic errors. In addition, all
the models had very low offset values. The RMSECV values were
also very low: not more than 1% in all models. The model perfor-
mance and prediction accuracy were assessed by their RMSEP, SEP,
bias, and Q2
values. The RMSEP value is a measure of the total error
or the average uncertainty expected when the model is used in
predicting the concentration of an independent sample. The RMSEP
values were less than 1.25% wt/wt and 0.5% wt/wt for PSLR and PCR
models of the NIR and Raman spectra, respectively. The biases
associated with all the models were very low and statistically
insignificant at p 0.05. This further confirmed the precision and
the low levels of systematic and random errors in the models.
The Q2
also referred to as the cross-validated R2
value, measures
the fraction of the total variation in the TMO content that can be
predicted by a component as estimated by cross validation at a
p value of 0.05. This serves as a measure of the predictive ability of
the model. The Q2
value was computed as:
Q2
¼ 1 À
PRESS
SSYðTotalÞ
where SSY is the residual sum of squares and PRESS is the predic-
tion error sum of squares.31,32
Models with a Q2
value !0.5 have
good predictivity. The Q2
values were greater than 0.98 which is an
indication of good predictive ability.
Moreover, the validity of the NIR and Raman models were
further confirmed by the permutation plots (Fig. 6). The permuta-
tion plots was used to assess the legitimacy of the low risk asso-
ciated with the models and how well the models will predict the
TMO content for new independent observations. In the permuta-
tion test, the order of the observations was randomly permuted,
new models developed from the data and their performance
assessed using the R2
and Q2
diagnostic statistic. The procedure is
then repeated several times (N ¼ 20) to obtain a null distribution of
each diagnostic statistic (R2
and Q2
) and the validity of the model
assessed by comparing the R2
and Q2
of the original (unpermuted)
model with the new models. The original model is said to be valid if
the Q2
and R2
are on the lower left of that of the original model in
the permutation plot.33
Furthermore, there was not any significant difference in statis-
tical parameters for NIR models in which the spectra data were
preprocessed using MSC-SG and those in which SNV-SG was used.
A similar observation was for models based on the Raman spectra.
Also the model fit parameters for PCR and PLSR models were
similar. However, the regression models of the Raman spectra had
better statistical parameters of RMSECV, RMSEP, bias, and offset
values than their corresponding NIR models. The better
performance of Raman spectroscopy could be attributed to Raman
spectra usually having sharper and better-defined spectra peaks
than NIR. This improves the discriminatory power and sensitivity of
Raman spectroscopy in the presence of excipients. In addition, most
APIs usually have much higher Raman activity than excipients and
Raman spectroscopy unlike NIR is sensitive to lattice vibrations
which are significantly altered when there are changes in the
crystal lattice due to polymorphism.
TMO Quantification
The prediction accuracy of the models was tested on indepen-
dent observations with known TMO contents. The independent
samples consisted of 2.5%, 5%, 7.5%, and 10% of the total API content
as TMO. This corresponds to 1.33%, 2.66%, 3.98%, and 5.33% of the
total solid content of the formulation, respectively. The indepen-
dent samples had the same formulation composition as the sample
matrices used in developing the models. The spectra were mean-
centered and preprocessed either by SNV-SG or MSC-SG as was
done for sample matrices used for actual model development. The
models had good precision and accuracy. The model predicted
values were very close to the actual TMO content in the indepen-
dent samples with an absolute difference between the model
predicted and actual TMO content less than 1.76% and 1.28% for NIR
and Raman models, respectively (Table 2). In addition, the RMSEP
values for the independent samples were low ( 0.44% wt/wt and
1.24% wt/wt) for both Raman and NIR models, respectively.
Conclusion
The relationship between the solid-state form of a pharma-
ceutical solid and its product performance has been well estab-
lished. The risk associated with these polymorphic changes in
pharmaceutical solids is even greater for products such as
theophylline, which has a narrow therapeutic index. Controlling
the quality of these products is therefore essential in ensuring
consistent product performance and product safety. Chemometric
models that allow easy, rapid, and accurate quantification of the
amount of TMO in extended release formulations were developed
using NIR and Raman spectroscopy. The model accuracy and
diagnostic statistics such as RMSECV, SEP, bias, and Q2 values were
not significantly different for spectra data pretreated by a combi-
nation of SNV and second derivative S-G and that in which multi-
plicative scatter correction and second derivative SavitzkyeGolay
Table 2
Model Predicted and Actual TMO Percentage Content for Independent Samples
Method Pretreatment Actual % TMO
(wt/wt)
Predicted % TMO content (wt/wt)
PLSR PCR
NIR MSC-SG 2.5 3.53 ± 0.69 3.12 ± 0.88
5 4.77 ± 0.08 4.06 ± 0.78
7.5 7.52 ± 0.46 7.13 ± 1.10
10 11.76 ± 0.63 9.76 ± 0.70
SNV-SG 2.5 3.72 ± 0.70 3.16 ± 1.93
5 5.20 ± 0.72 4.03 ± 0.92
7.5 7.47 ± 0.47 6.93 ± 1.30
10 10.16 ± 0.07 9.55 ± 0.79
Raman MSC-SG 2.5 2.72 ± 0.10 2.43 ± 0.54
5 4.85 ± 0.06 4.79 ± 0.45
7.5 6.95 ± 0.06 6.86 ± 0.41
10 10.10 ± 0.04 11.28 ± 0.53
SNV-SG 2.5 2.61 ± 0.20 2.48 ± 0.53
5 4.74 ± 0.10 4.74 ± 0.48
7.5 6.87 ± 0.21 6.88 ± 0.42
10 10.90 ± 0.02 11.0 ± 0.48
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105104
9. were used. The performance of PLSR and PCR models was similar.
The models were able to accurately quantify low levels of pseu-
dopolymorphic changes in extended release theophylline formu-
lations. However, the performance of the Raman chemometric
models was better than the ones based on the NIR spectra. The
performance of the chemometric models may be affected by
product-induced variations such as changes in formulation excip-
ients, the type, grade, and physical properties of excipients used,
and possible process-induced variations. All anticipated variations
must be accounted for in the sample matrix used in model devel-
opment to ensure the model accuracy and robustness. This work
further highlights the utility of NIR and Raman chemometric
models as quality control tools.
References
1. Harris RK. Polymorphism in the Pharmaceutical Industry. Weinheim, Germany:
WILEY-VCH Verlag; 2006.
2. Hilfiker R, Blatter F, Raumer Mv. Relevance of solid-state properties for phar-
maceutical products. In: Polymorphism. Weinheim, Germany: Wiley-VCH Ver-
lag GmbH Co. KGaA; 2006:1-19.
3. Raw AS, Furness MS, Gill DS, Adams RC, Holcombe Jr FO, Yu LX. Regulatory
considerations of pharmaceutical solid polymorphism in Abbreviated New
Drug Applications (ANDAs). Adv Drug Deliv Rev. 2004;56:397-414.
4. Shah B, Kakumanu VK, Bansal AK. Analytical techniques for quantification of
amorphous/crystalline phases in pharmaceutical solids. J Pharm Sci. 2006;95:
1641-1665.
5. Siddiqui A, Rahman Z, Khan MA. Application of chemometric methods to dif-
ferential scanning calorimeter (DSC) to estimate nimodipine polymorphs from
cosolvent system. Drug Dev Ind Pharm. 2014:1-5.
6. Rahman Z, Mohammad A, Akhtar S, Siddiqui A, Korang-Yeboah M, Khan MA.
Chemometric model development and comparison of Raman and 13C solid-
state nuclear magnetic resonanceechemometric methods for quantification
of crystalline/amorphous warfarin sodium fraction in the formulations. J Pharm
Sci. 2015;104:2550-2558.
7. Korang-Yeboah M, Akhtar S, Siddiqui A, Rahman Z, Khan MA. Application of
NIR chemometric methods for quantification of the crystalline fraction of
warfarin sodium in drug product. Drug Dev Ind Pharm. 2015:1-11.
8. Newman AW, Byrn SR. Solid-state analysis of the active pharmaceutical
ingredient in drug products. Drug Discov Today. 2003;8:898-905.
9. Bugay DE. Characterization of the solid-state: spectroscopic techniques. Adv
Drug Deliv Rev. 2001;48:43-65.
10. Roggo Y, Chalus P, Maurer L, Lema-Martinez C, Edmond A, Jent N. A review of
near infrared spectroscopy and chemometrics in pharmaceutical technologies.
J Pharm Biomed Anal. 2007;44:683-700.
11. Barnes PJ. Theophylline. Am J Respir Crit Care Med. 2003;167:813-818.
12. ZuWallack RL, Mahler DA, Reilly D, et al. Salmeterol plus theophylline
combination therapy in the treatment of COPD. Chest. 2001;119:1661-
1670.
13. Ando H, Ishii M, Kayano M, Ozawa H. Effect of moisture on crystallization of
theophylline in tablets. Drug Dev Ind Pharm. 1992;18:453-467.
14. Ando H, Takayuki O, Masaaki I, Suniio W, Yasuo M. Crystallization of theoph-
ylline in tablets. Int J Pharm. 1986;34:153-156.
15. Herman J, Visavarungroj N, Remon JP. Instability of drug release from anhy-
drous theophylline-imcrocrystalline cellulose formulations. Int J Pharm.
1989;55:143-146.
16. Adeyeye CM, Rowley J, Madu D, Javadi M, Sabnis SS. Evaluation of crystallinity
and drug release stability of directly compressed theophylline hydrophilic
matrix tablets stored under varied moisture conditions. Int J Pharm. 1995;116:
65-75.
17. Airaksinen S, Luukkonen P, Jørgensen A, Karjalainen M, Rantanen J, Yliruusi J.
Effects of excipients on hydrate formation in wet masses containing theoph-
ylline. J Pharm Sci. 2003;92:516-528.
18. Jørgensen A, Rantanen J, Karjalainen M, Khriachtchev L, R€as€anen E, Yliruusi J.
Hydrate formation during wet granulation studied by spectroscopic methods
and multivariate analysis. Pharm Res. 2002;19:1285-1291.
19. Wikstr€om H, Kakidas C, Taylor LS. Determination of hydrate transition tem-
perature using transformation kinetics obtained by Raman spectroscopy.
J Pharm Biomed Anal. 2009;49:247-252.
20. Amado AM, Nolasco MM, Ribeiro-Claro PJA. Probing pseudopolymorphic
transitions in pharmaceutical solids using Raman spectroscopy: hydration and
dehydration of theophylline. J Pharm Sci. 2007;96:1366-1379.
21. Otsuka M, Kinoshita H. Quantitative determination of hydrate content of
theophylline powder by chemometric X-ray powder diffraction analysis. AAPS
PharmSciTech. 2010;11:204-211.
22. FDA 2015 Development and Submission of Near Infra Red Analytical
Procedures Guidance for Industry. Available at: http://www.fda.gov/
downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM
440247.pdf. Accessed August 14, 2015.
23. Barnes RJ, Dhanoa MS, Lister SJ. Standard normal variate transformation and de-
trending of near-infrared diffuse reflectance spectra. Appl Spectrosc. 1989;43:
772-777.
24. Rinnan Å, Berg Fvd, Engelsen SB. Review of the most common pre-processing
techniques for near-infrared spectra. Trends Analyt Chem. 2009;28:1201-1222.
25. Martens H, Nielsen JP, Engelsen SB. Light scattering and light absorbance sepa-
rated by extended multiplicative signal correction. Application to near-infrared
transmission analysis of powder mixtures. Anal Chem. 2003;75:394-404.
26. Geladi P, MacDougall D, Martens H. Linearization and scatter-correction for
near-infrared reflectance spectra of meat. Appl Spectrosc. 1985;39:491-500.
27. Næs T, Martens H. Principal component regression in NIR analysis: viewpoints,
background details and selection of components. J Chemom. 1988;2:155-167.
28. Geladi P, Kowalski BR. Partial least-squares regression: a tutorial. Analytica
Chimica Acta. 1986;185:1-17.
29. Wold S, Esbensen K, Geladi P. Principal component analysis. Chemometr Intell
Lab Syst. 1987;2:37-52.
30. European Medicines Agency. Guideline on the use of near infrared spectroscopy
by the pharmaceutical industry and the data requirements for new submissions
and variations. Available at: http://www.ema.europa.eu/docs/en_GB/document_
library/Scientific_guideline/2014/06/WC500167967.pdf. Accessed August 14,
2015.
31. Consonni V, Ballabio D, Todeschini R. Comments on the definition of the Q2
parameter for QSAR validation. J Chem Inf Model. 2009;49:1669-1678.
32. Consonni V, Ballabio D, Todeschini R. Evaluation of model predictive ability by
external validation techniques. J Chemom. 2010;24:194-201.
33. Szymanska E, Saccenti E, Smilde A, Westerhuis J. Double-check: validation of
diagnostic statistics for PLS-DA models in metabolomics studies. Metabolomics.
2012;8:3-16.
M. Korang-Yeboah et al. / Journal of Pharmaceutical Sciences 105 (2016) 97e105 105