Talanta 178 (2018) 686–697

Contents lists available at ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Monitoring of tablet coating processes with colored coatings T

⁎
Shirin Barimani, Peter Kleinebudde
Heinrich Heine University Duesseldorf, Institute of Pharmaceutics and Biopharmaceutics, Universitaetsstr. 1, 40225 Duesseldorf, Germany

A R T I C L E I N F O A B S T R A C T

Keywords: Endpoints of coating processes for colored tablets were determined using in-line Raman spectroscopy. Coatings
Colored tablet coating were performed with six commercially available formulations of pink, yellow, red, beige, green and blue color.
Fluorescence The coatings were comprising pigments and/or dyes, some causing fluorescence and interfering the Raman
Process analytical technology (PAT) tool signal. Using non-contact optics, a Raman probe was used as process analytical technology (PAT) tool, and
In-line Raman spectroscopy
acquired spectra were correlated to the sprayed mass of aqueous coating suspension. Process endpoints were
Science-Based Calibration (SBC)
MCR-ALS
determined using univariate (UV) data analysis and three multivariate analysis methods, namely Projection to
Latent Structures (PLS)-regression, Science-Based Calibration (SBC) and Multivariate Curve Resolution (MCR).
The methods were compared regarding model performance parameters. The endpoints of all coating experiments
could be predicted until a total coating time of 50 min corresponding to coating thicknesses between 21 and
38 µm, depending on the density of the coat formulation. With the exception of SBC, all calibration methods
resulted in R2 values higher than 0.9. Additionally, the methods were evaluated regarding their capability for in-
line process monitoring. For each color, at least two methods were feasible to do this. Overall, PLS-regression led
to best model performance parameters.

1. Introduction thousandth part of the intensity of an incident beam is scattered due to

Tablet film coating is a common pharmaceutical process whereby
non-functional coatings are widely used [1]. The cores have to be
layered sufficiently to ensure safety and efficacy of the product. Higher
compliance can be achieved because of improvement of swallowing or
flavour seal in case of bitter active pharmaceutical ingredients (APIs).
Hereby, the coating thickness is of special importance. Colored, cos-
metics coatings are relatively thin. Nevertheless, a minimum thickness
has to be achieved to guarantee color saturation, as small thickness
differences can lead to observable color variations.
Analysing real-time information during the process helps to enhance
the production process and to detect possible problems [2,3]. It can be
obtained using process analytical technology (PAT) tools, as re-
commended by the Food and Drug Administration in a guidance in
2004 [4]. One possibility is the usage of Raman spectroscopy [5]. This
method was first introduced in 1928 [6]. The basic principle is an en-
ergy transfer between excitation source and molecules, resulting in
inelastic scattering of radiation. The method offers the advantages to
measure fast and non-destructive [7,8]. Additionally, no sample pre-
paration is needed. However, it is important to mention that only one
the weak Raman effect. Raman measurements are sensitive to fluores-
cence which occurs slower, but much stronger than the Raman effect
and often obscures the resulting peaks [9,10]. Fluorescence is more
common for colored coatings than for non-colored ones. Hence, Raman
measurements of tablet coatings are almost restricted to non-colored
coatings so far. Nevertheless, due to the fact that colored coatings
provide a number of benefits such as easier identification during the
manufacturing process as well as improved compliance by the patient,
they are relevant in pharmaceutical processes [11].
Raman spectroscopy was already used as off-line measurement
method for colored samples in the field of arts [12,13]. Measured
samples were pigments, dyes were not investigated yet. Fluorescence
effects, which were caused by the presence of pigments, could be re-
duced by adjusting Raman laser power [14] or the laser excitation
wavelength [15]. Raman spectroscopy has been already reported as
feasible PAT tool for in-line coating thickness predictions of tablets
[16–18]. An active coating was monitored in-line by Müller et al. [19].
Additionally, in-line Raman measurements were performed for an ac-
tive coating process of two-layer tablets by Wirges et al. [20,21]. In
2010, Cahyadi et al. [22] compared different methods for quantifying

Abbreviations: ALS, alternating least squares; API, active pharmaceutical ingredient; MCC, microcrystalline cellulose; MCR, multivariate curve resolution; NIR, near infrared; PEG,
polyethylene glycol; PVA, polyvinyl alcohol; PAT, process analytical technology; PLS, projection to latent structures; RMSEC, root mean square error of calibration; RMSEP, root mean
square error of prediction; SBC, science-based calibration; SEC, standard error of calibration; TiO2, titanium dioxide; UV, univariate
⁎
Corresponding author.
E-mail address: Kleinebudde@hhu.de (P. Kleinebudde).

http://dx.doi.org/10.1016/j.talanta.2017.10.008
Received 14 August 2017; Received in revised form 25 September 2017; Accepted 5 October 2017
Available online 07 October 2017
0039-9140/ © 2017 Elsevier B.V. All rights reserved.
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

the thickness of tablet coatings, such as off-line Raman spectroscopy, 2. Materials and methods
amongst others. The coat comprised red iron oxides in a small con-
centration and fluorescence was not reported. Using Raman data for 2.1. Tablet core materials
tablet coating thickness quantification in presence of fluorescence was
only reported in one article from Romero-Torres et al. [11]. The tablet cores comprised 49.5% (w/w) lactose monohydrate
The Raman signal was obscured after a thin layer of the fluorescent (Tablettose® 80; Meggle, Wasserburg, Germany), 49.5% (w/w) micro-
coat was applied. Several data transformation pretreatment methods crystalline cellulose ((MCC) Avicel® PH-102; FMC International, Little
were needed, although the amount of added fluorochrome is not re- Island Co., Cork, Ireland) and 1.0% (w/w) magnesium stearate (Peter
ported. Greven GmbH & Co. KG, Bad Münstereifel, Germany).
Changes of Raman band intensities occur during the coating pro-
cess, which provide quantitative information. One possibility to extract 2.2. Coating suspension materials
the information out of Raman measurements is the usage of chemo-
metrics [23,24]. Regarding data analysis methods, simple univariate All prepared coating suspensions were based on ready-to-use mix-
(UV) data analysis is often sufficient [25]. For more complex cases, tures comprising polymers and pigments and/or dyes. The resulting
multivariate methods are advantageous [26]. Projection to Latent color was used as batch name. Pink comprised hydroxypropyl methyl-
Structures (PLS)-regression is a well-known method and widely used cellulose, copovidone, polydextrose, polyethylene glycol (PEG), tita-
[27,28]. Nevertheless, the calibration step is complex, as it has to be nium dioxide (TiO2), medium chain triglycerides, iron oxide yellow and
representative and needs to be trained for all expected variations which iron oxide red (Aquarius™ Preferred HSP Pink; Ashland, Covington,
may occur [29]. An alternative suggested method by Marbach in 2005 USA). Yellow and Red consisted of a PEG-polyvinyl alcohol (PVA) co-
[30] is Science-Based Calibration (SBC), which is based on an article of polymer, vinyl pyrrolidone-vinyl acetate copolymer, TiO2, kaolin, so-
2002 comprising the theoretical background [31]. According to the dium lauryl sulfate. Additionally, yellow (Kollicoat® IR Yellow; BASF,
author, this direct calibration method does not need a widely changing Ludwigshafen, Germany) comprised iron oxide yellow and red con-
calibration set. Hence, the calibration step offers the advantages to be tained iron oxide red (Kollicoat® IR Red; BASF). Beige comprised PVA,
less time and cost consuming compared to PLS-regression. Marbach PEG, TiO2, Talc, iron oxide yellow and iron oxide red (Opadry® II Beige;
showed, that SBC is suitable to determine the distribution of APIs in Colorcon® GmbH, Idstein, Germany). Green contained PVA, PEG, TiO2,
tablets using off-line acquired near infrared (NIR) spectra [32]. In-line Talc, iron oxide yellow, tartrazine aluminium lake and brilliant blue
NIR spectroscopy was used by Möltgen et al. for tablet coating thickness FCF aluminium lake (Opadry® II Green, Colorcon®). Blue comprised a
determination [33]. Recently, in-line Raman spectra during tablet PEG-PVA copolymer, talc, TiO2, glycerol monocaprylocaprate and
coating with a non-colored suspension were acquired to compare the brilliant blue FCF aluminium lake (Opadry® QX Blue, Colorcon®).
suitability of SBC, PLS-regression and UV analysis for an endpoint de-
termination of the process. SBC analysis resulted in comparable model 2.3. Coating process
performance parameters and the method proved to have high predictive
power. All spectral changes were linear over time [34]. According to de For each of the six coating experiments, three batches of drug-free
Juan et al., model-free methods, so-called soft-modeling methods [35], cores were coated in a laboratory drum coater (BFC 5, L.B.B. Maschinen
allow to describe a system without the need of an underlying model. + Verfahren GmbH, Ennigerloh, Germany). The batch size for pink and
This is especially advantageous when unknown processes are in- beige was 3000 g and 4000 g for all the other colors. The suspensions
vestigated. A commonly used soft-modeling method is the Multivariate were applied using two 1 mm nozzles (Düsen-Schlick GmbH, Coburg,
Curve Resolution (MCR) [36]. Measuring the spectra of a mixture, it is Germany), with a working a distance of 10 cm from the rotating tablet
capable to decompose them mathematically into basis spectra and to bed. Following table (Table 1) shows the coating process parameters.
determine the corresponding concentration profiles. One solving ap- All aqueous coating suspensions were prepared with deionised
proach is the alternating least squares method (ALS) [37,38]. Applied to water and 20% of solids (w/w), except the beige one, for which the
infrared and Raman spectroscopic data, MCR-ALS was used to study fraction of solids was reduced to 15% (w/w) in order to avoid high
chemical reaction processes in solvent mixtures [39]. The method was viscosity. Spraying was done for 50 min and the pump speed was kept
also suitable to analyse quantitative mixtures of steroid drugs by constant. Due to different suspension viscosities, the amounts of
Holden et al. in 2003 [40]. In the study, MCR-ALS was compared to sprayed coat varied between 529.1 g and 722.0 g. Considering the
PLS-regression and was classified as comparable method. De Juan et al. densities of the coating films and surface to be coated, resulting coating
performed MCR-ALS analysis on images measured by Raman spectro- thicknesses were calculated and are shown in the results section. In
scopy for quality control purposes of tablets comprising different con- order to ensure constant spray rates, the weight changes of the coating
centrations of acetylsalicylic acid in 2004 [41]. Hereby, API distribu- suspensions were monitored during all experiments and linear regres-
tions and impurities could be detected. MCR was also evaluated for the sion was performed.
analysis of artificial food colors in spirits by Lachenmeier and Kessler
[42]. 2.4. Solvent casted film and density measurements
To investigate the feasibility of Raman spectroscopy as PAT tool for
colored coating processes, six commonly available coating formulations Representative of the film layer on the tablet cores, solvent casted
were chosen as tablet coats in the present study. Although cosmetics films were produced using the coating suspensions with solvent casting
coatings are relatively thin, tablets were coated until 50 min coating
time, in order to test model robustness. Depending on the coat densities, Table 1
the minimum calculated thickness was 21 µm. The coating colors (pink, Coating process parameters.
yellow, red, beige, green and blue) resulted from the presence of pig-
Step Pan Spray Inlet air Exhaust air Inlet air
ments and dyes which partly caused fluorescence. Modeling was per-
speed rate /g/ volume temperature /°C temperature /°C
formed in order to predict the coating endpoints, using PLS-regression, /rpm min /m3/h
SBC, MCR-ALS and UV analysis. Additionally, the methods were tested
for their suitability of process monitoring. Warm-up 2 – 100 40 60
Coating 15 12 100 40 60
Drying 15 – 100 45 65
Cooling 2 – 140 30 25

687
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

method on a film drawing bench (Coatmaster 510, Erichsen, Hemer,
Germany). The knife was moved with a speed of 6 mm/s, while the gap
width between coating knife and suspension was about 1000 µm. To
obtain a resulting film under similar conditions as the film on the ta-
blets, the metal plate was heated up to 40 °C corresponding to exhaust
temperature during spraying phase of coating. The drying time was
30 min.
Density measurements of pieces of the solvent casted film were
performed using a helium pycnometer (AccuPyc 1330, Micromeritics
Instrument Corporation, Norcross, USA). The temperature was set to
25 °C ± 0.5 °C. Each sample with a weight of approximately 3 g, was
measured ten times and the resulting density was averaged.
2.5. Raman measurement
Coating processes were monitored in-line using a Raman probe
(RXN2™ Analyser, Kaiser Optical Systems, Ann Arbor, USA) which was
installed through the front door of the coater with a working distance of
21 cm to the tablet bed. Compressed air was blown through a pipe in
the front part of the probe to protect against dust. The spectrometer was
equipped with a non-contact optic sampling device and a PhAT system.
The 785 nm diode laser with 400 mW power forms a circular illumi-
nation area with 6 mm diameter corresponding to 28.3 mm2. A large
sampling area is covered hereby. For data acquisition, the iC Raman™
4.1 software package (Kaiser Optical Systems) was used. Every minute,
three spectra were measured from 150 to 1890 cm−1 in 1 cm−1 steps
and averaged. The exposure time of the laser varied between 0.5 and 3 s
because the different suspension materials showed different Raman
scattering properties. Cosmic ray filtering and dark subtraction were
performed on the spectra.
2.6. Data analysis methods
2.6.1. Calibrated wavenumber ranges and coating time
Four different ranges were calibrated for every color and method:
150–1890 cm−1, 300–800 cm−1, 300–1200 cm−1 and
300–1600 cm−1. 150–1890 cm−1 is the whole measured range, the
choice of the other calibrated waveranges is further described in 3.4.
The data was calibrated until 50 min coating time.
2.6.2. PLS-regression
PLS-regression was performed in Simca® 13.0.2 (Sartorius Stedim
Biotech, Sweden). Raman spectra were mean-centered and standard
normal variate (SNV)-transformed. The observed spectra were used as X
variable and the mass of sprayed aqueous coating suspension as cor-
responding Y variable.
2.6.3. SBC
SBC was carried out on the SNV-transformed spectra using an own
code in Matlab® R2015a (MathWorks, USA). The method is described in
more detail in [34]. For basic understanding, only the main equations
of the method are described here.
The measured spectrum XT at time t is given by:
X T (t ) = y (t )*gT + XNT /AU
with y(t) as concentration of analyte at t (corresponding to the mass of
applied suspension), gT the shape of the response spectrum (spectrum of
the film formed by the coating suspension) and XNT as additional pro-
cess noise. The covariance matrix ∑, which is necessary for the calcu-
lation of the regression vector b, includes the noise X with n as their
total number.
T
XSNV *XSNV
Σ= /AU 2
n−1 (2)
B is needed to convert the observed spectrum into the amount of
applied suspension and is calculated by:
Σ −*g
b= /g/AU
gT *Σ −*g (3)
In a last step, the predicted mass of applied coating in g, ypred, is
calculated. The subscript op means operation point. The SBC method is
optimised for a specific point, the operation point. For the present
study, the amount of sprayed suspension in the last minute of the
coating process of the training set was used. Xpred is the predicted
spectrum corresponding to the measured spectrum Xop at the chosen
operation point, the last coating minute with yop as the corresponding
applied mass.
ypred = yop + (Xpred − Xop )*b/ g (4)
2.6.4. MCR-ALS
MCR analysis with alternating least squares was performed in
Matlab® R2015a with MCR-ALS GUI 2.0 [38]. To determine the number
of components, single value decomposition was used. All MCR models
were built based on two components, namely core and film, as no other
component showed strong Raman scattering properties and additional
components had low eigenvalues. Non-negativity constraint was ap-
plied just to concentration profiles of the sprayed mass of suspension.
This was done because the suspension mass cannot be negative. The
definition of the initial estimations was done manually and a spectrum
of the core and a spectrum of the coated tablet were used. As correlation
constraint, the amount of sprayed suspension was used. Iterative cycles
were repeated 50 times or when the convergence criterion of 0.1% was
achieved. Hence, the iterative process was stopped when the differences
of two consecutive model fits were lower than 0.1%.
2.6.5. UV analysis
To perform UV analysis, the spectra of four wavenumber ranges
were pretreated by SNV transformation, from 150 to 1890 cm−1,
300–800 cm−1, 300–1200 cm−1 and 300–1600 cm−1. Subsequently, a
regression was performed in Microsoft Excel®. The calibration wave-
numbers were chosen depending on the most changing Raman peak
(Table 2). It was obtained by performing principle component analysis
in Simca® 13.0.2 and choosing the wavenumber with the highest
loading. For the range from 150 to 1890 cm−1, two calibration wave-
numbers were chosen, as two wavenumbers with high loadings were
identified. As some peak intensities were not changing linear to the
time, linear regression was not sufficient in all cases. In the table, the
order of polynomial regression is indicated by the superscript after x,
(1) while no superscript indicates linear regression.

Table 2
Selected wavenumbers for calibration of UV for different colors in different ranges.

Color Regression order Range: 150–1890 /cm−1 Range: 300–800 /cm−1 Range:300–1200 /cm−1 Range:300–1600 /cm−1

Pink x 395, 638 638 638 638

Yellow x 173, 515 640 640 640
Red x 357, 642 642 642 642
Beige x3 173, 357 642 642 642
Green x3 514, 1087 638 638 638
Blue x 581, 1087 358 477 358

688
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

2.7. Model performance parameters

The observed Raman spectra were correlated to the mass of applied

aqueous coating suspension in g. Using the data of one performed
coating run as training set the coefficients of determination (R2) were
calculated. Furthermore, following expressions were used to evaluate
the calibration model: root mean square error of calibration (RMSEC),
standard error of calibration (SEC) and the systematic error bias, which
were calculated as:
n
∑i = 1 (yi − yˆi )2
RMSEC = /%
n (5)

n
∑i = 1 (yi − yˆi − bias )2
SEC = /%
n−1 (6)
n Fig. 1. Scaled spectra of core excipients and resulting tablet.
∑i = 1 (yi − yˆi )
Bias = /%
n (7)
core excipients, providing sharp Raman peaks (Fig. 1).
with yi and ŷi as the known and calculated mass of coating suspension in
Magnesium stearate has peaks with high intensity, but is used in
sample i and n as the number of samples. Furthermore, the coefficient of
comparable low concentration as lubricant. In contrast, MCC is just
prediction (Q2) for PLS models was calculated. Q2 was obtained by
slightly Raman active, hence the signal is not intense. Acquiring the
cross-validation of the training set. Hereby, the data was separated into
spectra shown in Fig. 1, the exposure time of 3 s was not changed. The
seven groups and the leave-one-out technique was used. Each 1/7th in
unscaled Raman spectrum of MCC showed factor 20 higher intensity
turn was removed and a model was built based on the 6/7th data left in.
than lactose monohydrate (data not shown). Therefore, the spectra of
From the new model, the left out data was predicted. A number of
Fig. 1 are scaled for better comparability. Additional to the lacking
parallel models were developed by the software. This was repeated,
Raman activity, MCC shows a baseline drift which can be also observed
until all data was predicted. This predicted data was compared with the
in the resulting tablet spectrum, occurring due to fluorescence. Never-
original one and sum of squared errors were calculated. The data of two
theless, MCC was used as one core excipient in this study, because it is a
other experiments, which were performed on different days to check
common excipient in pharmaceutical tablet formulations and in order
model robustness, were used as test sets (test set 1, 2) and root mean
to increase model robustness. The Raman spectra of the solvent casted
square errors of prediction (RMSEP) were calculated for each test set
films of the coat suspensions are shown in Fig. 2. While measuring
(RMSEP1, 2). Eq. (5) can be used for calculation, while ŷi is replaced by
them, the exposure time of 2 s was not changed and the obtained
the predicted mass based on the calibration model. As the calculated
spectra of blue showed the highest intensity (data not shown). Spectra
error values were in g and the amounts of sprayed suspensions were
were scaled to arbitrary units to show the ratios of the resulting peaks.
different, all errors were converted to percentual amount of applied
The formulations contained polymers without sharp Raman peaks
suspension for each experiment in % for better comparability.
as well as coloring pigments such as iron oxides and dyes (Table 4).
All formulations comprised the white pigment TiO2. The poly-
3. Results and discussion
morphic form anatase causes peaks at 396, 515 and 638 cm−1, which
can be observed for pink, yellow and green. These peaks are not clearly
3.1. Resulting coating thicknesses
visible for beige and blue. Also for red, no intense peaks can be ob-
served. One possible reason can be a low concentration of TiO2 in the
The following table (Table 3) shows the masses of applied suspen-
formulation. Another possible reason could be the particle size of the
sions, the densities of the solvent casted films, determined by helium
pigment iron oxide. If the particles are in nm-range, fluorescence can
pycnometry, and the resulting coating thicknesses. A further explana-
occur [43,44]. Nevertheless, all coating formulations were included in
tion about thickness calculation is given in the Supplemental material
the present study, even the ones which do not provide sharp, intense
of [34].

3.2. Choice of core and coat materials

Changes of Raman bands before and after coating provide quanti-

tative information for modeling. Therefore, it is advantageous to use
highly Raman active materials in order to obtain signal changes during
the process. In this study, lactose monohydrate was used as one of the

Table 3
Amount of sprayed suspensions, densities of the films and resulting calculated coating
thicknesses.

Name Core Coating Mass Density of Resulting

mass /g time /min suspension /g film /g/cm3 thickness /μm

Pink 3000 50 653.8 1.477 38.0

Yellow 4000 50 715.2 1.521 30.2
Red 4000 50 722.0 1.536 30.2
Beige 3000 50 557.6 1.619 22.2
Green 4000 50 590.5 1.612 23.6
Blue 4000 50 529.1 1.613 21.1
Fig. 2. Spectra of coat formulations, scaled.

689
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

Table 4 minute corresponding to the core signal (t = 0, bright grey line), the
Overview about comprised pigments / dyes in used coating formulations. observed signal after half amount of the suspension was sprayed (t =
25 min, dark grey line) and the signal of the coated tablets obtained in
Color White pigment Coloring pigment Dye
the last coating minute (t = 50 min, black line).
Pink + + The signal change over time for pink is shown in Fig. 3a. While the
Yellow + + core signal decreases with increasing coating time between 800 and
Red + +
1200 cm−1, the coat signal increases, reinforced by the fact that the
Beige + +
Green + + + pink formulation shows intense Raman peaks as shown in Fig. 2. As a
Blue + + consequence, the signal at t = 50 min contains three intense peaks
between 380 and 780 cm−1, caused by TiO2 in the suspension. Intensity
changes like these were used as quantitative information for modeling.
Raman peaks. Regarding the pink coating, the distances of the spectra between t =
0 min and t = 25 min as well as between t = 25 min and t = 50 min
3.3. Raman spectra intensity changes during coating appear to be equal. Hence, a linear Raman signal intensity increase over
time can be assumed. In contrast, the signals of the other coatings in
Fig. 3 shows the SNV-pretreated Raman spectra of the first coating

Fig. 3. SNV-transformed Raman spectra of first, mid and last coating minute of all colors; pink (a), yellow (b), red (c), beige (d), green (e) and blue (f).

690
S. Barimani, P. Kleinebudde
Fig. 3 do not change linearly over time. The spectral intensities of
yellow, red, beige and green change less in the first half of the process
than in the second half. This can be observed most clearly for beige. In
contrast, the spectral intensity change of blue is steeper in the begin-
ning of the process. The fact that the spectra of t = 25 min and t =
50 min change only slightly, indicates signal saturation. Although all
formulations comprise Raman active TiO2, the sharp, corresponding
peaks cannot be observed for all colors. This could be due to lower
concentrations of this pigment in some formulations compared to the
pink one and correlates to the shown spectra in Fig. 2. Regarding the
spectral intensity changes over time, it can be expected that modeling
red, beige and blue will be more challenging due to non-linear changes
over time and the additional fact that the coat formulations do not lead
to intense Raman scattering. Nevertheless, a change of Raman signal
intensity is necessary for modeling.
To understand the changes in more detail, the spectra of pink and
blue were investigated exemplarily in more detail. These two colors
were chosen, because the signal intensity change of pink appears to be
linear over time and blue shows saturation. Therefore, the untreated
and SNV transformed spectra of every 10th coating minute were plotted
against a wavenumber range between 300 and 1600 cm−1 (Fig. 4). This
range was selected, because lower and higher wavenumber ranges in-
clude noise. Due to different scattering properties of the coating com-
positions, the exposure times of the Raman laser had to vary in order to
avoid detector under- and oversaturation. To get an optimised signal, it
was adjusted between 0.5 and 3 s, depending on the formulation, before
Fig. 4. Spectrum of every 10th coating minute; untreated (a, c) and SNV-transformed (b, d) for pink and blue.
691
Talanta 178 (2018) 686–697
the measurement was started and not changed during the experiment.
In order to find the optimum exposure time, the solvent casted film was
measured, representative for the film layer on the tablets. Hence, the y-
axes of the untreated spectra (Fig. 4a, c) show absolute values.
As shown in Fig. 4a for pink, the intensity of the core signal de-
creases during the coating process to approximately half of the initial
intensity. It can be observed, that the signal decrease is not perfectly
linear. But, pretreating the spectra by SNV-transformation leads to a
linear signal change over coating time (Fig. 4b). This leads to the ex-
pectation that the calibration and prediction errors of pink will be
comparably low. In contrast, the intensities of the Raman spectra of
blue increase more than factor 4. Nevertheless, the signal at t = 50 min
does not provide any sharp Raman peaks. This effect can be explained
by the dye brilliant blue FCF, which is included in the blue coating
formulation. It causes fluorescence [45] and obscures the Raman signal
and hence, the resulting peaks. Although the same dye was among other
colorants in green, the strong fluorescence effect in blue indicates that it
is higher concentrated in blue as in green. In case of blue, pretreating
the spectra (Fig. 4d) does not lead to a lower signal to noise ratio and a
linear intensity change over time. Moreover, intensity changes of the
spectra can be only observed until 30 coating minutes. Afterwards, the
signal saturates. Therefore, using SNV-transformed spectra, modeling
would be only possible with restrictions until 30 min. As raw spectra's
intensities show high linearity, modeling of blue was performed based
on untreated spectra.
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

3.4. Choice of calibration range
The four chosen wavenumber ranges for calibration were
150–1890 cm−1, 300–800 cm−1, 300–1200 cm−1 and
300–1600 cm−1. The first range is the whole measured spectrum. It was
analysed to investigate, if a simpler analysis without further selection of
narrower wavenumber ranges is useful for model improvement.
However, it can be obtained in Fig. 3 that it includes noise. Therefore,
spectra below 300 cm−1 and above 1600 cm−1 were excluded from the
three remaining analysed ranges. The region between 300 and
800 cm−1 comprises three intense peaks from TiO2 which shows strong
Raman scattering properties. Hence, all calibration ranges include this
region. It was extended in 400 cm−1 steps to 1200 and 1600 cm−1 as
these regions comprise peaks of two core materials, namely lactose
monohydrate and magnesium stearate. By analysing these two ranges,
it should be tested if including a broader analysis range, containing
additional information, leads to improved model performance para-
meters.
3.5. Color saturation
A colored coating process of a non-functional coat can be finished
when the color does not change anymore. For the present study, sa-
turation was detected visually. The tablet colors did not change after
20 min for pink, 25 min for red and all other colors were saturated after
30 min.
3.6. Calibration model results obtained by Raman analysis
3.6.1. Comparison of calibration model performance parameters
All calibration models were built up to a process time of 50 min. The
wavenumber ranges leading to the highest R2, Q2 and the lowest errors,
are shown in Table 5 for each color and each method. The calibration
and prediction performance parameters for the remaining wavenumber
ranges can be found in the Supplemental material. The value in
brackets after PLS indicates the number of components and the number
in brackets for UV analysis is the calibrated wavenumber in the given
pretreated range.
A low systematic error (SEC) and hence, a small bias, lead to RMSEC
and SEC with similar values. It is conspicuous that the range between
150 and 1890 cm−1 is only present once in the table. This can be ex-
plained by the fact that it includes the noise range, leading to com-
parable higher errors. It is striking that SBC analysis leads to compar-
able higher error values than the other methods. As an example, the
RMSEC for red obtained by SBC is factor 12 higher than from the
method leading to lowest RMSEC (PLS). Correspondingly, this method
is not suitable to analyse some colors at all. Surprisingly, simple UV
analysis leads to error values, which are comparable to the ones ob-
tained by PLS-regression or MCR. All analyses methods resulted in R2
values higher than 0.9 except SBC analysis in case of red. However,
some colors lead to overall low calibration errors, e.g., pink and green.
In contrast, some others like red or blue show generally higher errors.
Independently from the used analysis method, lowest calibration
errors can be observed for pink, resulting in errors below 2.35% of the
sprayed aqueous suspension. MCR and PLS-regression using two com-
ponents are the most suitable methods for this color, followed by simple
UV analysis. SBC led to comparable higher errors, this can be also ob-
served for yellow. Regarding yellow, PLS resulted in the lowest RMSEC
(0.62%) and SEC (0.77%), followed by UV analysis (1.26% / 1.29%)
and MCR (3.38% / 3.36%). SBC analysis has the highest bias and the
calibration errors are more than factor two higher compared with MCR.
To build a calibration model for red, SBC is not suitable at all, as the R2
is 0.27% and the errors are around 25%. Nevertheless, RMSECs and
SECs below 5% can be obtained by PLS and MCR, while UV analysis
leads to errors around 6.5%. Interestingly, simple UV analysis is suffi-
cient too as method to model beige, followed by PLS-regression. The
errors resulting from SBC and MCR are higher than 6.8%. Analysing
green, all methods resulted in error values below 5%, while SBC
showed a negative trend, indicated by a negative bias. Although un-
treated spectra were included in the models for blue and the coat
showed fluorescence, acceptable calibration errors could be obtained.
With the exception of SBC, all errors were below 2.30%. SBC analysis
led to the highest error values also for this color (RMSEC = 8.47% and
SEC = 9.63%). The fact that the whole measured range is only present
once in the table and the varying error values depending on different
wavenumber ranges indicate that it is of importance to choose a cali-
bration wavenumber range carefully. The resulting errors from simple
UV analysis, which are often comparable to the results obtained by
multivariate analysis methods, lead to the assumption that using com-
plex, analysis methods is often unnecessary. Additionally, the bias

Table 5
Model performance parameters of calibration, R2, Q2, RMSEC, SEC and bias for each color and each method, value in percentual amount of applied aqueous coating suspension (%).

Color Model Range /cm−1 R2 Q2 RMSEC /% SEC /% Bias /%

Pink PLS (2) 300–800 0.9997 0.9996 0.50 0.66 0.43

SBC 300–800 0.9960 2.35 2.30 0.70
MCR 300–1600 0.9997 0.49 0.50 0.00
UV 300–1600 (638) 0.9997 0.52 0.53 0.00
Yellow PLS (3) 300–800 0.9996 0.9988 0.62 0.77 0.47
SBC 300–1200 0.9643 7.78 7.61 2.88
MCR 300–800 0.9871 3.38 3.46 0.28
UV 300–1600 (640) 0.9982 1.26 1.29 0.00
Red PLS (3) 300–1200 0.9958 0.9917 2.00 2.56 1.61
SBC 300–800 0.2731 25.30 25.89 1.25
MCR 300–1200 0.9819 4.00 4.10 0.33
UV 300–1600 (642) 0.9543 6.45 6.58 0.00
Beige PLS (3) 300–1600 0.9958 0.9917 2.02 2.61 1.67
SBC 300–1600 0.9436 9.40 9.58 1.38
MCR 300–1200 0.9513 6.79 6.95 0.51
UV 300–1200 (642) 0.9988 0.99 1.07 0.00
Green PLS (3) 300–1600 0.9988 0.9982 1.05 1.32 0.82
SBC 300–1600 0.9913 4.25 2.89 −2.84
MCR 300–800 0.9794 4.28 4.37 0.06
UV 300–1200 0.9990 1.00 1.05 0.00
Blue PLS (2) 300–1200 0.9987 0.9979 1.13 1.11 0.89
SBC 300–1200 0.9205 8.47 8.63 0.41
MCR 300–1200 0.9960 1.89 0.04 1.15
UV 150–1890(1424) 0.9941 2.30 1.77 0.06

692
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

Fig. 5. Observed versus predicted masses of applied aqueous suspensions resulting from the calibration models of a) pink, b) yellow, c) red, d) beige, e) green and f) blue. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

resulting from UV analysis has the lowest value in all cases. To build
calibration models for the presented colored coatings, SBC seems to be
the least sufficient method.
Further information about the calibration models can be obtained
by plotting the observed versus predicted amounts of sprayed suspen-
sion, which was used y-variable, against each other, as shown in Fig. 5.
For instance, the plots provide information if the calibration models are
also valid to monitor the process additional to endpoint determination.
It is advantageous to print out Fig. 5 in color.
The plots of Fig. 5 support the error values of Table 5. Regarding
process monitoring, the accuracy of all analysis methods for pink and
green was higher than for the other colors, disregarding the analysis
method. This is corresponding to the low error values and can be caused
by a more linear signal intensity change during the coating, as discussed
for Fig. 3. At first, this can be caused by comparably intense peaks the
coat formulations of pink and green provide (Fig. 2). Secondly, the
resulting colors of both coatings are brighter than from the other
coatings, what makes an interference with the Raman signal less
probable.
Similarly, it is possible to monitor the blue coating process, except
by SBC. This is contrary to the initial assumption that the fluorescence
compound in the blue formulation would disturb the Raman signal in a
manner that sufficient modeling would not be possible at all.
Altogether, the plots confirm the assumption, that the SBC calibration
models deviate more than the other methods from the observed mass of
the training set. This corresponds to the comparatively high error va-
lues, as discussed for Table 5.
Likewise, the calibration of red was challenging. Besides the fact

693
S. Barimani, P. Kleinebudde
Fig. 7. Spectrum of every 10th coating minute of red, SNV-transformed.
that SBC was not suitable at all as calibration method for this color
(Fig. 5c, blue prediction line), the process could also not be monitored
in-line by UV analysis. A possible reason could be the resulting color of
red, which showed the highest intensity amongst all colors (Fig. 6).
leading to a non-linear spectral change. Additionally, red reflects higher
wavelengths of the visible spectrum, namely between 620 and 780 nm.
The used Raman excitation wavelength of 785 nm could explain dis-
ruptions and resulting high errors of the calibration models. The pre-
treated Raman spectra, observed every 10th coating minute during the
coating for red, are shown in Fig. 7.
The spectra from 300 to 600 cm−1 change slightly during the first
coating minutes, whereas the spectra between 600 and 1600 cm−1
seem to change linearly. Because of the linear change in the region,
642 cm−1 was chosen as calibration wavenumber for UV analysis.
These variations of intensity changes could explain the varying cali-
bration errors in the table. However, process monitoring can be per-
formed with PLS and MCR. Similar spectra were obtained for beige. The
spectral intensities were changing less in the beginning of the process.
The prediction lines for SBC and MCR in Fig. 5d correspond to this, as
they are less steep in the beginning. Therefore, these methods over-
estimate the beginning of the process and underestimate it in the
second half. Nevertheless, process monitoring of beige is possible by
PLS and UV analysis. Overall, regarding process monitoring as well as
endpoint determination, PLS shows minor deviations, whereas the
largest variations were achieved by SBC analysis. The SBC method is
optimised regarding the prediction of the endpoint. Consequently, this
point is predicted most precisely, but monitoring of the whole coating
process is not possible. Nevertheless, MCR showed the largest variations
in terms of endpoint determination.
3.7. Prediction model results obtained the analysis methods
3.7.1. Prediction errors of both test sets
For each color, two additional experiments were performed on two
different days in order to test calibration model performance. Based on
the chosen calibration waveranges, predictions were performed using
all methods and RMSEPs were calculated. The resulting error values,
RMSEP 1 and 2, are shown in Fig. 8.
694
Talanta 178 (2018) 686–697
Fig. 6. Uncoated and coated tablets, from left to
right: uncoated, pink, yellow, red, beige, green and
blue. (For interpretation of the references to color in
this figure legend, the reader is referred to the web
version of this article.)
Small differences between RMSEP 1 and 2 indicate model robust-
ness. It is noticeable that the prediction errors 1 and 2, calculated by
SBC, show the largest variations. This is corresponding to the errors
resulting from the calibration models and indicates, that SBC models
are not robust.
With the exception of SBC, all RMSEP values for pink are lower than
3%. PLS, MCR and UV analysis led to comparable prediction errors. The
RMSEP values obtained for yellow are higher. PLS led to the lowest
values (1.42% and 3.23%), followed by UV (3.77% and 5.12%) and
MCR (7.21% and 7.46%). This trend corresponds to the calibration
errors. Neglecting SBC, all methods resulted in similar errors for red in a
range from 6.35% to 9.16%. Nevertheless, the errors are comparatively
high. Regarding the beige coating, the low RMSEP 2 value, obtained by
SBC, could be an outlier, as SBC analysis results in the highest errors for
all other colors. The lowest RMSEPs were obtained by UV analysis and
PLS, while MCR errors were slightly higher. The same trend can be
observed for green and blue. The highest RMSEP for blue was 10.16%,
resulting from SBC analysis. This is in contrast to the expectation that
prediction would lead to much higher errors because of the fluores-
cence.
However, coating thickness quantification of a formulation com-
prising a fluorescent dye was already reported by Romero-Torres et al.
[11]. The data was acquired off-line. Sulfanilamide cores, providing
sharp Raman peaks, were coated with a formulation, comprising an
unknown amount of brilliant blue FCF. The relative intensity of the
sulfanilamide bands decreased with increasing amount of coat and sa-
turation occurred after 23 µm calculated coating thickness. Calibration
was performed in a range from 50 to 151 µm and data was analysed by
UV and PLS. Therefore, a lot of data preprocessing had to be performed,
such as mean-centering, SNV-transformation and Savitzky-Golay
second derivative smoothing. Although preprocessing methods are
performed to improve the signal-to-noise ration of data, there is a risk
to decrease the amount of data information. The lowest RMSEC value
which could be obtained in the study was 2.6%. To achieve this, the
first two data points (0 µm and 23 µm) had to be left out, as the signal
change was not linear with increasing coat. This should be assessed
critically, as leaving single data points out can lead to totally different
results. Additionally, most colored coatings are thinner than 50 µm. In
contrast, in-line acquired spectra were used in the present study. They
have the advantage to provide real-time information. Data analysis
could be performed using drug-free cores. The Raman peaks of API
cores usually provide more intense peaks, hence, the modeling process
here was more challenging. Although the same dye as reported by
Romero-Torres et al. was used in the blue formulation of this study, no
spectra had to be left out during the calibration process. Therefore, the
presented models are also valid for thin coatings with film thicknesses
up to 21.1 µm in case of blue. Cahyadi et al. [22] compared different
methods for quantitative coating thickness determination, such as X-ray
fluorescence, Raman and NIR-spectroscopy. They analysed off-line ac-
quired spectra of tablets, which were coated with a suspension com-
prising 1% red iron oxide until 3% weight gain. The authors did not
report fluorescence caused by the pigment. Resulting from PLS analysis,
calibration models with a R2 of 0.89 and prediction models with R2 of
0.88 were obtained. As reference value for the modeling, coat thick-
nesses were determined using direct optical microscopy. Thickness
prediction could be made with a RMSEP of 10.9%. The relatively high
error can be explained by the fact that only eight tablets were mea-
sured. Another reason could be the presence of the red iron oxide
S. Barimani, P. Kleinebudde
Fig. 8. RMSEP values for test set 1 and 2 (RMSEP 1, 2), a) pink, b) yellow, c) red, d) beige, e) green, f) blue.
pigment in the coat, which could cause the comparable high error
value. However, using PLS and MCR analysis, slightly lower errors
could be obtained in the present study. Additionally, the sprayed mass
of suspension was used as reference value in the present study, as
especially in case of thin coatings high thickness variations occur. Each
measured spectrum was correlated to the mass of suspension at the time
point, why the error values could be more precise. A multi-layering
coating process of pellets was monitored in-line in a recent study by
Hisazumi and Kleinebudde [46] and the data was analysed by MCR and
PLS-regression. The Raman spectra were correlated to thicknesses ob-
tained by image analysis. All calibration models resulted in R2 values
695
Talanta 178 (2018) 686–697
higher than 0.98. MCR gave more appropriate results than PLS. This
differs from the results of the present study, as the errors of PLS-re-
gression had lower values and the process endpoints could be de-
termined more precisely. This could be a consequence of the used
materials as well as from the analysis methods. First of all, the Pellets of
Hisazumi and Kleinebudde were layered inter alia with an API and
TiO2, both providing sharp Raman peaks. Secondly, MCR was first in-
troduced as a method to recover pure response profiles for structure
determination. As the spectra were a combination of four components,
namely pellet cores and three different composed layers, MCR could
probably better deal with the data than PLS. The determination of
S. Barimani, P. Kleinebudde
yellow and blue food colors in absinthe spirits was done by Lachen-
meier and Kessler [42]. They analysed the spectrophotometric data
with MCR. Amongst others, it was possible to resolve the pure com-
pound spectra of the blue dye which was also comprised in the blue
coat of the present study, namely brilliant blue FCF. MCR led to com-
parable model performance parameters as PLS-regression and was ad-
ditionally able to provide chemical information. A further component
could be detected by the method in the study. Also this example shows,
that MCR provides additional benefits such as structure elucidation.
Nevertheless, error values in acceptable ranges were obtained by this
method in the present study, although PLS performed slightly better.
The usage of Raman spectroscopy for the identification of yellow pig-
ments in paintings was reported by Ropret et al. [47]. In contrast to the
present study, where the laser power was set to 400 mW, 10 mW were
used. Hence, reducing the laser power could improve results of ex-
periments as presented in this study.
The impact of adjustments of the Raman spectrometer settings for
colored coatings, especially in presence of fluorescence compounds,
should be investigated in further experiments. Overall, colored coating
processes could be predicted until 50 min of coating, although visual
color saturation was achieved much earlier. Although process endpoints
could be determined most precisely by SBC analysis, which can be
optimised in terms of a specific points, it does not provide a sufficient
tool for process monitoring. In total, it is restricted to linear changing
Raman signals over time.
Best model performance parameters were obtained by PLS-regres-
sion. The results of this study implicate that simple UV analysis leads to
comparable error values to multivariate tools such as PLS-regression
and MCR. Taken together, PLS-regression is a well-known analysis
method which can be used for the prediction of colored coating pro-
cesses as well as for process monitoring, while SBC is limited to linear
spectral intensity changes as shown previously [34].
To implement Raman spectroscopy as PAT tool, a test run must first
confirm that the amount of chosen suspension is enough to cover and
color all tablets properly. As the Raman signal contains an averaged
signal of the passing tablets, single outliers are not detected. Using the
optimised coating process parameters, data for a training set should be
collected. The same parameters should be used for the test sets. As an
example, fluctuations of spray rate or drum rotation speed of the test
sets lead to differing signals from training set. Hereby, in-line mon-
itoring is possible. The process endpoint is detected when the signals of
training and test set have same intensities and all mass is applied. So
far, endpoint prediction of colored tablet coatings were detected vi-
sually or based on empirical knowledge. These methods are biased and
using a PAT tool leads to higher product quality corresponding to
regulatory demanding's. Additionally, using the outcome of the present
study, process automatisation is possible also for colored coatings.
Although modeling should be performed carefully and the right para-
meters such as calibration wave range and analysis method needs to be
defined first, this study demonstrated that using Raman spectroscopy as
in-line PAT tool is not limited to non-colored processes anymore. All six
presented colors could be analysed, even in presence of fluorescence, if
modeling is performed properly.
4. Conclusion
It was possible to predict the process endpoints of six colored
coatings using in-line acquired Raman spectra with PLS-regression,
SBC, MCR and UV analyses. As the colors of none of the coatings was
changing after 30 min visually and the process endpoints were 50 min
for all colors, the obtained RMSEP values in this study are acceptable
for the purpose of cosmetic coatings and the processes could be finished
after 30 min. Additionally, for each color, at least one analysis method
besides PLS-regression provided the possibility to monitor the process.
An exception is SBC analysis, which is only suitable for linear changing
Raman signals, like in pink and green. However, Raman spectroscopy is
696
Talanta 178 (2018) 686–697
not limited to processes with the absence of colors and can be used as
PAT tool for colored coatings as well.
Acknowledgments
This work was supported by L.B.B. Maschinen + Verfahren GmbH.
The authors would also like to acknowledge Ashland, BASF and
Colorcon® for providing the coating materials as free samples.
Conflict of interest
The authors declare that they have no conflict of interest to disclose.
Appendix A. Supporting information
Supplementary data associated with this article can be found in the
online version at http://dx.doi.org/10.1016/j.talanta.2017.10.008.
References
[1] K. Knop, P. Kleinebudde, PAT-tools for process control in pharmaceutical film
coating applications, Int. J. Pharm. 457 (2) (2013) 527–536.
[2] T. Peng, Y. Huang, L. Mei, L. Wu, L. Chen, X. Pan, C. Wu, Study progression in
application of process analytical technologies on film coating, Asian J. Pharm. Sci.
10 (3) (2015) 176–185.
[3] K.A. Bakeev, Process Analytical Technology: Spectroscopic tools and
Implementation Strategies for the Chemical and Pharmaceutical Industries, Wiley,
2005.
[4] Food and Drug Administration, Guidance for Industry: PAT - A Framework for
Innovative Pharmaceutical Development, Manufacturing and Quality Assurance.
< http://www.fda.gov/downloads/Drugs/…/Guidances/ucm070305.pdf > (last
accessed 14 August 2017).
[5] J.A.S. Williams, W. Bonawi-Tan, Online quality control with Raman spectroscopy in
pharmaceutical tablet manufacturing, J. Manuf. Syst. 23 (4) (2004) 299–308.
[6] C.V. Raman, A new radiation, Indian J. Phys. 2 (1928) 387–398.
[7] R. Müller, Ramanspektroskopie als PAT-Methode beim Coating von Tabletten im
Trommelcoater, Heinrich Heine Universität, 2007.
[8] A. Paudel, D. Raijada, J. Rantanen, Raman spectroscopy in pharmaceutical product
design, Adv. Drug Deliv. Rev. 89 (2015) 3–20.
[9] F.M. Schnepel, Physikalische methoden in der chemie: Raman‐spektroskopie,
Chem. Unserer Zeit 14 (5) (1980) 158–167.
[10] C.M. McGoverin, T. Rades, K.C. Gordon, Recent pharmaceutical applications of
Raman and terahertz spectroscopies, J. Pharm. Sci. 97 (11) (2008) 4598–4621.
[11] S. Romero-Torres, J.D. Pérez-Ramos, K.R. Morris, E.R. Grant, Raman spectroscopy
for tablet coating thickness quantification and coating characterization in the pre-
sence of strong fluorescent interference, J. Pharm. Biomed. Anal. 41 (3) (2006)
811–819.
[12] C.B. Lauridsen, J. Sanyova, K.P. Simonsen, Raman analysis of complex pigment
mixtures in 20th century metal knight shields of the order of the elephant,
Spectrochim. Acta Part A: Mol. Biomol. Spectrosc. 150 (2015) 54–62.
[13] P. Vandenabeele, A. Hardy, H.G. Edwards, L. Moens, Evaluation of a principal
components-based searching algorithm for Raman spectroscopic identification of
organic pigments in 20th century artwork, Appl. Spectrosc. 55 (5) (2001) 525–533.
[14] N.C. Scherrer, Z. Stefan, D. Francoise, F. Annette, K. Renate, Synthetic organic
pigments of the 20th and 21st century relevant to artist's paints: raman spectra
reference collection, Spectrochim. Acta Part A: Mol. Biomol. Spectrosc. 73 (3)
(2009) 505–524.
[15] F. Schulte, K.W. Brzezinka, K. Lutzenberger, H. Stege, U. Panne, Raman spectro-
scopy of synthetic organic pigments used in 20th century works of art, J. Raman
spectrosc. 39 (10) (2008) 1455–1463.
[16] S. Laske, A. Paudel, O. Scheibelhofer, S. Sacher, T. Hoermann, J. Khinast, A. Kelly,
J. Rantannen, O. Korhonen, F. Stauffer, F. De Leersnyder, T. De Beer, J. Mantanus,
P.-F. Chavez, B. Thoorens, P. Ghiotti, M. Schubert, P. Tajarobi, G. Haeffler, S. Lakio,
M. Fransson, A. Sparen, S. Abrahmsen-Alami, S. Folestad, A. Funke, I. Backx,
B. Kavsek, F. Kjell, M. Michaelis, T. Page, J. Palmer, A. Schaepman, S. Sekulic,
S. Hammond, B. Braun, B. Colegrove, A Review of PAT strategies in secondary solid
oral dosage manufacturing of small molecules, J. Pharm. Sci. 106 (3) (2017)
667–712.
[17] A.S. El Hagrasy, Application of Raman spectroscopy for quantitative in-line mon-
itoring of tablet coating, Am. Pharm. Rev. 9 (1) (2006) 40–45.
[18] T. De Beer, A. Burggraeve, M. Fonteyne, L. Saerens, J.P. Remon, C. Vervaet, Near
infrared and Raman spectroscopy for the in-process monitoring of pharmaceutical
production processes, Int. J. Pharm. 417 (1–2) (2011) 32–47.
[19] J. Müller, K. Knop, J. Thies, C. Uerpmann, P. Kleinebudde, Feasibility of Raman
spectroscopy as PAT tool in active coating, Drug Dev. Ind. Pharmacy 36 (2) (2010)
234–243.
[20] M. Wirges, A. Funke, P. Serno, K. Knop, P. Kleinebudde, Active Coating of two-layer
Tablets - Raman Spectroscopy as PAT Tool for Process Monitoring, AAPS Annual
Meeting, Washington DC; 2011.
[21] M. Wirges, A. Funke, P. Serno, K. Knop, P. Kleinebudde, Monitoring of an active
S. Barimani, P. Kleinebudde Talanta 178 (2018) 686–697

coating process for two-layer tablets-model development strategies, J. Pharm. Sci.
102 (2) (2012) 556–564.
[22] C. Cahyadi, A.D. Karande, L.W. Chan, P.W.S. Heng, Comparative study of non-de-
structive methods to quantify thickness of tablet coatings, Int. J. Pharm. 398 (1–2)
(2010) 39–49.
[23] M.J. Pelletier, Quantitative analysis using Raman spectrometry, Appl. Spectrosc. 57
(1) (2003) 20A–42A.
[24] B.K. Lavine, Chemometrics, Anal. Chem. 70 (12) (1998) 209–228.
[25] J. Rantanen, Process analytical applications of Raman spectroscopy, J. Pharmacy
Pharmacol. 59 (2) (2007) 171–177.
[26] J.H. Kalivas, Multivariate calibration, an overview, Anal. Lett. 38 (14) (2005)
2259–2279.
[27] S. Wold, M. Sjöström, L. Eriksson, PLS-regression: a basic tool of chemometrics,
Chemom. Intell. Lab. Syst. 58 (2) (2001) 109–130.
[28] P. Geladi, B.R. Kowalski, Partial least-squares regression: a tutorial, Anal. Chim.
Acta 185 (1986) 1–17.
[29] S. Sharma, M. Goodarzi, L. Wynants, H. Ramon, W. Saeys, Efficient use of pure
component and interferent spectra in multivariate calibration, Anal. Chim. Acta 778
(2013) 15–23.
[30] R. Marbach, A new method for multivariate calibration, J. Near Infrared Spectrosc.
13 (5) (2005) 241–254.
[31] R. Marbach, Useful New Multivariate Calibration Technique and Application to IR
ATR Paper Coating Measurement (VTT Annual Report: Research Activities in
Optoelectronics and Electronics Manufacturing), 2001 (2002), pp. 49–53.
[32] R. Marbach, Multivariate Kalibrierung, Selektivität und die SBC-methode, Chem.
Ing. Tech. 82 (4) (2010) 453–466.
[33] C.V. Möltgen, T. Herdling, G. Reich, A novel multivariate approach using science-
based calibration for direct coating thickness determination in real-time NIR pro-
cess monitoring, Eur. J. Pharm. Biopharm. 85 (3, Part B) (2013) 1056–1063.
[34] S. Barimani, P. Kleinebudde, Evaluation of in–line Raman data for end-point de-
termination of a coating process: comparison of science–based calibration, PLS-re-
gression and univariate data analysis, Eur. J. Pharm. Biopharm. 119 (2017) 28–35.
[35] A. De Juan, E. Casassas, R. Tauler, Soft modeling of analytical data, in: R.A. Meyers
(Ed.), Encyclopedia of Analytical Chemistry: Instrumentation and Applications,
Wiley, Chichester, 2000, pp. 9800–9837.
[36] A. de Juan, R. Tauler, Multivariate curve resolution ((MCR)) from 2000: progress in
concepts and applications, Crit. Rev. Anal. Chem. 36 (3–4) (2006) 163–176.
[37] J. Jaumot, R. Gargallo, A. de Juan, R. Tauler, A graphical user-friendly interface for
MCR-ALS: a new tool for multivariate curve resolution in MATLAB, Chemom. Intell.
Lab. Syst. 76 (1) (2005) 101–110.
[38] J. Jaumot, A. de Juan, R. Tauler, MCR-ALS GUI 2. 0: new features and applications,
Chemom. Intell. Lab. Syst. 140 (2015) 1–12.
[39] C.A. Holden, S.S. Hunnicutt, R. Sánchez-Ponce, J.M. Craig, S.C. Rutan, Study of
complexation in methanol/Water mixtures by infrared and Raman spectroscopy
and multivariate curve resolution—alternating least-squares analysis, Appl.
Spectrosc. 57 (5) (2003) 483–490.
[40] T. Azzouz, R. Tauler, Application of multivariate curve resolution alternating least
squares (MCR-ALS) to the quantitative analysis of pharmaceutical and agricultural
samples, Talanta 74 (5) (2008) 1201–1210.
[41] Ad Juan, R. Tauler, R. Dyson, C. Marcolli, M. Rault, M. Maeder, Spectroscopic
imaging and chemometrics: a powerful combination for global and local sample
analysis, TrAC Trends Anal. Chem. 23 (1) (2004) 70–79.
[42] D.W. Lachenmeier, W. Kessler, Multivariate curve resolution of spectrophotometric
data for the determination of artificial food colors, J. Agric. Food Chem. 56 (14)
(2008) 5463–5468.
[43] D. Shi, M.E. Sadat, A.W. Dunn, D.B. Mast, Photo-fluorescent and magnetic prop-
erties of iron oxide nanoparticles for biomedical applications, Nanoscale 7 (18)
(2015) 8209–8232.
[44] C. Karunakaran, J. Jayabharathi, R. Sathishkumar, K. Jayamoorthy, Interaction of
fluorescent sensor with superparamagnetic iron oxide nanoparticles, Spectrochim.
Acta Part A: Mol. Biomol. Spectrosc. 110 (2013) 151–156.
[45] K. Nishi, S.-I. Isobe, Y. Zhu, R. Kiyama, Fluorescence-based bioassays for the de-
tection and evaluation of food materials, Sensors 15 (10) (2015) 25831.
[46] J. Hisazumi, P. Kleinebudde, In-line monitoring of multi-layered film-coating on
pellets using Raman spectroscopy by MCR and PLS analyses, Eur. J. Pharm.
Biopharm. 114 (2017) 194–201.
[47] P. Ropret, S.A. Centeno, P. Bukovec, Raman identification of yellow synthetic or-
ganic pigments in modern and contemporary paintings: reference spectra and case
studies, Spectrochim. Acta Part A: Mol. Biomol. Spectrosc. 69 (2) (2008) 486–497.

697