Pharmaceutical Development and Technology, 2012; 17(4): 389–397

© 2012 Informa Healthcare USA, Inc.

ISSN 1083-7450 print/ISSN 1097-9867 online
DOI: 10.3109/10837450.2010.550623

RESEARCH ARTICLE

Quality by design approach in the optimization of the

spray-drying process
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13

Arnaud Baldinger, Lucas Clerdent, Jukka Rantanen, Mingshi Yang, and Holger Grohganz

Faculty of Pharmaceutical Sciences, Department of Pharmaceutics and Analytical Chemistry, University of Copenhagen,
Copenhagen, Denmark

Abstract
Context/objective: The aim of this study was to illustrate the influence of the processing parameters, inlet temperature,
atomization air flow rate and feed flow rate, on critical quality attributes of spray-dried powders using design of
experiments (DoE).
Methods: Spray-dried powders were characterized by laser diffraction, X-ray powder diffraction (XRPD) and near-
infrared spectroscopy (NIR). Multivariate analysis of two different experimental designs was performed to elucidate
the optimal process conditions.
Results and discussion: XRPD revealed that the spray-dried powders consisted of crystalline β-mannitol and amorphous
trehalose. Non-invasive NIR measurement was successfully used for correlating the critical quality attribute particle
For personal use only.

size with size determined by laser diffraction. The full factorial design proved to be unsuitable due to the non-linear
influence of factors. The composite face-centered design improved the quality of the models and showed both linear
and non-linear influence of the parameters on the outcomes. A model explaining the influence of the factors on all
quality attributes showed similar results as the models optimized for a single response.
Conclusion: This study showed the applicability of DoE for the investigation of spray-dried powders. The knowledge
of the interplay between process parameters and quality attributes will enable rational process design to achieve a
desired outcome.
Keywords:  Design of experiments, quality by design, multivariate analysis, spray-drying, solid state
characterization

Introduction be utilized when defining a design space.[2] A compo-

The quality by design (QbD) concept is a major new
development in pharmaceutical processing and pro-
duction. QbD can be expected to have a major impact
on the way pharmaceutical companies will perform
in the future. This approach is used in pharmaceuti-
cal development to reduce the development costs, but
especially to ensure the quality of the product. The
pharmaceutical QbD consists of choosing and develop-
ing relevant formulations and production processes to
ensure a predefined product quality.[1] Process param-
eters are identified through a risk assessment, based on
the extent to which their variation can have an impact
on the quality of the drug product. This assessment can
nent of QbD allows an understanding of factors and
their interaction effects, using a set of experiments
based on design of experiments (DoE). DoE is used in
several industrial sectors and statistical experimental
designs such as full factorial design and composite face-
centered design are the core elements of DoE.[3] These
designs are employed to determine the relation, which
exists between process parameters and the outcome of
a process. In the current project, DoE was applied to the
production of fine inhalable particles as it is an obvious
process that requires thorough understanding of the
influence of process parameters on the product attri-
butes. Different techniques are available to produce dry

Address for Correspondence: Holger Grohganz, Faculty of Pharmaceutical Sciences, Department of Pharmaceutics and Analytical Chemistry,
University of Copenhagen, Universitetsparken 2, DK-2100 Copenhagen, Denmark. Tel: +45 35 33 64 73. E-mail: hgr@farma.ku.dk
(Received 23 September 2010; revised 10 December 2010; accepted 10 December 2010)

389
 390  A. Baldinger et al.
powder for inhalation, such as jet milling, freeze-drying
or spray-drying.
Spray-drying is a one-step continuous process with
the possibility of optimizing the physicochemical char-
acteristics of the obtained dry powder for inhalation.
Besides producing fine powder for inhalation, spray-
drying is one of the most commonly used techniques
in pharmaceutical industry for the purposes of micro-
encapsulation, drying thermally sensitive materials and
improving the flow properties of granules. Pulmonary
drug delivery is one of the most promising alternative
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
routes for the administration of biomacromolecular
drugs such as therapeutic peptides and proteins. This is
due to the large alveolar epithelial surface area (up to
100 m2) in combination with an extremely thin (0.1–0.2
μm) absorptive mucosal membrane and good blood
supply in the lung. To promote the systemic absorption,
a deposition of the drug in the alveolar region of the
lung is necessary. Among the physicochemical param-
eters influencing the lung deposition, aerodynamic
particle size is the most important one. The optimum
aerodynamic particle size distribution for most inha-
lation aerosols has generally been recognized to be in
the range of 1–5 µm,[4] which can be achieved by spray-
drying.
With regard to the spray-drying process, several
process parameters are contributing to the outcome
For personal use only.
of the process such as inlet temperature, air flow rate
and feed flow rate. In addition, formulation parameters
should be considered in the design of inhalable dry
powder, which also greatly influences the physical char-
acteristics of the dry powder. The interaction between
the process variables and the formulation parameters
is extremely complicated to be interpreted. Hence,
it is very challenging to design inhalable dry powder
with optimum inhalation efficiency using an empirical
approach alone.
There has been an appeal for a shift from empirical
formulation efforts toward rational design of spray-dried
powder based on a better understanding of the interplays
between the formulation parameters and the process
variables in the spray-drying process.[5] Several studies
have used the DoE approach to develop the ­spray-drying
process of different materials. In the preparation of
spray-dried acetaminophen particles, a fractional facto-
rial screening design was built and the results showed
that inlet temperature and feed rate of the solution are
important factors influencing powder properties.[6]
Another study used a full factorial design for the optimi-
zation of spray-dried microparticles, and a considerable
improvement of product yield with minimization of the
residual moisture content was achieved.[7] Maltesen et al.
reported a central composite face-centered (CCF) design
could be created to investigate characteristics of spray-
dried insulin powder.[8]
The aim of the current study was to apply DoE to the
spray-drying process to design inhalable dry powder
with optimum physicochemical properties. Following a
 Pharmaceutical Development and Technology
characterization of the particles with regard to particle
size, shape, residual moisture content and solid state
analysis, a full factorial design was performed as a screen-
ing to estimate the impact of three process variables (inlet
temperature, atomization air flow rate and feed flow
rate) on powder characteristics and yield. Subsequently,
a composite face-centered design was applied to map
the optimal process variables for producing spray-dried
powder for inhalation.
Materials and methods
Materials
D (−)-mannitol and D (+)-trehalose dihydrate were
purchased from BDH Prolabo (Leuven, Belgium). Spray-
drying was performed using 100 mL of an aqueous
solution containing 10 g of a mixture of mannitol and
trehalose in mass ratio 90/10. Products were stored in
closed vials under desiccated conditions at 5% relative
humidity at room temperature.
Methods
Experimental design
In the first screening step, a two-level full factorial design
with three factors was used to evaluate the effects of dif-
ferent process parameters on particle size, residual mois-
ture content, density and yield. The outlet temperature
was not included as process parameter, but its value was
measured and treated as a response. In general, the out-
let temperature is strongly dependent on the inlet tem-
perature and is therefore not included in the modeling
approach. The process parameters investigated were the
inlet temperature (110–220°C), the atomization air flow
rate (7.3–29.1 L/min) of the atomization air and the solu-
tion feed flow rate (2.5–7.5 mL/min). Between the low
and high levels, a center-point was added in triplicate to
evaluate the validity of the models, which led to a total
number of 11 experiments.
In the second step, the factorial design was upgraded
to CCF design in order to optimize the fitting of the data.
This design consisted of 17 experiments. In addition to
the 11 points of the full factorial design, 6 star points were
added. The setup and the evaluation of the DoE were
performed using Modde 8.0.2.0 and Simca-P+ 12.0.1
software (both: Umetrics, Umeå, Sweden).
Spray-drying
A Büchi Mini Spray Dryer B-290 (Büchi Labortechnik
AG, Flawil, Switzerland) with a 0.7-mm two fluid nozzle
was used. The solution was sprayed in a co-current flow
with air as drying medium. The spray-drying param-
eters were varied as stated in section “Experimental
design.” The spray-dried particles were separated from
the drying air by an improved cyclone.[9] The aspirator
rates were set at 100% in all experiments, leading to a
drying air flow of approximately 35 m3/h. Spray-dried
powders were collected, weighed and stored in capped
glass vials.

Analytical techniques
Particle size measurement by laser diffraction.  The par-
ticle size distribution of the dry powders was measured
using a laser diffractometer Mastersizer 2000 connected
with a Scirrocco 2000 powder feeder (both: Malvern
Instruments, Malvern, UK). For the measurement of
the particles in air, a dispersion pressure of 1 bar was
used. The particle size distribution is characterized by
the diameters d(0.1) and d(0.5), as well as by the span. Each
sample was measured in triplicate.
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
Scanning electron microscopy.  The morphologies of the
spray-dried powders were examined by scanning electron
microscopy (SEM) using a JSM-5200 (JEOL Ltd., Tokyo,
Japan). Before the observation, samples were sprinkled
on aluminum stubs with adhesive carbon tape. The
samples were coated with a thin layer of gold by using a
gold sputter coater (E5200 Auto Sputter Coater; Bio-Rad,
Hemel Hempstead, Hertfordshire, UK).
Near-infrared spectroscopy analysis.  Each sample was ana-
lyzed using a Bomem MB Series Fourier-Transform-Near
Infrared Spectrometer MB-160 (ABB Bomem, Quebec,
Canada) in reflectance mode. Directly after spray-drying,
samples were measured in triplicate through the bottom
of a glass vial and rotated between the measurements.
Spectra were recorded over the range 4000–8000 cm−1.
For personal use only.
A resolution of 8 cm−1 was used and each spectrum was
an average of 32 scans. Data acquisition was performed
using GRAMS software (version 7.00/LT; Thermo Fisher
Scientific, Waltham, MA).
Data analysis.  Principle component analysis (PCA) and
partial least square projection to latent structure (PLS)
were used in two aspects of the investigation: First, the
correlation of near-infrared spectroscopy (NIR) spectra
to the particle size obtained by laser diffraction and sec-
ond the evaluation of the design space.
PCA and PLS are projection methods, which are used
to reduce the dimension in a multivariate data set and
detect the influence of variables on the variation within
a data set. The observations are projected in a multi-
dimensional space and are represented on a lower
dimensional space. This is done by finding the direction
of most variation in a data set which is then defined as
the first principal component (PC). If the first PC does
not explain a sufficient amount of variation, a second PC
can be added, which will be orthogonal to the first PC.
Additional PCs can be added if necessary, however, one
has to be careful not to over-fit the data.
In the case of NIR spectroscopy each wavenum-
ber will result in a variable, resulting in several hun-
dred variables. Due to the connection between the
wavenumbers, this amount of variables can easily be
reduced to few PCs by the application of PCA. PCA was
used to gain an overview of the data. PLS is a projec-
tion method related to PCA, but is using two matrixes.
Contrary to PCA, PLS does therefore not try to obtain
© 2012 Informa Healthcare USA, Inc.
QbD in spray-drying  391
the best fit for one data matrix, but rather calculates a
good fit, which also yields a high degree of co-variation
between the two matrixes. PLS was used for connect-
ing the information of the NIR spectra to the particle
size. In both cases, the scores show the clustering of
the data, whereas the loading (PCA) or weight (PLS)
will give the reason for the clustering by connecting the
projected data to the original data. Briefly summarized,
PCA shows the relationship among the observations,
whereas PLS is used to predict a set of Y-variables based
on a set of X-variables.
Before the analysis, different pretreatment can be
applied. In the evaluation of the NIR spectra, data were
mean centered in order to improve the interpretability of
the model, whereas no other scaling was applied as all
X-variables possessed the same unit. For the evaluation
of the process parameters, the variables were centered
and unit variance scaled due to the large variation in the
numerical values. In the optimization of CCF design–
based models, factors were excluded based on their
significance in the variable importance plot (VIP), the
correlation coefficient plot and their influence on the
predictability of the model.
Thermogravimetric analysis.  The residual moisture
content of the samples was investigated directly after
spray-drying by using a TGA 7 (Perkin Elmer, Norwalk,
CT). Powder samples between 3 and 7 mg were loaded
onto a platinum sample pan and heated from 25 to
150°C at a rate of 10°C/min.
X-ray powder diffraction analysis.  The X-ray powder
diffractograms were measured at ambient conditions
on a PANalytical X’Pert Pro diffractometer equipped
with a PIXcel detector (PANalytical B.V., Almelo, The
Netherlands). Samples were placed on a zero-background
silicon holders for the measurements, and a continuous
2θ scan was performed in a range of 2–40° using a Cu
Kα radiation (λ = 1.54187 Å). The Kβ radiation was elimi-
nated by a nickel filter. The voltage and current applied
were 45 kV and 40 mA, respectively. Each measurement
was done with a step size of 0.0130°2θ and at a speed
of 0.0336°/s. Sample spinning was employed during
measurements to avoid preferred orientation effects.
Data were collected using X’Pert data collector version
2.2 and were analyzed with X’Pert Highscore plus ver-
sion 2.2.4 (both from PANalaytical B.V., Almelo, The
Netherlands). In order to identify the solid state, diffrac-
tion patterns of samples were compared with literature
patterns.[10,11]
Results and discussion
Physical characterization of the obtained particles
As the first step in the evaluation of the production pro-
cess, several critical quality attributes were investigated.
X-ray powder diffraction (XRPD) was applied to
investigate the solid state of mannitol and trehalose
 392  A. Baldinger et al.
in the spray-dried samples and a typical diffractogram
is shown in Figure 1. Characteristic reflections for the
major component β-mannitol were observed at 10.5°,
14.6° and 16.8°2θ in the XRPD pattern, whereas no
reflections corresponding to trehalose were observed,
demonstrating that spray-drying of mannitol and
trehalose resulted in crystalline mannitol and mostly
amorphous trehalose. Similar diffractograms were
observed for all samples regardless of the process
parameters. Only in a single experiment, an additional
peak at 13.8° was observed which might be due to
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
minor crystallization of trehalose or the formation of
α-mannitol.
The characterization of all powders included the
analysis of the residual moisture by a thermogravimetric
analysis directly after the spray-drying process. According
to the TGA data, all samples had a residual moisture
content below 0.5%, excluding one sample in which 1%
of moisture was obtained. In addition, the water content
of the samples was also investigated by NIR. Applying
a model using standard normal variate corrected NIR
spectra, which was developed for the quantification of
Trehalose dihydrate
For personal use only.
Counts
β-Mannitol
Spray-dried sample
10 15 20 25
Position [degree 2θ]
Figure 1.  XRPD pattern of a typical spray-dried sample in
comparison with β-mannitol and trehalose dihydrate reference.
Figure 2.  SEM pictures of spray-dried particles, inlet temperature 110°C (left side) and 220°C (right side).
 Pharmaceutical Development and Technology
water in freeze-dried samples,[12] confirmed the results
from the TGA analysis.
The morphology of the particles was investigated
by SEM (Figure 2). SEM pictures showed that spherical
smooth particles were obtained at low inlet temperature.
Due to a faster evaporation rate, increasing the inlet tem-
perature led to a more uneven and rough surface and
agglomeration could be observed.
In the following, the applicability of NIR for particle size
analysis and related density analysis was investigated. It
is known that the signal obtained via diffuse reflectance
NIR spectroscopy contains information of both chemical
information as well as physical information. The physi-
cal information obtained is influenced by various factors
such as the particle size, the density, the refractive index
(influenced by the energy of the light beam) and the
energy of the light beam itself. Mathematical pretreat-
ment such as derivatization or standard normal variate
transformation is often applied to reduce the physical
influence on the spectra.[13,14] As the aim was to investi-
gate the physical influence, it was decided to correlate
the untreated data with the particle size. An example of
typical NIR spectra is given in Figure 3A.
First, a PLS model was built on the range 4000–
8000 cm−1. The correlation between the observed size
measured by laser diffraction and the prediction based
on the NIR data showed a slope of 1.04 (standard error
0.04), a y-intercept of 2.3 µm with an R2 value of 0.953.
The root mean square error of estimation (RMSEE) of the
multivariate analysis was found to be 3.8 µm. The weight
plot showed peaks which are characteristic for chemical
information (data not shown). Due to the high content
of chemical information, an area with apparent stronger
physical information was selected, in order to focus on
the physical effects.
For this purpose, a shorter wavenumber range with
30
lower chemical information was investigated. Based on
visual inspection of the spectra, the range 5362–5905 cm−1
was chosen. The weight plot presented in Figure 3B rep-
resents the correlation between the particle size and the

spectra, that is to say the differences in the spectra, which
are connected to the particle size. The majority of the
variations can be seen to be a baseline offset, with the first
component explaining 98% of the variation of the data.
The comparison between the particle size that was
measured by laser diffraction vs. the particle size pre-
dicted by the NIR spectra is shown in Figure 4. The
baseline offset is related to the particle size as linearity in
the model can be seen. For the correlation between the
observed and predicted sizes, a slope of 1.00 (standard
error 0.03) and a y-intercept of 0 µm with an R2 value of
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
QbD in spray-drying  393
present a systematic overview over the relevant factors.
These diagrams are usually developed by a group of
experts applying and combining existing knowledge. The
complexity of the diagram will depend on the complex-
ity of the process and the desired degree of knowledge.
The Ishikawa diagram that was the basis for the current
investigation is given in Figure 5. The parameters which
were decided to be investigated, i.e. inlet temperature of
the process air, atomization air flow rate of the atomiza-
tion air and feed flow rate of the solution are marked in
bold italic script.

0.975 were achieved. The multivariate analysis reported
a RMSEE of 1.8 µm, showing the feasibility of NIR spec-
troscopy in the prediction of the particle size of the spray-
dried particles.
Experimental design
Ishikawa diagram (screening step)
The first step of applying the QbD principle to pharma-
ceutical manufacturing is to identify critical factors which
can influence the properties of the product. For this pur-
pose, Ishikawa diagrams (also called fishbone diagrams)
A whereas the feed flow rate only showed a marginal, not
In order to investigate the influence of the identified
process parameters on the quality of the end product,
DoE was applied. For the first screening step, a full fac-
torial design was applied, whereas a CCF design was
applied to a closer investigation. The process parameters
are shown in Table 1.
Full factorial design
For the initial screening, a full factorial design was applied.
Based on the coefficient plot, the inlet temperature and
the atomization air flow rate could be identified as the
critical process parameters for the various responses,

40
For personal use only.

Particle size (laser diffreaction) [µm]

30
Absorbance

20

10

4000 5000 6000 7000 8000

Wavenumber [cm−1] 0
0 10 20 30 40
B 0.14 Particle size (NIR) [µm]

0.12 Figure 4.  Comparison between particle size measured by laser

Weight PLS component 1

diffraction and predicted particle size based on NIR.

0.10
Process air Analytical Method
0.08
Temperature Sample preparation
0.06 Humidity
Drying air flow rate Measurement type
0.04 Atomization air flow rate
Gas composition Spray-dried
0.02 powder
Solvent type properties
0.00 Nozzle type
5400 5500 5600 5700 5800 5900 Solute concentration
Spray-dryer dimensions
Wavenumber [cm−1] Temperature
Feed rate Operator
Figure 3.  (A) NIR spectra of various spray-dried batches. The
gray area marks the wavenumber range 5362–5905 cm−1 which Feed solution Environment
was used for the optimized model for particle size determination.
(B) Weight plot of the first PLS component. Wavenumber range Figure 5.  Ishikawa diagram presenting causes which may affect
5362–5905 cm−1. the quality of the spray-dried powder.

© 2012 Informa Healthcare USA, Inc.

 394  A. Baldinger et al.
significant influence (data not shown). However, a poor
quantitative connection between the observed and the
predicted values was obtained for all responses, show-
ing the limitations of the factorial design. In all cases, the
three center points were away from the regression line,
as can be seen exemplified in Figure 6 which shows the
observed vs. predicted values for the particle size. The
observed phenomenon is due to non-linear influence
of one or several factors and can not be modeled cor-
rectly with a factorial design. It thus became necessary to
extend the experimental setup toward a CCF design.
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
also non-linear influence, as a quadratic polynomial
function permits to estimate curvature. Generally, a bet-
ter correlation was achieved for all factors as shall be
discussed in detail in the following sections. When cor-
relating the process parameters with the responses, two
alternatives can be applied. First, a model can be built for
each variable in order to optimize the parameters for this
variable (PLS 1) or, second, an overall model can be built
(PLS 2) to evaluate the whole picture without optimiza-
tion for each single variable.

Particle size.  Laser diffraction was applied to measure

CCF design the particle size which is a major quality attribute of inhal-
In order to cope with the short-comings of the full fac- able formulations. The descriptors in the experimental
torial design and to model non-linear influence, a CCF setup varied between 0.6 and 7.4 µm for the d(0.1), 1.5 and
model was built in a second step. Six additional experi- 35.3 µm for the d(0.5) and 1.19 and 6.97 for the span. In the
ments were added to the previous experiments from the evaluation of the span, one outlier had to be excluded.
full factorial design. The CCF allows modeling linear but The optimization of the models shall be explained
in more detail for the particle size model for d(0.5). All
responses were treated in the same manner. In the opti-
Table 1.  Spray-drying variables.
mization of the model, factors which do not show a sig-
Inlet Solution feed Atomizing air
Experiment temperature flow rate flow rate
nificant influence are removed in order to improve the
number (°C) (mL/min) (L/min) quality of the model. This exclusion is performed based on
1 110 2.5 7.3 the VIP and the coefficient plot (Figure 7A and 7B). It can
2 220 2.5 7.3 be seen that all interactions between the three primary
3 110 7.5 7.3 parameters and the non-linear term for the atomization
For personal use only.

4 220 7.5 7.3

5 110 2.5 29.1 A 2.0
6 220 2.5 29.1
7 110 7.5 29.1 1.5
8 220 7.5 29.1
9 110 5 13.8
10 220 5 13.8 1.0
VIP

11 165 2.5 13.8

12 165 7.5 13.8 0.5
13 165 5 7.3
14 165 5 29.1
15-17 165 5 13.8 0.0
Experiment numbers 1–8 and 15–17: full factorial design.
Experiment numbers 1–17: composite face-centered design.
−0.5 Ti FR * FR Ti * Ti AA FR

B 1.0
35
Observed particle size [µm]

30
0.5
25
Coefficient

20
0.0
15

10

5 −0.5

0 Ti FR AA Ti * Ti FR * FR
0 5 10 15 20 25 30 35
Predicted particle size [µm] Figure 7.  (A) Variable importance plot (VIP) for the particle size
based on CCF. (B) Coefficient plot for the particle size based on
Figure 6.  Correlation between observed vs. predicted particle size CCF. AA, atomization air flow rate; FR, feed flow rate; Ti, inlet
based on full factorial design. temperature.

 Pharmaceutical Development and Technology


1.0
D0.5
0.8
Ti
Weights PLS component
0.6
FR * FR
0.4
Ti * Ti
0.2
0.0
FR
−0.2
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
−0.4 AA
Figure 8.  Weight plot for the CCF model particle size d(0.5). AA,
atomization air flow rate; FR, feed flow rate; Ti, inlet temperature.
air flow rate were found to be non-significant and have
been removed from the model. The feed flow rate shows
no significance in the VIP, however, the quadratic term
investigating a non-linear influence of the feed rate does.
It is therefore necessary to keep the primary parameter
in the model. The atomization air flow rate does not show
significant influence in the coefficient plot. However, it
does have an influence on the quality of the model, as can
be seen in the VIP, and therefore remained in the model.
Using this setup, the optimal model with an R2 (fit) of
For personal use only.
75.6% and a Q2 (predictability) of 61.9% was achieved.
The influence of the single parameters can either be
seen in the VIP or be taken directly from the weight plot
achieved by the PLS (Figure 8). In both cases it can be
seen that the diameter increases with increased inlet
temperature. Furthermore, the inlet temperature shows
a positive non-linear influence, which means that the
particle size will not increase in a linear manner but over
proportional with increasing inlet temperature. The feed
flow rate and the atomization air flow rate show small
negative influences, meaning that an increase in either
of them will lead to a slight reduction in the particle size.
Based on the coefficient plot, the feed flow rate shows a
small, negative and non-significant influence, however,
the quadratic term shows a positive effect. This means
that there is a curvature in the parameter feed flow rate,
which, however, is not affecting the parameter itself
significantly in the area investigated (2.5–7.5 mL/min).
Considering the nature of curvature, one can, however,
expect that the factor feed flow rate in itself will become
significant at higher values than the one investigated
here.
The quality of the model can finally be evaluated by
comparing the predicted and the observed diameter of
the particles, which is shown in Figure 9. In the current
model, a RMSEE of 5.0 µm was found for d(0.5).
With regard to d(0.1), the same factors as described
above for the d(0.5) were found to be significant and show
the same type and strength of impact on the model. The
model can be described by an R2 of 69.1%, a Q2 of 56.2%
and a RMSEE of 1.1 µm. For the span, the inlet tempera-
ture and the quadratic term of the feed flow rate showed
© 2012 Informa Healthcare USA, Inc.
QbD in spray-drying  395
40
Observed particle size [µm]
30
20
10
0
0 10 20 30 40
Predicted partile size [µm]
Figure 9.  Observed vs. predicted particle size d(0.5) by CCF.
a positive influence on the span, whereas the feed flow
rate itself showed a non-significant influence just as in
the models above. The quality of the model is described
by an R2 of 75.2%, a Q2 of 62.7% and a RMSEE of 1.0.
All models thus show a Q2 above 50% and a difference
between R2 and Q2 of less than 20% and are thus judged
to be good models for the description of a preparative
operation.
Other critical quality attributes.  Another variable in
the quality assessment of powders for inhalation is the
tapped density. The absolute values varied between
0.312 and 0.628 g/cm3, however, the majority (13 out
of 17 samples) were found in the range 0.419–0.493 g/
cm3. This concentration of the samples in middle
ranges with only few samples at extreme values can
be assumed to hamper the model. Both atomization
air flow rate and feed flow rate were found to have a
negative effect on the tapped density of the samples,
whereas the only positive influence was found in the
inlet temperature. Although this influence was not sig-
nificant according to the VIP and the coefficient plot,
a clear decrease in model quality was observed when
excluding the inlet temperature. Neither interactions
nor non-linear terms showed any significance. The
model can be described by an R2 of 67.9%, a Q2 of 45.9%
and a RMSEE of 0.04 g/cm3, rendering it an acceptable
model.
In the pharmaceutical industry, one factor of extreme
importance is the yield of a process, calculated by the
ratio between the obtained powder and the expected
amount of powder. The yield varied between 7.7 and 91%
in this study.
The critical parameters which have a statistically sig-
nificant influence on the yield are shown in the coeffi-
cient plot (Figure 10). The model showed an R2 of 75.6%
and a Q2 of 66.8% and can thus be judged as good. For
the observed vs. prediction, an RMSEE of 16.2% was
found.
According to the coefficient plot, the inlet temperature
has a negative influence on the yield and the curvature
 396  A. Baldinger et al.
Density
0.5 0.6

0.4
D0.5

Weight PLS component 2

0.2 D0.1

Yield
Coefficient

0.0 0.0
FR * FR Span
−0.2 Ti

−0.4
FR
−0.5 −0.6
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13

−0.8
AA
Ti FR AA TI * Ti FR * FR −1.0
−1.0 −0.8 −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0
Figure 10.  Coefficient plot for the yield based on CCF. AA, weight PLS component 1
atomization air flow rate; FR, feed flow rate; Ti, inlet temperature.
Figure 11.  Weight plot of PLS 2 covering all responses. Closed
squares represent input factors, open circles represent responses.
indicates that this effect is getting overly pronounced
AA, atomization air flow rate; FR, feed flow rate; Ti, inlet
when increasing the temperature. An increase in the temperature.
inlet temperature will decrease the yield, which can be
explained as follows: with the rather low feed flow rate
7.5 30
applied and an aspirator setting kept constant at 100%,
an increase in the inlet temperature will lead to a sub- 7.0
60
stantial increase in the outlet temperature. Depending 6.5
on the solid state of the materials, the outlet temperature 6.0
will influence the physical properties of the compounds.
5.5
Feed rate

As shown in Figure 1, mannitol will remain crystalline

For personal use only.

upon spray-drying and thus show a melting point in the 5.0

range 165–169°C, whereas trehalose will be amorphous.
The glass transition temperature of trehalose has been
reported to be around 107°C.[15] When the inlet tempera-
ture was set at 110°C, the outlet temperature was found to
be maximum 62°C, that is to say under the glass transition
temperature of trehalose. When the inlet temperature
was set at 220°C, the outlet temperature was 117.5°C at its
highest, which is above the glass transition temperature of
trehalose. Thus, trehalose would be in the rubbery state.
Therefore, agglomeration occurs, which may explain the
decrease in yield as some sticky powder agglomerated in
the cyclone and did not reach the collection vessel.
Overall model and predictive abilities.  In addition to an
optimization of the model with regard to the influence
of the factors on single responses, an overall model (also
called PLS 2), which models all responses, is possible. In
this case, two PLS components were necessary for the
overall model and an R2 of 69.6% and a Q2 of 46.8% were
found. These values are slightly lower than the corre-
sponding values found for single-response models, which
can be expected as the factors influence the responses in
a different manner and are not optimized for a specific
response. The advantage of the overall model is its useful-
ness in the improvement of several responses at the same
time. A clear overview of the connection between the fac-
tors and the responses can be seen in the weight plot of
the PLS 2 where both factors and responses are included
(Figure 11). In the overall model, only the inlet tempera-
ture, the atomization air flow rate and the quadratic term
4.5
4.0 90
80
70 50
3.5
40
3.0
2.5 20
110 120 130 140 150 160 170 180 190 200 210 220
Inlet temperature
Figure 12.  Contour plot for the prediction of the yield based on
the inlet temperature and the feed rate calculated by the PLS 2
model. The values show the predicted yield in %.
of the feed flow rate showed a significant influence. The
obtained RMSEE values for the various responses in the
overall model were as follows: d(0.1): 1.1 µm; d(0.5): 4.7 µm;
span: 1.2; tapped density: 0.04 and yield: 20.0%. In sum-
mary, by applying a CCF model it was shown that inlet
temperature and atomization air flow rate are the most
significant parameters influencing the yield and the par-
ticle size of the spray-dried powder. The particle size was
increased with increasing inlet temperature and by low
atomization air flow rate. The yield was negatively influ-
enced mainly by the inlet temperature.
In order to visualize predicted values for unknown
conditions, contour plot can be applied. As an example, a
contour plot is given for the prediction of the yield based
on the PLS 2 model (Figure 12). This plot clearly shows
the influence of the inlet temperature on the yield and
that of the curvature on the feed flow rate. After the estab-
lishment of the models, an easy alternative to optimize

 Pharmaceutical Development and Technology


the process conditions can be found by comparing the
plots for the various responses by visual evaluation of the
obtained contour plots.
Conclusion
In this study, the benefits of DoE for the settings of the
spray-drying process were analyzed. The spray-dried
powders were initially characterized and particles within
the inhalable range with a low density and with a high
yield were produced. Particle with spherical and smooth
Pharmaceutical Development and Technology Downloaded from informahealthcare.com by National University of Singapore on 06/22/13
morphology could be obtained with low inlet tempera-
tures according to SEM pictures. Water content analysis
of the powders by TGA and NIR showed that the residual
moisture was usually less than 1%. The use of XRPD
showed that spray-drying of mannitol and trehalose
resulted in crystalline (β-)mannitol and amorphous tre-
halose. It was also shown that an acceptable correlation
between the particle size measured by laser diffraction
and calculation based on NIR data was achieved. This
approach provides a non-invasive approach for process
monitoring and control purposes.
A full factorial design proved to be insufficient to
describe the effects of the process parameters on the
outcomes. This is due to the non-linear influence of
the process parameters. Therefore, a CCF design was
applied to predict particle size and yield. This design was
For personal use only.
more suitable and shed light on the need of a strict con-
trol of inlet temperature and atomization air flow rate.
Although the particle size was influenced by both inlet
temperature as well as atomization air flow rate, the yield
mainly depended on the inlet temperature. It can be
concluded, that DoE can be used in the rational choice
of spray-drying conditions and to predict the spray-dried
product attributes, based on the spray-drying process
conditions.
Declaration of interest
The authors report no declarations of interest.
© 2012 Informa Healthcare USA, Inc.
QbD in spray-drying  397
References
1. International Conference on Harmonisation of Technical
Requirements for Registration of Pharmaceuticals for Human Use.
Topic Q8(R1): Pharmaceutical Development, 2008.
2. Naelapää K, Veski P, Kristensen HG, Rantanen J, Bertelsen P.
Building quality into a coating process. Pharm Dev Technol
2010;15:35–45.
3. Montgomery DC. Design and Analysis of Experiments. New York:
John Wiley, 2005.
4. Hickey AJ. Pharmaceutical Inhalation Aerosol Technology. New
York: Marcel Dekker Inc, 1996.
5. Vehring R. Pharmaceutical particle engineering via spray drying.
Pharm Res 2008;25:999–1022.
6. Billon A, Bataille B, Cassanas G, Jacob M. Development of spray-
dried acetaminophen microparticles using experimental designs.
Int J Pharm 2000;203:159–168.
7. Tewa-Tagne P, Degobert G, Briançon S, Bordes C, Gauvrit
JY, Lanteri P et  al. Spray-drying nanocapsules in presence of
colloidal silica as drying auxiliary agent: formulation and process
variables optimization using experimental designs. Pharm Res
2007;24:650–661.
8. Maltesen MJ, Bjerregaard S, Hovgaard L, Havelund S, van de
Weert M. Quality by design - Spray drying of insulin intended for
inhalation. Eur J Pharm Biopharm 2008;70:828–838.
9. Maury M, Murphy K, Kumar S, Shi L, Lee G. Effects of process
variables on the powder yield of spray-dried trehalose on a
laboratory spray-dryer. Eur J Pharm Biopharm 2005;59:565–573.
10. Burger A, Henck JO, Hetz S, Rollinger JM, Weissnicht AA, Stöttner
H. Energy/temperature diagram and compression behavior of the
polymorphs of D-mannitol. J Pharm Sci 2000;89:457–468.
11. Nunes C, Suryanarayanan R, Botez CE, Stephens PW.
Characterization and crystal structure of D-mannitol hemihydrate.
J Pharm Sci 2004;93:2800–2809.
12. Grohganz H, Fonteyne M, Skibsted E, Falck T, Palmqvist B,
Rantanen J. Role of excipients in the quantification of water in
lyophilised mixtures using NIR spectroscopy. J Pharm Biomed
Anal 2009;49:901–907.
13. 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.
14. Rinnan A, van den Berg F, Engelsen SB. Review of the most common
pre-processing techniques for near-infrared spectra. Trends Anal
Chem 2009;28:1201–1222.
15. Simperler A, Kornherr A, Chopra R, Bonnet PA, Jones W,
Motherwell WD et  al. Glass transition temperature of glucose,
sucrose, and trehalose: an experimental and in silico study. J Phys
Chem B 2006;110:19678–19684.