Journal of Pharmaceutical and Biomedical Analysis 114 (2015) 208–215
Contents lists available at ScienceDirect
Journal of Pharmaceutical and Biomedical Analysis journal homepage: www.elsevier.com/locate/jpba
A PAT-based qualification of pharmaceutical excipients produced by batch or continuous processing A. Hertrampf a,c , H. Müller b , J.C. Menezes c , T. Herdling a,∗ a b c
Quality Operations, Laboratory for Liquid Products, Merck Serono, Frankfurter Str. 250, 64293 Darmstadt, Germany Operations, Laboratory for Process Development, Merck Millipore, Frankfurter Str. 250, 64293 Darmstadt, Germany Institute for Biotechnology and Biosciences, Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais, 1049-001 Lisbon, Portugal
a r t i c l e
i n f o
Article history: Received 30 January 2015 Received in revised form 29 April 2015 Accepted 14 May 2015 Available online 1 June 2015 Keywords: Pharmaceutical excipients Pharmaceutical engineering Process analytical technology Multivariate data analysis Quality control strategy Supplier qualification
a b s t r a c t Pharmaceutical excipients have an influence on the main requirements for medicinal products (viz., quality, safety and efficacy) but also on their manufacturability. During product lifecycle it may become necessary to introduce minor changes (e.g., to continuously improve it) or major changes in the validated process (e.g., moving it to a new production site, replacing process version or even disruptively changing processing type). Those changes can influence the critical to quality attributes of the product. Therefore, it is important to enhance process understanding to avoid the risk of any significant quality changes. Process analytical technology can support better decision making and risk-management as required in quality by design – viz., by many pharmaceutical regulatory authorities. This study compares the quality of the pharmaceutical excipient sodium carbonate (anhydrous) produced either in a batch or a continuous process. For continuous processing two different production lines were available that differed on the dryer and crystallizer types used. Therefore their influence on critical to quality attributes of sodium carbonate was investigated for each of the three processing alternatives. The overall goal was to identify which of the continuous processes ensures a similar product quality to batch processing. Namely, changes on chemical and physical attributes of the product were investigated with Raman spectroscopy, laser diffraction and X-ray powder diffraction. Principal component analysis, a very common multivariate analysis technique, was applied to extract relevant information from small differences at multiple spectral regions from samples from each process type and from each analytical technique used. Changing processing from batch to continuous improved consistency of certain attributes (e.g., particle size distribution) but affected others. However, the increased process/product knowledge gained can lead to an enhanced control strategy and ensure a similar product quality is obtained from distinct process versions. © 2015 Elsevier B.V. All rights reserved.
1. Introduction The requirements for the release of medicinal products are very high in terms of pharmaceutical quality, safety, and efficacy [1].
Abbreviations: MVA, multivariate data analysis; PSD, particle size distribution; CQA, critical to quality attribute; PCA, principal component analysis; PAT, process analytical technology; QbD, quality by design; MSPC, multivariate statistical quality control; LDA, linear discriminant analysis. ∗ Corresponding author at: Merck KGaA, Frankfurter Str. 250, Postcode PH028/001, 64293 Darmstadt, Germany. Tel.: +49 06151 72 8569; fax: +49 06151 72 3155. E-mail addresses: [email protected] (A. Hertrampf), [email protected] (H. Müller), [email protected] (J.C. Menezes), [email protected] (T. Herdling). http://dx.doi.org/10.1016/j.jpba.2015.05.012 0731-7085/© 2015 Elsevier B.V. All rights reserved.
Medicinal products contain active pharmaceutical ingredients (API) and several excipients. The excipients perform different functions in the formulation. For example, they can be used as solvent for the API, formulation filler, coating material or for flavour masking, among other functions. Depending on their type, concentration, and characteristics, excipients affect processability and the drug product functions (stability, solubility, and bioavailability of the API in medicinal products) [2,3]. Therefore, excipients impact general requirements for medicinal products through several of their quality attributes. Changes of a production process can affect those CQAs. Excipient manufacturers have the duty to inform about such changes, as they can potentially impact processability as well as the quality of the final drug product. One possible way to ensure the necessary consistency of CQAs over a process lifecycle, is through a high-level of process understanding obtained – viz., QbD as described in ICH Q11 [4]. Within such approaches process units
A. Hertrampf et al. / Journal of Pharmaceutical and Biomedical Analysis 114 (2015) 208–215
with a high effect on the CQAs are identified. Then a comprehensive effort in product and process characterization, comparing chemical and physical CQAs between “old” and “new” lots is carried out. In general many analytical measurements are performed and the results evaluated from a univariate point of view. However, most analytical techniques used can provide more information than that captured by analysing single attributes. Therefore it can be very helpful to use multivariate data analysis techniques [5,6]. In our study, a risk analysis based on ICH guideline Q9 was performed as FMEA. An interdisciplinary team identified critical process units and parameters for the anhydrous sodium carbonate produced by each of the three different manufacturing lines. Due to confidentiality and proprietary aspects, the risk analysis is not described herein. Sodium carbonate is an excipient that is mainly used as alkaline component in effervescent tablets. The crystallization and drying steps were identified as the most critical process steps because of a high influence on the excipients particle size, shape, and surface. The criticality of crystallization steps was also identified by Rantanen [7] and Datta et al. [3]. One of the manufacturing lines is operated in batch, while the other two are continuous processes but with different equipment as mentioned earlier. To identify possible chemical and physical changes, Raman spectroscopy, X-ray powder diffraction and laser diffraction were performed on finished excipient samples. Raman spectroscopy is a rapid and non-invasive technique, capturing chemical as well as physical attributes of the material [7–10]. Laser diffraction was used to determine the volume-based particle size [11]. Although two laser sources are used – viz., blue to measure small particles and red to detect larger particles, the method is not capable to discriminate between crystals and crystal agglomerates [12]. Consistency of particle size distribution is very important. Smaller particles have a higher surface to volume ratio than larger particles, which can affect dissolution kinetics. Mixing and segregation are also particle size dependent. X-ray powder diffraction is used to determine the structure of molecules. Differences between amorphous and crystalline forms as well as between polymorphic forms of a substance [13] are detectable. Crystallinity affects processing through differences in flowability, compressibility, and solubility thus affecting bioavailability of the medicinal product [3,14]. To determine the crystallite size, the reflex broadening was calculated by applying a Rietveld refinement [15]. Therefore the amount of reflections underneath each peak is relevant. Crystallites are very small crystals, which form conglomerates [16]. Thus, low crystallite size does not mean low crystallinity. The particles surface texture can be important concerning water resorption and therefore stability, dissolution, and ultimately bioavailability. Particles shape can also influence powder flowability. The influence of the crystallization process (batch and continuous crystallizers) on particle size and shape distribution as well as on crystal form was analyzed by Vetter et al. [17]. The importance of particle size is also known for its influence on powder colour intensity [18,19], stability, bioavailability, and dissolution [20–22]. Yu et al. [23] presented a strategy to control crystallite size and shape during crystallization processes by using process analytical technology. Particle shape and surface differences are only detectable with microscopy techniques. For example, scanning electron microscopy (SEM) has the resolution and depth of focus necessary [24]. Suvakanta et al. [25] described the characterization of a purified polysaccharide with different analytical methods, namely SEM, XRPD, differential scanning calorimetry, Fourier-transform infrared spectroscopy, and nuclear magnetic resonance spectroscopy. But they did not analyze these results with multivariate techniques. Balcerowska-Czerniak et al. [26] applied multivariate analysis, namely principal component analysis and multivariate curve resolution, on XRPD patterns. The focus was to enhance the information of XRPD and to characterize the substance in detail.
209
Pizarro et al. [27] deals with the geographical traceability of olive oils. The focus is on the variance within natural products caused by weather, soil etc. and not on the oil extraction process. The goal of this publication is to increase the effectiveness of analytical techniques by applying multivariate data analysis. Therefore two data matrices were concatenated and analyzed using PCA and LDA. The first data matrix, called quality matrix, contained the results of physical chemical parameters. The second matrix contained the data of UV spectroscopy and is therefore named as spectral matrix. The application of MVA results in the discrimination of three different types of Spanish olive oils by their origin. Other approaches examined specific processing steps but not into the whole process and therefore do not analyze or compare the end-product from different manufacturing lines. Kumar et al. [28] reviewed the change from batch to continuous for wet granulation processes using a model-based analysis. The review showed that the shift to continuous production is challenging but has the advantage to use online measurement tools instead of off-line analyses that make real-time-release feasible. But this is only one step of a complex manufacturing process. Our study focuses on the characterization of sodium carbonate by different analytical methods followed by multivariate data analysis, then relating those findings with each of the process versions used. 2. Materials and methods 2.1. Samples The pharmaceutical excipient is the inorganic salt sodium carbonate. It was produced on three different production lines during this study, which will be referred as Processes A, B, and C. Process A is overall a batch process, and Processes B and C are continuous production lines. The dryer and crystallizer were identified as critical process units for all process versions investigated. For Processes A and B the same type of dryer is used. Process C is equipped with a dryer, working under a different drying principle than the one used by A and B. Since Processes B and C are continuous they use a different type of crystallizer from Process A. 2.2. Raman spectroscopy Raman spectra were collected using RamSys FT-Raman instrument by Bruker together with the software OPUS 6.5 (Bruker Optic GmbH, Ettlingen, Germany). The exciting source was a Nd:YAG laser emitting invisible radiation at 1064 nm. The laser light was transmitted by optical fibres to the probe. The laser output at the exit port of the probe was >1000 mW. The spectrometer was equipped with a liquid nitrogen cooled Ge detector (D418-T). The measured spectral range was between 3600 and 10 cm−1 with a 8 cm−1 resolution and averaging 10 scans. For each batch 10 spectra of the sampling point were measured to reduce the variance caused by the analytical method (i.e., different pressure of the probe on the sampled powder). The mean of spectra replicates for each batch was calculated before subsequent multivariate data analysis has been applied. 2.3. Laser diffraction To determine the particle size distribution for all three manufacturing processes the laser diffractometer Mastersizer 2000 v.5.60 with the dry powder dispersion unit Scirocco 2000 (both by Malvern Instruments Ltd., Worcestershire, UK) was used. The light sources are a He–Ne laser (632.8 nm, red, 4 mW) and a LED laser (470 nm, blue, 0.3 mW). The measured particle size range is from 0.02 to 2000 m and is displayed as a volume based distribution. Particle size is calculated using Fraunhofer approximation.
210
A. Hertrampf et al. / Journal of Pharmaceutical and Biomedical Analysis 114 (2015) 208–215
To destroy the agglomerates the dispersion pressure was 3.5 bar. A sample of 2 g was measured within 10 s.
Venetian blind cross-validation was performed with three splits. [6,29,30]
2.4. X-ray powder diffraction
3. Results and discussion
X-ray powder diffraction was performed with STOE StadiP 611 KL powder diffraction system with Mythen-PSD detector (both STOE & Cie GmbH, Darmstadt, Germany). Samples were measured in transmission with Cu-K␣1 radiation (40 kV, 40 mA). The angle range is from 1◦ to 90◦ 2 with a resolution of 0.015◦ 2. The PSD steps were 2.0◦ 2 with a speed of 15 s/step. The measurements were evaluated using the softwares Igor Pro v.6.2.1 (WaveMetrics Inc., Portland, OR, USA) and STOE WinXPow v.2.23.2 (STOE & Cie GmbH, Darmstadt, Germany). To determine the crystallite size, the reflex broadening was calculated by applying a Rietveld refinement using the software General Structure Analysis System v.12.08 (Los Alamos National Laboratory, NM, USA).
PCA was applied Based on Raman spectroscopy, laser diffraction, and X-ray powder diffraction data. As discussed below through PCA clustering and other trends of the three types of samples (Processes A, B, and C) could be identified.
2.5. Scanning electron microscopy SEM images were recorded by using the Zeiss system Supra 35 for Processes A and B and the Leo 1530, which is also by Zeiss (Carl Zeiss Microscopy GmbH, Jena, Germany) for samples from Process C. These images allowed the determination of the surface morphology. Both microscopes used a field emission cathode and an Everhardt–Thornley detector. That detector is secondary electron detector. The scans have been performed with a resolution of 2 nm. The acceleration voltage was set to 5 kV, 8 kV or 10 kV and magnification of 50- and 500-fold was chosen. For preparation the samples were applied on top of aluminium samples holders, which were covered by a double-faced adhesive and conductive foil. Afterwards the samples were coated with platinum. This layer had a thickness of 10 nm or 30 nm. 2.6. Multivariate data analysis The above mentioned analytical techniques data were analyzed by applying multivariate data analysis, using MATLAB® R2010a v.7.10.0 (MathWorks, Inc., Natick, MA, USA) and the PLS Toolbox® v.7.5 (Eigenvector Research Inc., Wenatchee, WA USA). For the Raman spectroscopy with a large wavenumber range it is possible to have bands dominating over others; therefore it can be helpful to choose narrower ranges and/or use some signal processing to improve discrimination. The wavenumber regions that provided the best discrimination were: 1460–1379 cm−1 , 781–619 cm−1 and 441–360 cm−1 . Pre-processing techniques like standard normal variate, first derivatives preceded by Savitzky–Golay smoothing or mean centring, were used. Mean centred spectra gave the best results. Cross-validation was performed using venetian blinds with ten splits. Laser diffraction particle size distributions were pre-processed by mean centring only. The whole range of the particle size distribution was used to build a MVA model. Venetian blinds with four splits were used for cross-validation. In XRPD the layer thickness of the samples can vary due to the preparation. This can affect the intensities in the diffractogram. To compensate this effect area normalization (area = 1) was applied. The peak structure, which includes the information about crystallinity and crystallite size, is not affected by the normalization used. In addition the data was mean centred before further analyses were applied. The range considered was between 22◦ and 90◦ 2. After pre-processing the data, PCA was used to project the data into lower dimensional spaces and in that way help classification from each analytical technique. Therefore PCA was used to support the exploratory and unsupervised analysis of the data (i.e., not imposing a process version or any direct process knowledge).
3.1. Raman spectroscopy The raw spectra of the products from Processes A, B, and C are almost indistinguishable. The main peak at 1078 cm−1 can be assigned to a symmetric 1 (CO) stretching vibration [31]. However a closer look at the spectra revealed differences in the smaller peaks around 1420 cm−1 , 700 cm−1 and 420 cm−1 (cf. Fig. 1). Applying PCA showed a clearer discrimination if only the specific regions around the before mentioned peaks were chosen. The 52–245 cm−1 peak probably contains information about the crystal structure. However this region was excluded. Afterwards no other chemical differences could be detected. The remaining differences within the selected regions seem to be caused by the particle size distribution (PSD) – see below. Fig. 2 displays the score (a) and loading (b) plots of the first two principal components (PCs). All batches can be described by a single PCA model, with all samples inside the 95% confidence interval. The first two principal components explain 94.64% of the samples variance. Additional principal components do not allow any discrimination of differences among process versions. The loading plot shows that all three peaks are described by PC1. The 700 cm−1 peak intensity differences explain most of the variance within PC1, while the peak around 420 cm−1 is very well captured in PC2. The 1420 cm−1 peak has almost no contribution to PC2. The two small “neighbour” peaks of the 700 cm−1 peak, which can hardly be seen on the raw spectra, are well captured and contribute significantly to PC2. The correlation is positive and therefore in the opposite direction as for the 700 cm−1 peak. Nevertheless, the score plot shows that Processes A and B are very similar to each other and that Process C can be discriminated (Processes A and B use the same type of dryer). The influence of process version (batch vs. continuous) and the type of crystallizer does not seem to be large, in terms of Raman spectra. The residuals of each sample (distance to the PCA model) are very low and non-significant. Only two Process A batches show larger residuals, meaning that these batches may be different from the remaining ones and not well described by a same PCA model. 3.2. Laser diffraction The particle size distribution in Fig. 3 shows that Processes A and B have fines in the range between 3.6 m and 11.25 m with an average size around 5.6 m. Two batches of Process B do not show these fines. There is a slight increase of particles in the range between 12 m and 100 m followed by the main particle size between 100 m and 1260 m. Process C does not show fines like the other processes but it shows a small fraction of particles between 28 m and 89 m (max. 56.4 m). But the volume (%) for this size is still lower as for Processes A and B. It should be mentioned that the Frauenhofer approximation is not completely accurate for small particles below ten times of the laser light wavelength (i.e.,
Contents lists available at ScienceDirect
Journal of Pharmaceutical and Biomedical Analysis journal homepage: www.elsevier.com/locate/jpba
A PAT-based qualification of pharmaceutical excipients produced by batch or continuous processing A. Hertrampf a,c , H. Müller b , J.C. Menezes c , T. Herdling a,∗ a b c
Quality Operations, Laboratory for Liquid Products, Merck Serono, Frankfurter Str. 250, 64293 Darmstadt, Germany Operations, Laboratory for Process Development, Merck Millipore, Frankfurter Str. 250, 64293 Darmstadt, Germany Institute for Biotechnology and Biosciences, Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais, 1049-001 Lisbon, Portugal
a r t i c l e
i n f o
Article history: Received 30 January 2015 Received in revised form 29 April 2015 Accepted 14 May 2015 Available online 1 June 2015 Keywords: Pharmaceutical excipients Pharmaceutical engineering Process analytical technology Multivariate data analysis Quality control strategy Supplier qualification
a b s t r a c t Pharmaceutical excipients have an influence on the main requirements for medicinal products (viz., quality, safety and efficacy) but also on their manufacturability. During product lifecycle it may become necessary to introduce minor changes (e.g., to continuously improve it) or major changes in the validated process (e.g., moving it to a new production site, replacing process version or even disruptively changing processing type). Those changes can influence the critical to quality attributes of the product. Therefore, it is important to enhance process understanding to avoid the risk of any significant quality changes. Process analytical technology can support better decision making and risk-management as required in quality by design – viz., by many pharmaceutical regulatory authorities. This study compares the quality of the pharmaceutical excipient sodium carbonate (anhydrous) produced either in a batch or a continuous process. For continuous processing two different production lines were available that differed on the dryer and crystallizer types used. Therefore their influence on critical to quality attributes of sodium carbonate was investigated for each of the three processing alternatives. The overall goal was to identify which of the continuous processes ensures a similar product quality to batch processing. Namely, changes on chemical and physical attributes of the product were investigated with Raman spectroscopy, laser diffraction and X-ray powder diffraction. Principal component analysis, a very common multivariate analysis technique, was applied to extract relevant information from small differences at multiple spectral regions from samples from each process type and from each analytical technique used. Changing processing from batch to continuous improved consistency of certain attributes (e.g., particle size distribution) but affected others. However, the increased process/product knowledge gained can lead to an enhanced control strategy and ensure a similar product quality is obtained from distinct process versions. © 2015 Elsevier B.V. All rights reserved.
1. Introduction The requirements for the release of medicinal products are very high in terms of pharmaceutical quality, safety, and efficacy [1].
Abbreviations: MVA, multivariate data analysis; PSD, particle size distribution; CQA, critical to quality attribute; PCA, principal component analysis; PAT, process analytical technology; QbD, quality by design; MSPC, multivariate statistical quality control; LDA, linear discriminant analysis. ∗ Corresponding author at: Merck KGaA, Frankfurter Str. 250, Postcode PH028/001, 64293 Darmstadt, Germany. Tel.: +49 06151 72 8569; fax: +49 06151 72 3155. E-mail addresses: [email protected] (A. Hertrampf), [email protected] (H. Müller), [email protected] (J.C. Menezes), [email protected] (T. Herdling). http://dx.doi.org/10.1016/j.jpba.2015.05.012 0731-7085/© 2015 Elsevier B.V. All rights reserved.
Medicinal products contain active pharmaceutical ingredients (API) and several excipients. The excipients perform different functions in the formulation. For example, they can be used as solvent for the API, formulation filler, coating material or for flavour masking, among other functions. Depending on their type, concentration, and characteristics, excipients affect processability and the drug product functions (stability, solubility, and bioavailability of the API in medicinal products) [2,3]. Therefore, excipients impact general requirements for medicinal products through several of their quality attributes. Changes of a production process can affect those CQAs. Excipient manufacturers have the duty to inform about such changes, as they can potentially impact processability as well as the quality of the final drug product. One possible way to ensure the necessary consistency of CQAs over a process lifecycle, is through a high-level of process understanding obtained – viz., QbD as described in ICH Q11 [4]. Within such approaches process units
A. Hertrampf et al. / Journal of Pharmaceutical and Biomedical Analysis 114 (2015) 208–215
with a high effect on the CQAs are identified. Then a comprehensive effort in product and process characterization, comparing chemical and physical CQAs between “old” and “new” lots is carried out. In general many analytical measurements are performed and the results evaluated from a univariate point of view. However, most analytical techniques used can provide more information than that captured by analysing single attributes. Therefore it can be very helpful to use multivariate data analysis techniques [5,6]. In our study, a risk analysis based on ICH guideline Q9 was performed as FMEA. An interdisciplinary team identified critical process units and parameters for the anhydrous sodium carbonate produced by each of the three different manufacturing lines. Due to confidentiality and proprietary aspects, the risk analysis is not described herein. Sodium carbonate is an excipient that is mainly used as alkaline component in effervescent tablets. The crystallization and drying steps were identified as the most critical process steps because of a high influence on the excipients particle size, shape, and surface. The criticality of crystallization steps was also identified by Rantanen [7] and Datta et al. [3]. One of the manufacturing lines is operated in batch, while the other two are continuous processes but with different equipment as mentioned earlier. To identify possible chemical and physical changes, Raman spectroscopy, X-ray powder diffraction and laser diffraction were performed on finished excipient samples. Raman spectroscopy is a rapid and non-invasive technique, capturing chemical as well as physical attributes of the material [7–10]. Laser diffraction was used to determine the volume-based particle size [11]. Although two laser sources are used – viz., blue to measure small particles and red to detect larger particles, the method is not capable to discriminate between crystals and crystal agglomerates [12]. Consistency of particle size distribution is very important. Smaller particles have a higher surface to volume ratio than larger particles, which can affect dissolution kinetics. Mixing and segregation are also particle size dependent. X-ray powder diffraction is used to determine the structure of molecules. Differences between amorphous and crystalline forms as well as between polymorphic forms of a substance [13] are detectable. Crystallinity affects processing through differences in flowability, compressibility, and solubility thus affecting bioavailability of the medicinal product [3,14]. To determine the crystallite size, the reflex broadening was calculated by applying a Rietveld refinement [15]. Therefore the amount of reflections underneath each peak is relevant. Crystallites are very small crystals, which form conglomerates [16]. Thus, low crystallite size does not mean low crystallinity. The particles surface texture can be important concerning water resorption and therefore stability, dissolution, and ultimately bioavailability. Particles shape can also influence powder flowability. The influence of the crystallization process (batch and continuous crystallizers) on particle size and shape distribution as well as on crystal form was analyzed by Vetter et al. [17]. The importance of particle size is also known for its influence on powder colour intensity [18,19], stability, bioavailability, and dissolution [20–22]. Yu et al. [23] presented a strategy to control crystallite size and shape during crystallization processes by using process analytical technology. Particle shape and surface differences are only detectable with microscopy techniques. For example, scanning electron microscopy (SEM) has the resolution and depth of focus necessary [24]. Suvakanta et al. [25] described the characterization of a purified polysaccharide with different analytical methods, namely SEM, XRPD, differential scanning calorimetry, Fourier-transform infrared spectroscopy, and nuclear magnetic resonance spectroscopy. But they did not analyze these results with multivariate techniques. Balcerowska-Czerniak et al. [26] applied multivariate analysis, namely principal component analysis and multivariate curve resolution, on XRPD patterns. The focus was to enhance the information of XRPD and to characterize the substance in detail.
209
Pizarro et al. [27] deals with the geographical traceability of olive oils. The focus is on the variance within natural products caused by weather, soil etc. and not on the oil extraction process. The goal of this publication is to increase the effectiveness of analytical techniques by applying multivariate data analysis. Therefore two data matrices were concatenated and analyzed using PCA and LDA. The first data matrix, called quality matrix, contained the results of physical chemical parameters. The second matrix contained the data of UV spectroscopy and is therefore named as spectral matrix. The application of MVA results in the discrimination of three different types of Spanish olive oils by their origin. Other approaches examined specific processing steps but not into the whole process and therefore do not analyze or compare the end-product from different manufacturing lines. Kumar et al. [28] reviewed the change from batch to continuous for wet granulation processes using a model-based analysis. The review showed that the shift to continuous production is challenging but has the advantage to use online measurement tools instead of off-line analyses that make real-time-release feasible. But this is only one step of a complex manufacturing process. Our study focuses on the characterization of sodium carbonate by different analytical methods followed by multivariate data analysis, then relating those findings with each of the process versions used. 2. Materials and methods 2.1. Samples The pharmaceutical excipient is the inorganic salt sodium carbonate. It was produced on three different production lines during this study, which will be referred as Processes A, B, and C. Process A is overall a batch process, and Processes B and C are continuous production lines. The dryer and crystallizer were identified as critical process units for all process versions investigated. For Processes A and B the same type of dryer is used. Process C is equipped with a dryer, working under a different drying principle than the one used by A and B. Since Processes B and C are continuous they use a different type of crystallizer from Process A. 2.2. Raman spectroscopy Raman spectra were collected using RamSys FT-Raman instrument by Bruker together with the software OPUS 6.5 (Bruker Optic GmbH, Ettlingen, Germany). The exciting source was a Nd:YAG laser emitting invisible radiation at 1064 nm. The laser light was transmitted by optical fibres to the probe. The laser output at the exit port of the probe was >1000 mW. The spectrometer was equipped with a liquid nitrogen cooled Ge detector (D418-T). The measured spectral range was between 3600 and 10 cm−1 with a 8 cm−1 resolution and averaging 10 scans. For each batch 10 spectra of the sampling point were measured to reduce the variance caused by the analytical method (i.e., different pressure of the probe on the sampled powder). The mean of spectra replicates for each batch was calculated before subsequent multivariate data analysis has been applied. 2.3. Laser diffraction To determine the particle size distribution for all three manufacturing processes the laser diffractometer Mastersizer 2000 v.5.60 with the dry powder dispersion unit Scirocco 2000 (both by Malvern Instruments Ltd., Worcestershire, UK) was used. The light sources are a He–Ne laser (632.8 nm, red, 4 mW) and a LED laser (470 nm, blue, 0.3 mW). The measured particle size range is from 0.02 to 2000 m and is displayed as a volume based distribution. Particle size is calculated using Fraunhofer approximation.
210
A. Hertrampf et al. / Journal of Pharmaceutical and Biomedical Analysis 114 (2015) 208–215
To destroy the agglomerates the dispersion pressure was 3.5 bar. A sample of 2 g was measured within 10 s.
Venetian blind cross-validation was performed with three splits. [6,29,30]
2.4. X-ray powder diffraction
3. Results and discussion
X-ray powder diffraction was performed with STOE StadiP 611 KL powder diffraction system with Mythen-PSD detector (both STOE & Cie GmbH, Darmstadt, Germany). Samples were measured in transmission with Cu-K␣1 radiation (40 kV, 40 mA). The angle range is from 1◦ to 90◦ 2 with a resolution of 0.015◦ 2. The PSD steps were 2.0◦ 2 with a speed of 15 s/step. The measurements were evaluated using the softwares Igor Pro v.6.2.1 (WaveMetrics Inc., Portland, OR, USA) and STOE WinXPow v.2.23.2 (STOE & Cie GmbH, Darmstadt, Germany). To determine the crystallite size, the reflex broadening was calculated by applying a Rietveld refinement using the software General Structure Analysis System v.12.08 (Los Alamos National Laboratory, NM, USA).
PCA was applied Based on Raman spectroscopy, laser diffraction, and X-ray powder diffraction data. As discussed below through PCA clustering and other trends of the three types of samples (Processes A, B, and C) could be identified.
2.5. Scanning electron microscopy SEM images were recorded by using the Zeiss system Supra 35 for Processes A and B and the Leo 1530, which is also by Zeiss (Carl Zeiss Microscopy GmbH, Jena, Germany) for samples from Process C. These images allowed the determination of the surface morphology. Both microscopes used a field emission cathode and an Everhardt–Thornley detector. That detector is secondary electron detector. The scans have been performed with a resolution of 2 nm. The acceleration voltage was set to 5 kV, 8 kV or 10 kV and magnification of 50- and 500-fold was chosen. For preparation the samples were applied on top of aluminium samples holders, which were covered by a double-faced adhesive and conductive foil. Afterwards the samples were coated with platinum. This layer had a thickness of 10 nm or 30 nm. 2.6. Multivariate data analysis The above mentioned analytical techniques data were analyzed by applying multivariate data analysis, using MATLAB® R2010a v.7.10.0 (MathWorks, Inc., Natick, MA, USA) and the PLS Toolbox® v.7.5 (Eigenvector Research Inc., Wenatchee, WA USA). For the Raman spectroscopy with a large wavenumber range it is possible to have bands dominating over others; therefore it can be helpful to choose narrower ranges and/or use some signal processing to improve discrimination. The wavenumber regions that provided the best discrimination were: 1460–1379 cm−1 , 781–619 cm−1 and 441–360 cm−1 . Pre-processing techniques like standard normal variate, first derivatives preceded by Savitzky–Golay smoothing or mean centring, were used. Mean centred spectra gave the best results. Cross-validation was performed using venetian blinds with ten splits. Laser diffraction particle size distributions were pre-processed by mean centring only. The whole range of the particle size distribution was used to build a MVA model. Venetian blinds with four splits were used for cross-validation. In XRPD the layer thickness of the samples can vary due to the preparation. This can affect the intensities in the diffractogram. To compensate this effect area normalization (area = 1) was applied. The peak structure, which includes the information about crystallinity and crystallite size, is not affected by the normalization used. In addition the data was mean centred before further analyses were applied. The range considered was between 22◦ and 90◦ 2. After pre-processing the data, PCA was used to project the data into lower dimensional spaces and in that way help classification from each analytical technique. Therefore PCA was used to support the exploratory and unsupervised analysis of the data (i.e., not imposing a process version or any direct process knowledge).
3.1. Raman spectroscopy The raw spectra of the products from Processes A, B, and C are almost indistinguishable. The main peak at 1078 cm−1 can be assigned to a symmetric 1 (CO) stretching vibration [31]. However a closer look at the spectra revealed differences in the smaller peaks around 1420 cm−1 , 700 cm−1 and 420 cm−1 (cf. Fig. 1). Applying PCA showed a clearer discrimination if only the specific regions around the before mentioned peaks were chosen. The 52–245 cm−1 peak probably contains information about the crystal structure. However this region was excluded. Afterwards no other chemical differences could be detected. The remaining differences within the selected regions seem to be caused by the particle size distribution (PSD) – see below. Fig. 2 displays the score (a) and loading (b) plots of the first two principal components (PCs). All batches can be described by a single PCA model, with all samples inside the 95% confidence interval. The first two principal components explain 94.64% of the samples variance. Additional principal components do not allow any discrimination of differences among process versions. The loading plot shows that all three peaks are described by PC1. The 700 cm−1 peak intensity differences explain most of the variance within PC1, while the peak around 420 cm−1 is very well captured in PC2. The 1420 cm−1 peak has almost no contribution to PC2. The two small “neighbour” peaks of the 700 cm−1 peak, which can hardly be seen on the raw spectra, are well captured and contribute significantly to PC2. The correlation is positive and therefore in the opposite direction as for the 700 cm−1 peak. Nevertheless, the score plot shows that Processes A and B are very similar to each other and that Process C can be discriminated (Processes A and B use the same type of dryer). The influence of process version (batch vs. continuous) and the type of crystallizer does not seem to be large, in terms of Raman spectra. The residuals of each sample (distance to the PCA model) are very low and non-significant. Only two Process A batches show larger residuals, meaning that these batches may be different from the remaining ones and not well described by a same PCA model. 3.2. Laser diffraction The particle size distribution in Fig. 3 shows that Processes A and B have fines in the range between 3.6 m and 11.25 m with an average size around 5.6 m. Two batches of Process B do not show these fines. There is a slight increase of particles in the range between 12 m and 100 m followed by the main particle size between 100 m and 1260 m. Process C does not show fines like the other processes but it shows a small fraction of particles between 28 m and 89 m (max. 56.4 m). But the volume (%) for this size is still lower as for Processes A and B. It should be mentioned that the Frauenhofer approximation is not completely accurate for small particles below ten times of the laser light wavelength (i.e.,
