in-line and real-time monitoring of resonant acoustic

ANALYTICAL SCIENCES JANUARY 2017, VOL. 33 41

Introduction

The application of process analytical technology (PAT) and a quality by design (QbD) approach is particularly preferred by the European Medicines Agency (EMA) and the US Food and Drug Administration (FDA).1,2 This is a technology to ensure the quality of the final product by measuring the critical quality attributes (CQA) of the products in real-time by non-destructive and non-contact methods in the course of the manufacturing process.3–5 Process monitoring by off-line methods requires multiple interruptions in order to perform frequent manual sampling between manufacturing processes, and this increases the cost and the risk of loss of important samples. Therefore, in order to perform real-time process monitoring and prediction, the application of in-line or on-line analysis methods is recommended. In the near future, PAT systems may be introduced into actual manufacturing processes.

Near-infrared (NIR) spectroscopy,6 Raman spectroscopy,7 and non-destructive and non-contact spectroscopic analysis methods are candidate analytical methods, and these are currently

attracting the most attention. In particular, NIR has attracted attention in a very wide range of fields in combination with chemometric technology.8 NIR spectroscopy is used in pharmaceutical analysis to extract information on the chemical and physical factors of ingredient content,9–13 moisture,14–16 crystalline polymorphism,17–19 particle size,15,20,21 and density.21 The energy level of NIR light is between visible light and infrared light (IR). Because many absorption peaks overlap each other in this wavelength region, it is difficult to identify an absorption peak. However, quantitative analysis of absorption peaks in the NIR wavelength range can be performed by use of multivariate analysis.22 NIR spectra have a significantly lower absorption intensity (molar extinction coefficient) than IR spectra. Therefore, sample dilution is not required before measurements. In addition, NIR absorption spectra can simultaneously quantify multiple components by use of Beer–Lambert’s law.

Content uniformity of pharmaceutical products in drug manufacturing is an key factor that must always be guaranteed. There are many reports on drug uniformity of the pharmaceutical products during the ordinary mixing process by high performance liquid chromatography23 and ultraviolet spectroscopy.24 In particular, one of the most accurate and effective methods to characterize drug content uniformity during pharmaceutical manufacturing processes is partial least squares regression

2017 © The Japan Society for Analytical Chemistry

† To whom correspondence should be addressed.E-mail: [email protected]

In-line and Real-time Monitoring of Resonant Acoustic Mixing by Near-infrared Spectroscopy Combined with Chemometric Technology for Process Analytical Technology Applications in Pharmaceutical Powder Blending Systems

Ryoma TANAKA,* Naoyuki TAKAHASHI,** Yasuaki NAKAMURA,** Yusuke HATTORI,* Kazuhide ASHIZAWA,*** and Makoto OTSUKA*†

* Research Institute of Pharmaceutical Sciences, Faculty of Pharmacy, Musashino University, 1-1-20 Shinmachi, Nishi-Tokyo, Tokyo 202–8585, Japan

** Life Science Division, Daiwa Can Co., Ltd, 2-7-2 Marunouchi, Chiyoda, Tokyo 100–0005, Japan *** SSCI Laboratory, Musashino University, 1-1-20 Shinmachi, Nishi-Tokyo, Tokyo 202–8585, Japan

Resonant acoustic® mixing (RAM) technology is a system that performs high-speed mixing by vibration through the control of acceleration and frequency. In recent years, real-time process monitoring and prediction has become of increasing interest, and process analytical technology (PAT) systems will be increasingly introduced into actual manufacturing processes. This study examined the application of PAT with the combination of RAM, near-infrared spectroscopy, and chemometric technology as a set of PAT tools for introduction into actual pharmaceutical powder blending processes. Content uniformity was based on a robust partial least squares regression (PLSR) model constructed to manage the RAM configuration parameters and the changing concentration of the components. As a result, real-time monitoring may be possible and could be successfully demonstrated for in-line real-time prediction of active pharmaceutical ingredients and other additives using chemometric technology. This system is expected to be applicable to the RAM method for the risk management of quality.

Keywords Resonant acoustic mixing (RAM), process analytical technology (PAT), powder blending, powder technology, direct compression formulation, near infrared (NIR) spectroscopy, chemometrics, real-time monitoring, partial least squares regression, quality by design

(Received August 19, 2016; Accepted November 9, 2016; Published January 10, 2017)

42 ANALYTICAL SCIENCES JANUARY 2017, VOL. 33

(PLSR) analysis using NIR spectra. However, in order to construct drug production processes with high uniformity, it is required to have an NIR calibration model based on many standard samples and with sufficient mixing time using an ordinary blend mixer.9

Resonant acoustic® mixing (RAM) technology (Resodyn, MT) is a system that performs mixing using vibration by means of controlling the acceleration and frequency according to the equation of simple harmonic motion. The system is explained in more detail in previous reports.25–29 Previously, we reported that RAM is estimated to repeatedly throw the powder upward into the air and perform mixing by utilizing free fall (Fig. 1).30 In brief, RAM can produce an ideal mixing state with low drug contents over a very short period of time in comparison with ordinary methods using a modified V-shape blender. Therefore, RAM has been applied to a broad field of mixing applications.28–30 RAM can be applied to blending processes with low drug content formulations in pharmaceutical manufacturing processes. However, the risk management of quality in actual production has not yet been established for the RAM method.

In this study, we examined the application of PAT for the combination of RAM, NIR, and chemometric technology as tools to support their introduction into manufacturing processes. Content uniformity was based on the PLSR model. The objective included quantitative monitoring, prediction of formulations for uniform blending by RAM for in-line and real-time processes, and risk management of quality with RAM in the establishment of actual production methods.

Experimental

MaterialsAnhydrous lactose (Super Tab® 21AN, DFE Pharma, Goch,

Germany) was used as a diluent, potato starch (Kosakai Pharmaceutical, Tokyo, Japan) was used as a disintegration agent, and anhydride theophylline bulk powder (Shizuoka Caffeine, Shizuoka, Japan) was used as an active pharmaceutical ingredient (API) as an anti-asthma drug. All powder samples were sieved through a mesh screen (850 μm, Tokyo Screen, Tokyo, Japan) after mixing. The formulations for direct compression were used as ideal standard mixed formulation samples. As shown in Table 1, 30 pharmaceutical formulations were weighed and placed in individual 50 mL glass sample containers. These samples were mixed completely by RAM at 100g (1.0g = 9.8 m/s2) for 300 s.25,30 In addition, the samples in Run 1 – 4 for mixing process monitoring were prepared on the basis of the formulations in Table 2, and the mixing processes were monitored using the NIR method.

Resonant acoustic mixing conditionsThe equipment used was a Resonant Acoustic LabRAM

(Resodyn, MT), with the frequency fixed at 60 Hz. Table 2 shows each vibration force. In this case, the frequency was set at approximately 60 Hz in order to transmit the force directly to the mixing plate.28,29

Near-infrared spectroscopy and multiple data analysis conditionNIR spectra were obtained by the diffuse reflection method.

In this method, the diffuse reflection lights are repeatedly used to illuminate the sample surface and reflection then, refraction, transmission, and scattering are measured. A filter spectroscopic system interference filter type MicroNIR 1700 spectrometer (S1-00498, JDSU, Milpitas, CA) was used for NIR monitoring.31 The wavelength range and integration number were 980 – 1600 nm and 50, respectively. The integration time and scan

Fig. 1　Resonant acoustic mixing (RAM) system used in this study.

Table 1　Content of the different mixtures used for calibration

Lactose, wt% Potato starch, wt% Theophylline, wt%

70 25 565 30 575 20 570 20 1065 25 1075 15 1070 15 1565 30 1575 10 1570 10 2065 15 2075 5 2070 5 2565 10 2560 15 2560 10 3065 5 3055 15 3080 15 560 35 580 10 1060 30 1080 5 1560 25 1550 30 2060 20 2055 15 2050 25 2560 15 2555 20 25


speed were 10 ms and 1 scan/s, respectively. MicroNIR Pro 2.0.1 (JDSU) software was used.

Figure 2 shows a schematic diagram of the measurement system. The glass vessel (50 mL volume) containing the sample was fixed by clamping to the sample stage coupled to the RAM vibration plate. After clamping, the input controls were set to correct for the weight balance of the whole tube and stabilize the equipment. Spectra were obtained continuously after securing the MicroNIR.

Calibration was modeled on the basis of the NIR spectra of standard mixed samples using PLSR. For calibration model building, samples of the 30 types of calibration model stated previously were examined, and which were assumed to be completely mixed by RAM (vibration force was 100g; blending time 300 s).25,30 A total of 540 NIR spectra were obtained six times during vibration at each vibration force setting (40, 60, or 80g). The software used for spectral analysis was Pirouette 4.5 (Infometrix, Bothell, WA). The obtained NIR spectra were subjected to pretreatment with various functions. The objective variable for each component content (wt%) analyzes multivariate data by PLSR, and from this constructs a calibration model. Treatment was carried out in order to reduce the effect of unnecessary variables and noise.32,33 PLSR was performed using untreated (Non) spectra and treated spectra using normalization (Nor), second derivative, and multiplicative scattering correlation (MSC). In addition, centering was performed before all of the analyses, and this dimension reduced the average of the data to the origin. The calibration models were validated by the one-leave-out-cross-validation method.

Quantitative monitoring was performed using Runs 1 – 4. Measurements were performed three times. The calibration

models were evaluated by comparing chemometric parameters associated with the measured and predicted values, such as the standard error (SE, Eq. (1)), prediction residual error sum of squares (PRESS, Eq. (2)), and linear correlation coefficients (r). The powder sample mixing processes in Runs 1 – 4 samples were evaluated quantitatively using the most accurate calibration model involving MSC.

SE PRESS=n (1)

PRESSi

n

i i= ∑ −=1

2{ ( )}y f x (2)

Results and Discussion

PLSR model overviewFigure 3 shows the raw NIR spectral changes of the

theophylline formulation obtained during the mixing process using the RAM equipment, and their second derivative spectra. In the raw spectra, the broad band peaks at 1450 – 1500 nm were attributable to the O–H groups of the free water in between the particles. The baseline related to light scattering from the powder samples might have fluctuated because of noise and changes in the local density of each powder component due to the influence of vibration from RAM. Second derivative processing was carried out in order to eliminate the baseline shift and separate the overlapping peaks.

From second derivative spectra, the peaks based on intramolecular hydrogen bonds of starch and lactose were observed at 1430 and 1480 nm, respectively, and then the first overtone of the O–H stretching vibration of intermolecular hydrogen bond between lactose and starch was observed at 1580 nm. In addition, a peak based on the second overtone of the C–H stretching vibration in lactose and starch was observed in the vicinity of 1225 nm, and C–H2 produced combination tones of stretching vibration and deformation vibration. Theophylline had weak peaks attributed to the first overtone of N–H bonds at

Table 2　Process parameters and contents of the formulation parameters

Run

Parameter Component

Force/gaLactose Potato starch Theophylline

g wt% g wt% g wt%

1 40 3.78 63 1.62 27 0.6 102 60 3.78 63 1.62 27 0.6 103 80 3.78 63 1.62 27 0.6 104 60 4.08 68 1.62 27 0.3 5

a. 1.0g = 9.8 m/s2.

Fig. 2　Schematic diagram of the near-infrared spectroscopy (NIR) measurement system for the resonant acoustic mixing (RAM) system.

Fig. 3　Near-infrared spectroscopy (NIR) spectra obtained for the preparation of the partial least squares regression (PLSR) calibration models. Light solid lines are raw data; solid lines are the second derivatives (n = 540).


1490 nm, and C–H3 combination tones of stretching vibration and deformation vibration were at 1360 nm, with the second overtone of C–H3 stretching vibration at 1150 nm. This was because of the low content of theophylline. To improve the information on chemical substances in the spectral data set and to construct the PLSR model, various pretreatment methods were combined with NIR spectroscopy. In this study, the spectrum was converted and pretreatment performed as a way to eliminate the physical information of the spectral data set. When optimizing the obtained model, it is generally required to remove outliers and unnecessary wavelength ranges. However, as a result of examining the spectral data set, further means to improve the model quality were not found. Table 3 shows the effect of a variety of pretreatments on the analytical results of the PLSR models.

From this result, the standard error based on the training set (SEC) and standard error based on cross-validation set (SEV) was reduced, and the linear correlation coefficient constants

indicated that a highly linear model has been constructed. The reason for this was because the baseline was corrected by excluding physical factors such as changing density and noise. Thus, the results showed a straight line.

Figure 4 shows the regression plots in each model and the loadings plots for variables (LV) 1 – 3 of each composition. By checking the regression plot and loading plot, the values of the explanatory variables can be used to evaluate the respective importance in the calibration model. Absolute values of the vectors indicated a high contribution of the functional group, indicating that it contained chemical information. A regression plot included the peaks caused by the O–H and C–H groups of lactose and starch. This result indicates that lactose and starch are the major components and caused further fluctuations in the concentrations in the formulation. This was particularly the case for the content of lactose, yet both can be inferred because the variation in the concentration was larger. Looking at the loadings plots, each LV1 showed a high contribution ratio, and

Table 3　Error analysis by the partial least squares regression (PLSR) model based on multiplicative scattering correlation (MSC)-treated spectra

Pretreatment LVa SECb, wt% PRESS Calc, wt%2 r Cald SEVe, wt% PRESS Valf, wt%2 r Valg

Lactose Nonh 2 3.031 4.932 × 103 0.933 3.042 4.998 × 103 0.932Nori 3 2.585 3.581 × 103 0.952 2.606 3.668 × 103 0.9512ndj 3 2.110 2.387 × 103 0.968 2.126 2.440 × 103 0.967MSCk 3 2.078 2.314 × 103 0.969 2.088 2.354 × 103 0.969

Starch Nonh 2 1.902 1.943 × 103 0.974 1.909 1.967 × 103 0.974Nori 2 1.649 1.46 × 103 0.981 1.657 1.482 × 103 0.9812ndj 3 1.555 1.296 × 103 0.983 1.568 1.328 × 103 0.983MSCk 3 1.959 2.056 × 103 0.973 1.970 2.096 × 103 0.972

Theophylline Nonh 2 2.237 2.686 × 103 0.956 2.244 2.719 × 103 0.955Nori 1 2.018 2.190 × 103 0.964 2.022 2.208 × 103 0.9642ndj 2 1.728 1.604 × 103 0.974 1.732 1.621 × 103 0.973MSCk 3 1.363 0.996 × 103 0.984 1.379 1.027 × 103 0.983

a. Latent variables. b. Standard error of calibration. c. Prediction residual error sum of squares of calibration. d. Linear correlation coefficient of calibration. e. Standard error of cross-validation. f. Prediction residual error sum of squares of cross-validation. g. Linear correlation coefficient of cross-validation. h. Untreated spectrum. i. Treated spectrum by normalization. j. Treated spectrum by second derivative. k. Treated spectrum by multiplicative scattering correlation.

Fig. 4　Regression vector and loadings of latent variables (LV) 1 – 3 of each composition based on multiplicative scattering correlation (MSC)-treated spectra. Solid lines are regression vectors, dark solid lines are LV1, light solid lines are LV2, and broken lines are LV3 of (a) lactose, (b) potato starch and (c) theophylline.


also a characteristic peak of each component. In LV2 and LV3, the peaks associated with lactose or starch were confirmed. Therefore, each of the PLSR models was represented by the characteristic peaks of each component, and construction of a reliable model was confirmed. Quantitative monitoring was carried out using Runs 1 – 4 of the sample.

In-line real-time monitoring and blending profilesFigure 5 shows the blending profile of each sample by RAM.

These samples made it possible to predict the satisfactory monitoring of each component. Run 1 had a vibration force of 40g, and each component was stable in the range of 95 – 105% of the theoretical values for about 40 s. Furthermore, each component was confirmed to be stable around the theoretical density value at about 20 and 12 s by further increasing the acceleration (Runs 2 and 3). Although the API was 5% in Run 4 and the low content uniformity mixture was more difficult, each component was confirmed to be stable around the theoretical density value at 30 s when vibration force was 60g.

These results demonstrated that the resulting PLSR calibration model was a robust model capable of measuring the nature of the mixing in various conditions. It was demonstrated that the RAM apparatus could mix various pharmaceutical powders simply, rapidly, precisely and uniformly.

Conclusions

RAM was able to produce uniform mixing of the test powders in just 12 s. In addition, in a short period of time RAM could achieve uniform mixing by increasing the acceleration (vibration force). Furthermore, using NIR spectroscopy was possible to measure mixing in-line and in a non-destructive and non-contact manner. A robust PLSR model with regard to the component concentrations was built, and the model managed the RAM configuration parameters (vibration force) and changing concentrations. These methods allowed real-time monitoring and successfully demonstrated the in-line real-time prediction of API and other additives using chemometric technology. The present study has shown the effectiveness of uniform drug

powder mixing by the application of RAM and the usefulness of PAT application using NIR spectroscopy and chemometrics. These systems are expected to expand the use of the RAM method in the risk management of quality in actual production for PAT applications and QbD approaches.

References

1. Database of ICH Guidelines, Compilation prepared by International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use (ICH), http://www.ich.org/products/guidelines.html.

2. Database of FDA Guidance for Industry, Compilation prepared by Food and Drug Administration, http://www.gmp-compliance.org/guidemgr/files/PAT-FDA-6419FNL.

3. M. J. Maltesen, S, Bjerregaard, L. Hovgaard, S. Havelund, and M. Weert, Eur. J. Pharm. Biopharm., 2008, 70, 828.

4. A. S. Rathore and H. Winkle, Nat. Biotechnol., 2009, 27, 26.

5. A. S. Rathore, Trends Biotechnol., 2009, 27, 546. 6. Y. Roggo, P. Chalus, L. Maurer, C. L. Martinez, A. Edmond,

and N. Jent, J. Pharm. Biomed. Anal., 2007, 44, 683. 7. D. S. Hausman, R. T. Cambron, and A. Sakr, Int. J. Pharm.,

2005, 298, 80. 8. D. L. Massart, B. G. Vandeginste, L. M. C. Buydens, P. J.

Lewi, J. Smeyers-Verbeke, and S. De Jong, “Handbook of Chemometrics and Qualimetrics: Part A”, 1997, Elsevier, Amsterdam.

9. Y. Hattori, Y. Tajiri, and M. Otsuka, J. Near Infrared Spectrosc., 2013, 21, 1.

10. W. Plugge and C. Vliest, J. Pharm. Biomed. Anal., 1992, 10, 797.

11. S. S. Sekulic, H. W. Ward, D. R. Brannegan, E, D. Stanley, C. L. Evans, S. T. Sciavolino, P. A. Hailey, and P. K. Aldridge, Anal. Chem., 1996, 68, 509.

12. D. J. Wargo and J. K. Drennen, J. Pharm. Biomed. Anal., 1996, 14, 1415.

13. M. Blanco, R. Bañó, and E. Bertran, Talanta, 2002, 56, 203.

Fig. 5　Quantitative predictions of lactose (◇ diamonds), starch (△ light triangles) and theophylline (● dark circles) as a function of the powder mixing time. Solid lines indicate 95 and 105% of theoretical content (n = 3). (a) Run 1, (b) Run 2, (c) Run 3 and (d) Run 4.


14. M. Otsuka, Y. Kanai, and Y. Hattori, J. Pharm. Sci., 2014, 103, 2924.

15. M. Otsuka, A. Koyama, and Y. Hattori, RSC Adv., 2014, 4, 17461.

16. Y. Hayashi, T. Sato, and M. Otsuka, J. Near Infrared Spectrosc., 2013, 21, 107.

17. G. Buckton, E. Yonemochi, J. Hammond, and A. Moffat, Int. J. Pharm., 1998, 168, 231.

18. K. Kamada, S. Yoshimura, M. Murata, H. Murata, H. Nagai, H. Ushio, and K. Terada, Int. J. Pharm., 2009, 368, 103.

19. K. DeBraekeleer, F. Sánchez, P. A. Hailey, D. C. A. Sharp, A. J. Pettman, and D. L. Massart, J. Pharm. Biomed. Anal., 1998, 17, 141.

20. H. Tanabe, K. Otsuka, and M. Otsuka, Anal. Sci., 2007, 23, 857.

21. M. Otsuka, Powder Technol., 2004, 141, 244. 22. Y. Ozaki, S. Morita, and Y. Du, “Near-Infrared Spectroscopy

in Food Science and Technology, Spectral Analysis”, ed. Y. Ozaki, W. F. McClure, and A. A. Christy, 2007, John Wiley & Sons, NJ.

23. K. Krummen and R. W. Frei, J. Chromatogr. A, 1977, 132, 429.

24. K Järvinen, W. Hoehe, M. Järvinen, S. Poutiainen, M. Juuti,

and S. Borchert, Eur. J. Pharm. Sci., 2013, 48, 680. 25. J. G. Osorio and F. J. Muzzio, Powder Technol., 2015, 278,

46. 26. T. L. Pourpoint, T. D. Wood, M. A. Pfeil, J. Tsohas, and S.

F. Son, Int. J. Aerosp. Eng., 2012, 2012, 1. 27. R. T. Mclaughlin, A. P. Parker, S. A. M. Howes, C. Smith,

and T. Brown, Poster: Taste-mask an API, “Taste Masking Active Pharmaceutical Ingredients Using a Novel Acoustic Mixing Process”, 2015, Catalent Pharma Solutions, NJ.

28. Compilation prepared by Resodyn Acoustic Mixers, Inc., http://www.resodynmixers.com/wp-content/uploads/2009/ 05/ram-technical-white-paper1.pdf.

29. W. Lattin, R. Arbon, J. Bickley and P. Lucon, in Proceeding of Waste Management (WM) Conference, 2010, AZ.

30. R. Tanaka, N. Takahashi, Y. Nakamura, Y. Hattori, K. Ashizawa, and M. Otsuka, RSC Adv., 2016, 6, 87049.

31. N. A. O’Brien, C. A. Hulse, D. M. Friedrich, F. J. Van Milligen, M. K. von Gunten, F. Pfeifer, and H. W. Siesler, in Proceeding of SPIE 8474, ed. M. A. Druy and R. A. Crocombe, 2012, 1 – 8.

32. K. H. Norris and P. C. Williams, Cereal Chem., 1983, 61, 158.

33. Å. Rinnan, F. Berg, and S. B. Engelsen, TrAC, Trends Anal. Chem., 2009, 28, 1201.

in-line and real-time monitoring of resonant acoustic

Documents