An Integrated Process Analytical Technology (PAT) Approach for

Jul 7, 2014 - Analysis of the moving powder bed dynamics suggested that the .... such as Process Analytical Technology (PAT)(13) and Quality-by-Design...
1 downloads 0 Views 1MB Size
Article pubs.acs.org/OPRD

An Integrated Process Analytical Technology (PAT) Approach for Process Dynamics-Related Measurement Error Evaluation and Process Design Space Development of a Pharmaceutical Powder Blending Bed Huiquan Wu,* Maury White, and Mansoor Khan Division of Product Quality Research (DPQR, HFD-940), Office of Testing and Research (OTR), Office of Pharmaceutical Sciences (OPS), Center for Drug Evaluation and Research (CDER), Food and Drug Administration (FDA), FDA White Oak Campus Building 64, 10903 New Hampshire Avenue, Silver Spring, Maryland 20993, United States ABSTRACT: In this work, a model pharmaceutical powder blending system consisting of ibuprofen (drug), MCC, and lactose anhydrous was monitored in real-time via inline near-infrared (NIR) spectroscopy for dual purposes: (1) to examine the effects of formulation variables (drug and MCC contents) and a process variable (impeller rotation speed) on powder blending process kinetics via a 33 full factorial design, and (2) to examine the measurement errors associated with the real-time NIR monitoring environment. The NIR probe was in direct contact with the powder bed for process monitoring. Selected powder blend samples were collected at certain prespecified time points for UV analysis. Three consecutive spectra were used to calculate the standard deviation (stdev) of NIR absorbance at each wavelength. A moving window average was applied to establish the evolution of stdev over the course of powder blending. Two distinct process segments were found: an initial period during which the stdev rapidly decreases and a following fluctuation period during which both the mean and the stdev vary with the formulation and process parameters. Analysis of the process thermodynamics indicated that the initial period of rapid decrease was due to rapid decrease of the thermodynamic driving force, i.e., the powder component concentration gradient. Analysis of the moving powder bed dynamics suggested that the subsequent period of relative stability punctuated by minor fluctuation corresponds to the powder bed’s microstructure fluctuation, i.e., dynamics in compactness, density, and porosity, due to mechanic rotation of the impeller. The analysis was confirmed by ANOVA results. ANOVA shows that the formulation compositions are primary factors dictating how fast the powder system could achieve the macro-homogeneity (often within 1−2 min); both the impeller rotation speed and the formulation composition are the primary factors dictating both the powder blending homogeneity at microlevel and the measurement error associated with real-time dynamic PAT monitoring environment. General Linear Models (GLM) were used to link the critical formulation and process parameters (CPPs) with the derived response variables and to construct a powder blending process design space. For the model powder blending system, it was shown that selection of appropriate impeller rotation speed range is critical to ensure optimal powder blending performance with practically acceptable dynamic noise. Therefore, this work provided an integrated PAT approach and methodology to address practical powder blending challenges from both process engineering and regulatory science perspectives.



INTRODUCTION Blend uniformity has significant regulatory science implications due to a number of reasons: (1) powder blending is one of the most critical unit operations for the two major manufacturing routes of tablets (e.g., controlled release tablet manufacturing route and immediate release tablet manufacturing route); and (2) blend uniformity issues could be translated to content uniformity issues and product quality issues in final tablets (e.g., subpotency or overpotency). This is especially important for high-risk category products such as highly potent drugs, drugs with narrow therapeutic indices (NTI), or drugs that have strong segregation or cohesion tendencies. To address product quality compliance issues, the FDA published its Guidance for Industry “Investigating Out-of-Specification (OOS) Test Results for Pharmaceutical Production”.1 Current Good Manufacturing Practice (CGMP) also requires that content uniformity and standard deviation are reported. In the past decade, many warning letters issued by the FDA were due to GMP testing failures on uniformity testing and batch failures. © 2014 American Chemical Society

Blending operations involve competing processes that lead to mixing and segregation. Mixing in powders generally results from relative motions of groups of particles. Segregation or demixing occurs because the motion of individual particles is a result of their particular characteristics: size, shape, composition, etc. In theory, powder mixing can take place via three mechanisms:2 (i) diffusive mixing, for which particles are distributed over a freshly developed surface; (ii) convective mixing, for which larger particle groups are transferred from one location to another; and (iii) shear mixing, for which slipping planes are set up within the mixture. However, for most manufacturing settings, powder mixing is complex since several mechanisms are invoked simultaneously. As such, fundamental powder mixing mechanisms can be categorized on a more sophisticated basis, for example, a distinction between macromixing and micromixing,3 Special Issue: Process Analytical Technologies (PAT) 14 Received: March 13, 2014 Published: July 7, 2014 215

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

delayed due to off-line testing of in-process samples; and (3) wet chemistry assay is laborious and time-consuming. Innovative pharmaceutical approaches13−15 such as Process Analytical Technology (PAT)13 and Quality-by-Design (QbD)14 have stimulated broad interest for pharmaceutical innovation, promoted science-based regulatory policy development, and helped to ensure or improve pharmaceutical product quality. Regulatory science discussions on topics such as pharmaceutical product quality,16 process understanding,17 and process control strategies18 are available. As a fast and nondestructive technology, near-infrared (NIR) spectroscopy has been applied for pharmaceutical powder blending process monitoring.19−25 Depending on the data analysis and modeling strategies, the use of NIR for pharmaceutical powder blending study can be classified into two categories: (1) a quantitative approach, which typically involves establishing a calibration model (either univariate or multivariate); and (2) a qualitative approach, for which there is no need for establishing a calibration model but typically requires establishing process control limits. The qualitative approach has been commonly used for statistical process control (SPC) and statistical quality control (SQC)26 purposes. For real-world application of process monitoring and control, one practical challenge is how to extract meaningful information from the time-series spectral data, which often carries noise from process system perturbation and error from real-time measurement system simultaneously. In this regard, methodology to estimate error in the online measurement of blend uniformity27 is important. Successful differentiation of the contributions from true process variability, process dynamics, and measurement system error is a key to building a robust SPC model for process monitoring and control. To the best of our knowledge, few reports are available for addressing the practical challenge surrounding the qualitative approach for NIR pharmaceutical PAT applications. For example, how does the dynamic measurement environment (such as real time PAT monitoring a powder bed in motion) impact the measurement result? How does one develop criteria to establish an SPC limit for routine real time monitoring? In this work, the process dynamics and measurement system error of a model dynamic powder blending bed were evaluated systematically using NIR real time process monitoring and a Design of Experiments (DOE) approach. The real-time process data were analyzed in an integrated manner using process engineering principles, analysis of variance (ANOVA), general linear modeling (GLM), and partial leastsquares (PLS) algorithm, etc. This integrated approach allows us to identify critical formulation variables and critical process parameters (CPPs), to map critical formulation variables and CPPs to process design space, and to link them to significant process patterns for process control.

or a distinction according to their systematic or stochastic feature,4 etc. Furthermore, particulate solids are highly intricate systems with very complex dynamic flow and mixing behaviors during the interactions between their constituents and the various parts of the mixture. Therefore, while fundamental mixing mechanisms have been studied energetically by investigators for many years, clearing up these complex processes remains a great challenge.5 A recent review6 discussed the evaluation, mechanisms, and processes of mixing and segregation in powders. The development of new tools for identification of mixing mechanisms is helpful as it opens new ways for process optimization and process improvement. The powder blending process mechanism can be qualitative in nature, which characterizes the intermingling of components. Visual observation, such as using trace particles in conjunction with a video camera, is one of the simplest experimental tools for attaining information on the mechanism involved. During the past two decades, positron emission particle tracking (PEPT) has been used for a number of powder blending unit operation studies, such as examining the influence of blade speed on the performance of a powder mixer,7 characterizing powder motion in a number of types of mixer,8 predicting impeller torque in a high-shear mixer,9 and investigating the effect of impeller rotation speed on powder flow behavior in a continuous blender.10 Undoubtedly, PEPT and particle tracing are valuable tools for research and development. However, they may not be applicable in a pharmaceutical manufacturing setting due to CGMP compliance issues. On the other hand, in the case of different mixing mechanisms acting simultaneously, it is important to assess their relative contribution to the global result of operation, or to determine the dominant ones. To achieve such quantitative characteristics, detailed analysis of the data acquired either by experimental measurements or simulation should be performed. This quantitative analysis also helps to resolve the kinetics of the blending process, i.e., to determine the rate of blending and equilibrium homogeneity. From a process engineering perspective, the main questions for powder blending are how, at what rate, and to what extent is the mixing going on such that an informed process decision regarding when the blending homogeneity criteria are reached can be made appropriately. Obviously powder components’ properties, blending process parameters, blending vessel’s geometry and interior design features, are all important aspects and can interplay together to impact the blending progress, the formation of blending patterns, and the evolution of blend homogeneity. For a specific pharmaceutical powder formulation and blending vessel, understanding the mechanisms and kinetics of the blending process are critical for two aspects: (1) optimizing and controlling the powder blending process; and (2) ensuring final blend quality, especially for challenging formulations where there is either very cohesive components or significant particle size differences among the formulation components involved. The first scenario of very cohesive components may cause great difficulty to achieve homogeneity at the microlevel. The second scenario of significant particle size difference may have inherent segregation tendency.11 Both are challenging cases in terms of obtaining a final blend with excellent physical stability. The conventional method to assess pharmaceutical powder blending uniformity was to conduct wet chemistry assay of grabbed samples. While it is a standard practice, the method has been suffering from several limitations. For example, (1) thief sampling is prone to sampling error;12 (2) process decisions are



EXPERIMENTAL SECTION Materials and Methods. The following pharmaceutical materials were used as received for this study without further processing or purification prior to the powder mixing: USP/BP 70 grade ibuprofen as the active pharmaceutical ingredient (API) (Albemarle Corp., LA. Lot no. 7050-1115) (a BCS Class II drug); USP/NF microcrystalline cellulose (MCC) (JRS Pharma LP, Cedar Rapids, Iowa. Lot no. E5D7C30); lactose anhydrous (Kerry Bioscience, Norwich, NY. Lot no. 1320000007). The particle size distributions of these three powder components were characterized by a Sympatec QICPIC system (Sympatec GmbH, Clausthal-Zellerfeld, Germany). Bell-shaped normal 216

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Figure 1. Fishbone diagram for a powder blending process.

Table 1. 33 full factorial design for the powder blending study and derived response variables DOE pattern

X1 (API, grams)

X2 (MCC, grams)

X3 (impeller rotation speed, rpm)

blending batch code

Y1 (s)

Y2,dynamic (mean of averaged stdev)

Y2,static (mean of averaged stdev)

Y2,normalized (mean of averaged stdev)

Y3 (stdev of averaged stdev)

231 311 111 113 112 132 221 331 211 332 321 213 333 121 222 212 313 233 133 232 322 223 122 123 312 323 131

200 300 100 100 100 100 200 300 200 300 300 200 300 100 200 200 300 200 100 200 300 200 100 100 300 300 100

250 50 50 50 50 250 150 250 50 250 150 50 250 150 150 50 50 250 250 250 150 150 150 150 50 150 250

75 75 75 325 200 200 75 75 75 200 75 325 325 75 200 200 325 325 325 200 200 325 200 325 200 325 75

F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27

48 36 30 42 36 6 30 42 36 48 42 48 48 60 36 48 12 60 12 36 42 42 12 30 36 42 18

0.002389 0.004386 0.003174 0.003648 0.002139 0.001353 0.003003 0.003623 0.004577 0.001415 0.003617 0.006367 0.005355 0.003606 0.001894 0.002288 0.00813 0.007197 0.005555 0.002259 0.002578 0.006735 0.002069 0.006145 0.002161 0.00593 0.001843

0.000796 0.001313 0.004348 0.002643 0.001648 0.002568 0.001065 0.001239 0.00478 0.001318 0.00157 0.00114 0.001187 0.001901 0.001592 0.002205 0.001061 0.001338 0.001671 0.001742 0.001872 0.000918 0.000928 0.004253 0.000775 0.001644 0.001221

0.001593 0.003073 −0.00117 0.001005 0.0004906 −0.00121 0.001938 0.002384 −0.0002 9.74068E-5 0.002047 0.005017 0.004168 0.001705 0.0003017 8.2396E-5 0.007069 0.005859 0.003884 0.000517 0.000706 0.005817 0.001141 0.001901 0.001386 0.004286 0.000622

0.001763 0.002522 0.001576 0.000982 0.000344 0.000269 0.000666 0.000695 0.000857 0.00028 0.000763 0.001087 0.002219 0.000696 0.00053 0.000482 0.001871 0.001028 0.000831 0.000589 0.000742 0.001171 0.000312 0.001162 0.000368 0.001111 0.000405

distributions were obtained. The modes (the particle size corresponding to the peak in the normal distribution) are 67 μm, 85 μm, and 225 μm, respectively. Class 1B HPLC grade methanol (Fisher Scientific, USA. Lot no. 051796) was used for dissolving the powder samples prior to the UV analysis. Experimental Design. In this work, a risk-based approach was used to select variables for the DOE study. As shown in

Figure 1, a Fishbone diagram was created to structure the process of identifying possible causes that could have impacts on the two critical measures of powder blending process, i.e., how fast and what level will be achieved for the powder blending homogeneity? Even though at a first glance six categories of factors (materials, process, measurement, equipment, people, and environment) could have impact on the results of powder blending kinetics and blending process outcomes, in the research 217

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Table 2. Powder formulation density measurement results and estimated effective powder mass interrogated by NIR probe formulation compositions API (g)

MCC (g)

lactose anhydrous (g)

averaged bulk density, Dbulk (g/ mL)

averaged tap density, Dtap (g/ mL)

estimated minimum effective powder mass interrogated by NIR probe, Mmin (mg)

estimated maximum effective powder mass interrogated by NIR probe, Mmax (mg)

2.5 2.5 2.5 5 5 5 7.5 7.5 7.5

1.25 3.75 6.25 1.25 3.75 6.25 1.25 3.75 6.25

11.25 8.75 6.25 8.75 6.25 3.75 6.25 3.75 1.25

0.55 0.49 0.44 0.53 0.47 0.42 0.51 0.45 0.41

0.72 0.66 0.60 0.70 0.63 0.58 0.66 0.61 0.55

3.47 3.05 2.76 3.30 2.95 2.66 3.20 2.85 2.60

18.19 16.59 15.16 17.51 15.91 14.45 16.66 15.29 13.93

Figure 2. NIR spectra of the three pure components at static state.

laboratory setting, a large number of factors such as equipment, people, environment, and measurement systems and technologies can be fixed or controlled, thus being eliminated for further considerations. Therefore, we could focus our attention to the most important material properties and process variables. It has been well recognized that physical properties of the formulation components such as particle size difference and morphology can impact the powder blending process largely. However, particle size difference among the formulation components was not included in this study on the basis of the particle size characterization results. In addition, since we wanted to test the as-received raw materials without any further processing prior to use, particle morphology was not included as an independent variable in this study. Therefore, based on risk assessment and particle size characterization for a three-component powder blending system in a KG-5 blender, two main formulation variables (API weight (g) and MCC weight (g)) and one main process variable (impeller rotation speed (RPM)) were selected as independent

variables in this study. Each factor was tested with three levels. A 33 full factorial design was created using JMP9.0 (SAS Institute, Inc., Cary, NC) for experiments as shown in Table 1. Powder Blending Experiments. After weighing the formulation components using an analytical balance (Mettler Instrument Corp, Highstown, NJ), the components of each formulation (a total of 600 g) were added to the bowl (5 L) of a KG-5 blender (Key International, NJ) in the following sequence: MCC, ibuprofen, and lactose anhydrous. The KG-5 blender was operated at predefined impeller rotation speed. For blending batches of F2, F3, F4, F13, and F27, mixing was stopped at both 30 and 60 min. Those selected batches covered the scenarios of highest/lowest API concentrations and highest/lowest impeller speeds. Weigh boats were used to extract samples from the three midpoints between the three impeller blades of the KG-5 blender. These three points are at opposing sides of the bowl and make a triangle formation with each other. The samples were then taken to a balance and 2.4 g was taken from each and placed 218

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

ensure that the NIR spectroscopy and UV spectroscopy have similar scale of scrutiny22 for assessing powder homogeneity, an equivalent powder subsample mass of 16 mg powder blend was taken from the scintillation vial (described in Powder Blending Experiments subsection previously) for UV analysis to determine the API concentration in the powder blend. Each 16 mg subsample at each blending time point was transferred to a 25 mL volumetric flask, and was then dissolved in 25 mL of methanol. The dissolving process was facilitated by manual shaking for 10 s, followed by vortex with Scientific Industries’ Vortex-Genie 2 (Bohemia, New York) for 30 s, and again manually for 10 s. The solution was then filtered through a 0.45 μm PTFE filter from Millipore Corporation (Billerica, MA) and divided into three Fisherbrand disposable culture tubes (borosilicate glass 16 × 100 mm). Since both MCC and lactose anhydrous are practically insoluble in methanol, presumably their presence in the filtrate is minimal. Three consecutive UV measurements were made for filtrate in each testing tube. The UV absorbance spectra were recorded over the wavelength range from 200 to 300 nm, where all major and minor absorbance peaks associated with the components of interest are covered. Nine UV spectra were acquired for each powder sample and then averaged. The averaged UV spectrum was used for data analysis and modeling.

into a scintillation vial. Three subsamples from the scintillation vial were used for UV analysis. Powder Blend Density Measurement: Bulk Density and Tap Density. As a representative measurement of the bulk density (Dbulk) of the layered components in the KG-5 blender prior to mixing, the individual components were weighed to obtain a total of 15 g of each formulation. Then, the components were added to the graduated cylinder one at a time, in the same order as they were for the mixing experiments, and the volume was recorded. Tap density (Dtap) of the powder formulation was measured by Electrolab Tap Density Tester (USP) model ETD-1020 (GlobePharma, New Brunswick, NJ). Sixty g of each formulation was premixed by shaking in a 400 mL beaker for 2 min. A sample was taken from each formulation as well as the pure components and weighed in a 50 mL graduated cylinder. The graduated cylinder was filled until the volume reached somewhere between 30 and 50 mL. The mass and volume were then recorded. The cylinder was transferred to the tap density instrument and tapped 2000 times at about 250 taps per min. Once tapping was complete, the final volume was recorded. This procedure was repeated three times for each formulation and pure component. The measurement results for Dbulk and Dtap are listed in Table 2. NIR Spectroscopy. Near infrared (NIR) spectra of blending powders were acquired in real time continuously with a Luminar acoustic-optic tunable-filter (AOTF)-based NIR spectrometer (Brimrose Corporation of America, Baltimore, MD), equipped with a transflectance probe. To ensure representative and consistent sampling from the probe measurement perspective, the optic probe was inserted into the powder bed at a fixed position vertically (at about the middle portion of the powder bed height). The acquisition parameters for the NIR spectrometer are as follows: the number of spectra average was 50; no background correction; normal scan type; the gain was 4×; the NIR spectra was recorded every 6 s. The illuminated spot size for the NIR spectrometer was approximately 4 mm. The corresponding sample sizes being interrogated by NIR,22 Mmin and Mmax, were estimated on the basis of the data of the NIR radiation penetration depth (provided by the instrument supplier) and the powder blend density data (Dbulk and Dtap). As shown in Table 2, Mmin is in the range of 2.59−3.47 mg, and Mmax is in the range of 13.93−18.19 mg. The NIR spectra of pure formulation components at static state are shown in Figure 2. The measurement precision of the NIR spectrometer was tested for each formulation batch by the following procedure: (1) adding the three formulation components to the bowl of the KG5 blender; (2) inserting the NIR probe to the powder bed at a fixed position vertically; (3) acquiring the NIR spectra of the static powder bed for 10 times consecutively; and (4) calculating the standard deviation (stdev) of the 10 NIR absorbance values at each wavelength within the wavelength range of 1100− 2300nm; (5) calculating the average of the stdev of the static NIR spectra over the entire wavelength range of 1100−2300nm. This average was termed as a measurement precision of the NIR spectrometer in this study and was reported as Y2,static in Table 1. It can be seen from Table 1, that the range of Y2,static is of 7.96 × 10−4−9.28 × 10−3. UV Spectrometer. The UV spectra were collected using the Agilent UV−vis 8453 spectrophotometer (Santa Clara, CA) attached to a sipper system connecting the peristaltic pump and tubing. Given that the estimated maximum sample size interrogated by NIR is in the range of 13.93−18.19 mg, to



RESULTS AND DISCUSSION NIR Real-time Monitoring of Powder Blending Process: Measurement Science and Experimental Error Analysis. For contact mode NIR real-time monitoring of the powder blending process, the NIR probe can be treated as a sampling scheme. Only a very small portion of the powder bed that was exposed to NIR radiation travelling out of the NIR probe was sampled and measured. Therefore, technically, the NIR spectra acquired are only representative of the portion of powder bed being scrutinized at the particular moment when the spectra are taken, unless the entire powder bed is fully mixed well and homogenized. At the macro-level, a relatively large compositional inhomogeneity is expected initially once the blending process gets started. However, this compositional inhomogeneity is expected to decay quickly as blending progresses if there is no back-mixing. At the microlevel, the inhomogeneity (either compositional or structural) can be detected when the scale of scrutiny is decreased.28 Therefore, when the scale of scrutiny is decreased to microlevel, it is expected that the materials at different locations within the powder bed may have different compositions prior to the blending end-point, should the analytical technique or material characterization tool be sensitive enough. Under such circumstance, should multiple NIR probes be placed at different locations within the powder bed, a difference between the NIR spectral data acquired at the same time point is expected prior to the true blending end-point. Alternately, direct sampling at various locations within the powder bed, if feasible, would give different wet assay results. At one hand, this provides one advantage of evaluating the locationto-location variability within the powder bed. On the other hand, probe positioning difference will likely contribute to the overall experimental error which we might need to minimize from a measurement perspective, as discussed below. For the experimental setup in this work, the experimental measurement variance, σ2e , is a combination of the true variance resulting form the mixing process, σ2m, the variance introduced by the sampling error, σ2s , and the variance resulting from chemical analysis, σ2a . Mathematically,29 219

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Figure 3. Evolution of stdev profile of the moving window NIR absorbance of three consecutive time points (6 s apart) vs wavelength during the course of powder blending for batch F-20.

σe2 = σm2 + σs2 + σa2

instead both the formulation composition and impeller rotation speed collectively play their roles. Therefore, it should be treated with caution to avoid misinterpretation. Given the dynamic noise and fluctuation existing in the real-time NIR monitoring environment during the powder blending process, a rational approach to analyze the spectral data is critical. Doing so can help: (i) to gain insight about the process; (ii) to provide a scientifically sound strategy for data analysis and modeling during the implementation of pharmaceutical PAT; and (iii) to establish a good in-process control strategy for diagnosing false abnormal situations. As discussed in the data analysis section below, this kind of dynamic noise can be estimated systematically via a DOE approach. Assessment of Powder Blending Process Kinetics and Dynamics Based on Time Series of NIR spectra. A. Chemometric Visualization of Blending Profile Evolution. A moving block standard deviation method8 was employed to calculate the standard deviation of NIR absorbance values at each wavelength over the entire wavelength region (1100−2300 nm), for the averaged NIR spectra acquired at three consecutive time points. First, for each blending batch, a plot of the moving window standard deviation (MWSD) vs wavelength at various blending time windows was made to measure how fast the powder blending process progressed at each wavelength (local level). One representative example of such local progression plot is shown in Figure 3. If a fingerprint or characteristic peak of certain component (API would be ideal) is distinguishable on the MWSD plot, the MWSD plot may help to determine whether a blending end-point is achieved locally or globally. Second, for each blending batch, at each and every time point during the powder blending process, the averaged value of the standard deviations of the NIR absorbance values over the entire wavelength region was calculated. This averaged standard deviation (Stdev) was then plotted against the powder blending time (or NIR reading) as another measure of how fast the

(1)

The variance introduced by the sampling error of the NIR probe, σ2s , can be attributed to two factors: (1) the dynamic error due to dynamic powder bed movement during the blending process, σ2s,dynamic; and (2) the static sampling error due to inherent electronic noise of the NIR spectroscopy, σ2s,static. In this work, although advanced design was implemented and a large number of scans were averaged as well, practically σ2s,static is not zero and can be linked to measurement precision of the NIR spectrometer, Y2,static. The contact mode NIR probe was inserted into the powder bed via a hole engineered in the KG-5 blender cover. The vertical position of NIR probe was fixed by a screw knob over the KG-5 blender cover. With this experimental setup, the dynamic noise associated with the rotating powder bed may be attributed to the powder bed microstructure fluctuations such as: (1) possible voids or cavities existing in the moving powder bed due to powder bed expansion caused by mechanic rotation of the impeller. When the voided powder flows in front of the NIR probe tip, it will cause a variable optical length for the NIR signal. (2) The possibility of no powder present in the optical path within the effective radiation penetration depth when a NIR spectrum is acquired at certain time interval. For blade impellers at high powder Froude numbers, previous powder mechanics analysis based on PEPT results showed that9 because of wall friction, the circumferential velocity can be expected to decrease with height above the impeller, i.e., the assumption of ‘rigid’ body rotation is, at best, a first-order approximation. For our case, the KG-5 blender fitted with impeller blades having large bevel angles produces large powder Froude numbers, and deviation from the ideal ‘rigid’ body rotation is expected. This situation will naturally lead to some fluctuations or noise in the NIR spectra acquired. On the basis of ANOVA and GLM results presented in later subsections of this work, it is evident that the fluctuations are not solely due to the real inhomogeneity of powder material; 220

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Figure 4. Evolution of the averaged stdev profile of the moving window NIR absorbance of three consecutive time points over the entire wavelength range of 1100−2300 nm during the course of blending for batch F20.

depends on the sensitivities of the NIR spectrometer and the Stdev method. The blending process dynamics can be further visualized and characterized in Figure 4. It was found that for the uninterrupted continuous blending of the three-component powder bed in KG5 blender, all of the 27 DOE batches exemplified a similar profile of averaged standard deviation vs blending time. The profile consists of two distinguished stages: first an initial rapidly declining stage where the thermodynamic driving force is large, followed by a relative stable yet fluctuated stage where the process thermodynamic driving force is small. Several derived parameters, including the time elapsed for the first stage, the mean of the averaged Stdev of the second stage (Y2, normalized and Y2, dynamic), and the Stdev of the averaged Stdev of the second stage, were used to characterize the powder blending process kinetics and dynamics in this work, as discussed in the ANOVA subsection.

powder blending process progressed globally. Figure 4 is one representative example of such global progression plot. Figure 3 shows that within the first several minutes of blending, the Stdev profile changes rapidly from the initial highly fluctuated shape to a much less fluctuated shape. This not only indicates a very high mixing efficiency for the initial mixing stage, but also characterizes the initial powder blending process dynamics well. Once the blending is initiated, conceptually the thermodynamic driving force or concentration gradient at time t and specific location, ((ΔCi)/(Δh))t=t,local, can be expressed as: (Ci)t = t ,local − (Ci)t =∞ ,global ⎛ ΔCi ⎞ ⎜ ⎟ = ⎝ Δh ⎠t = t ,local h

(2)

where Ci is the concentration of powder component i, t = ∞ stands for time point when all powder components are fully mixed to form a homogeneous blend, h is the characteristic length that powder molecule should travel within the powder bed. Obviously, (Ci)t=∞, global is the targeted concentration which is fixed for a given powder blending experiment. In addition, as a first-order approximation, the characteristic length of the powder molecule can be considered as a constant for a given rotation speed and a given powder blending experiment. Therefore, the thermodynamic driving force at a specific time point is proportional to the concentration difference between the local value and the targeted value. For batch powder blending process, it is expected such thermodynamic driving force due to concentration difference will be decreased quickly, thus leading to reduced blending kinetics. With more mechanical energy being dispersed into the powder bed via the impeller, the blending process progresses to a state where the macro-level blending homogeneity is apparent, as evidenced by the gradually flattened Stdev profile where the thermodynamic driving force is approaching zero. This graphic determination of blending endpoint (in Figure 3) is qualitative. The accuracy of this method

Y2,normalized = Y2,dynamic − Y2,static

(3)

where Y2, static stands for Y2 derived from the NIR spectra of the initial powder bed acquired at static mode (prior to the rotation of the impeller); Y2, dynamic stands for Y2 derived from the NIR spectra of the powder bed acquired at dynamic mode (during the powder blending process while the impeller is rotating); Y2, normalized stands for Y2 due to measurement error other than NIR electronic noise. B. Powder Blending Process Thermodynamics and Dynamics Analysis: Assessment from a Transport Process Engineering Perspective. There are certain characteristics related to each formulation, which apparently lead to specific mixing profiles of each formulation. For example, formulation composition, powder flow properties, and impeller rotation speed, etc., could be attributed to the differences observed. From a transport process engineering standpoint, the thermodynamic driving force at a specific time point is proportional to the 221

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Table 3. ANOVA results dependent

hypothesis type

source

DF

SS

MS

F value

ProbF

Y1 Y1 Y1 Y1 Y1 Y1 Y2 Y2 Y2 Y2 Y2 Y2 Y3 Y3 Y3 Y3 Y3 Y3

3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3

X1 X2 X1*X2 X3 X1*X3 X2*X3 X1 X2 X1*X2 X3 X1*X3 X2*X3 X1 X2 X1*X2 X3 X1*X3 X2*X3

2 2 4 2 4 4 2 2 4 2 4 4 2 2 4 2 4 4

1138.667 18.66667 1805.333 114.6667 725.3333 405.3333 1.70 × 10−5 5.41 × 10−7 9.90 × 10−6 7.64 × 10−5 9.98 × 10−6 4.46 × 10−6 8.98 × 10−7 5.01× 10−7 8.62 × 10−7 3.54 × 10−6 4.34 × 10−7 1.05 × 10−6

569.3333 9.333333 451.3333 57.33333 181.3333 101.3333 8.52 × 10−6 2.71 × 10−7 2.48 × 10−6 3.82 × 10−5 2.50 × 10−6 1.12 × 10−6 4.49 × 10−7 2.5 × 10−7 2.15 × 10−7 1.77 × 10−6 1.08 × 10−7 2.62 × 10−7

4.8523 0.0795 3.8466 0.4886 1.5455 0.8636 10.4079 0.3306 3.0245 46.6481 3.0491 1.3639 1.7958 1.0011 0.8614 7.0778 0.4337 1.0492

0.0417 0.9243 0.0497 0.6306 0.2778 0.5250 0.00594 0.72784 0.08555 0.00004 0.08409 0.32759 0.2269 0.4093 0.5261 0.0170 0.7812 0.4398

were API concentration (X1), MCC concentration (X2), and impeller rotation speed (X3). With 27 runs of a 33 full factorial design, the analysis of variance (ANOVA) indicates that all main effects and two-way interactions among these main effects are estimable and the three-way interaction is used to estimate the underlying error. Therefore, the following statistical model was used:

concentration difference between the local value and the targeted value. During the initial blending stage, large inhomogeneity is expected from location to location within the powder bed; thus a larger thermodynamic driving force exists. This in turn will not only lead to a rapidly changing concentration profile within the powder bed as evidenced by the rapid decreasing stage on the Stdev plot but also push the process to reach the macrohomogeneity level in the powder bed quickly. After the initial rapid stage of reaching macrohomogeneity, a dynamic yet relatively stable stage with a much smaller order of fluctuation is expected for the powder bed, as evidenced by the fluctuation stage on the Stdev plot. During this stage, microhomogeneity can be achieved after a certain time of additional mixing. Although not examined specifically in this study, our previous study11 indicated that segregation tendency increased as the API/MCC particle size ratio increased. Therefore, selection of appropriate rotation speed may be warranted to achieve optimal blending outcome. With the aforementioned process engineering analysis, further evidence regarding the predominant factors that govern the primary process phenomena and trends at each process stage can be sought from ANOVA of the DOE data set and wet assay of the blends, as demonstrated in the subsequent discussions. Identification of Critical Formulation/Process Variables That Significantly Impact the Blending Process Kinetics and NIR Real Time Measurement Error. Statistical data analysis and general linear modeling were performed on the DOE data set as summarized in Table 1 via SAS9.1 (SAS Institute, Inc., Cary, NJ) to determine the following: (1) the critical formulation/process variables that have a statistically significant impact on the powder blending process responses; (2) if there are any relationships between the formulation/process variables and the response variables during the first and the second mixing stages, i.e., the initial rapidly declining stage through which the macrohomogeneity is reached quickly, and the subsequent stage of relative stability punctuated by minor fluctuations through which microhomogeneity is reached ANOVA for the DOE Data Set to Identify Critical Formulation/Process Variables and Interactions. Statistical analyses were carried out using linear models. The response variable (Y) was Y1 or Y2 or Y3. The independent fixed factors

yijk = μ + αi + βj + γk + (αβ)ij + (αγ )ik + (βγ )jk + εijk i = 1, 2, 3; j = 1, 2, 3; k = 1, 2, 3

(4)

where yijk is the observation (either time elapsed for the first mixing stage, the normalized mean of the averaged stdev for the second mixing stage, or the stdev of the averaged stdev) in the ith level of variable X1 and the jth level of variable X2 and the kth level of variable X3, μ is the overall mean, αi is the ith X1 effect, βj is the jth X2 effect, γk is the kth X3 effect, (αβ)ij is the two-way interaction between the ith X1 effect and the jth X2 effect, (αγ)ik is the two-way interaction between the ith X1 effect and the kth X3 effect, (βγ)jk is the two-way interaction between the jth X2 effect and the kth X3 effect, εijk, the random error, is normally independently identically distributed with mean 0 and variance σ2, which is written as εijk ≈ iid N(0, σ2). After the data were fitted to the model, the statistically important factors and interactions were identified. The interaction term was removed from the model when it was determined to be insignificant at the 0.05 level. On the basis of this statistical model, an ANOVA was conducted to screen significant factors and interactions for each response variable. The results are listed in Table 3. It was shown that at significance level α = 0.05, X1 and X1*X2 are statistically significant for Y1; X3 is statistically significant for Y2,normalized; X1 is marginally statistical significant for Y2, normalized; X3 is statistically significant for Y3. In other words, formulation composition X1 and interaction term X1*X2 dictate the time elapsed during the first mixing stage; impeller rotation speed dictates both Y2, normalized and Y3 of the second mixing stage. Therefore, within the scope of this study, the formulation compositions are primary factors which determine how fast the powder system can achieve macrohomogeneity, often within 1− 2 min; the impeller rotation speed is the primary factor that 222

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Table 4. GLM parameter estimate results dependent

parameter

estimate

std err

t value

prob t

Y1 Y1 Y1 Y1 Y2 Y2 Y2 Y3 Y3 Y3 Y3

intercept X1 X2 X1*X2 intercept X1 X3 intercept X1 X3 X3*X3

56.8889 −0.1008 −0.2133 0.0011 −0.0023 9.36 × 10−6 1.20 × 10−5 0.001787 2.22 × 10−6 −1.9 × 10−5 4.83 × 10−8

12.6477 0.0585 0.0741 0.0003 0.00107 3.98 × 10−6 3.19 × 10−6 0.0005 1.08 × 10−6 4.88 × 10−6 1.2 × 10−8

4.4980 −1.7223 −2.8807 3.0629 −2.1059 2.3504 3.7687 3.8898 2.0484 −3.8168 4.0190

0.0002 0.0984 0.0084 0.0055 0.045862 0.0273 0.000943 0.0007 0.0521 0.0009 0.0005

design space concept in the laboratory setting, a risk-based approach in conjunction with particle size distribution characterization was used to select the main formulation and process parameters as the independent variables. Final GLM models for Y1, Y2, nomalized, and Y3 were selected as follows:

determines both the powder blending homogeneity at the microlevel and the experimental error associated with real-time NIR monitoring for the dynamic powder blending process. A Practical Approach of Selecting Appropriate Impeller Rotation Speed to Achieve Optimal Blending Performance. Given the dynamic process behavior, it is important to select an appropriate impeller rotation speed range to achieve optimal powder blending performance, i.e., the desired level of microhomogeneity while keeping the experimental error associated with real-time NIR monitoring of a dynamic powder blending bed at an acceptable level. In other words, a balance is needed for the selection of appropriate impeller rotation speed. Instead of conducting another DOE optimization study, which is time-consuming and resource intensive, a practical approach was used to determine an appropriate impeller rotation speed range quickly based on the DOE data set already obtained in this work. Sorting out the existing DOE data set (Table 3) by Y2,dynamic and Y3 shows that if we, for example, set up the acceptable criteria for the powder blending homogeneity at the microlevel as follows:

(Y1)i = β0,1 + β1,1(X1)i + β2,1(X 2)i + β3,1(X1X 2)i + εi i = 1, ...., 27,

(5)

(Y2,normalized)i = β0,2 + β1,2(X1)i + β2,2(X3)i + εi i = 1, ...., 27

(6)

(Y3)i = β0,3 + β1,3(X1)i + β2,3(X3)i + β3,3(X3X3)i + εi i = 1, ...., 27

(7)

where (Y1)i, (Y2, normalized)i, and (Y3)i are the ith observation of Y1, Y2, normalized, and Y3; (X1)i, (X2)i, and (X3)i are the ith values for X1, X2, and X3, respectively; error term εi is normally independently identically distributed with mean 0 and variance σ2, which is written as εi ≈ iid N(0, σ2), i = 1,...,27. The GLM modeling was carried out using SAS9.2 (SAS Institute Inc., Cary, NC). The parameter estimate results are listed in Table 4. The GLM prediction results are summarized in Table 5.

{Y2,dynamic ≤ 0.0026} and {Y3 ≤ 0.00075}

the data set can then be classified accordingly. Our classification results show that: (1) All batches with 200 rpm and two batches with 75 rpm meet the above criteria. (2) None of the batches with 375 rpm meets the criteria. Therefore, apparently, a 200 rpm impeller rotation speed would provide a reasonable tradeoff between the powder blending performance (practically acceptable microlevel homogeneity) and practically acceptable dynamic noise for the realtime NIR monitoring of the dynamic powder blending bed. This is consistent with the previous report30 that intermediate rotation speeds exhibited lowest values of RSD (relative standard deviation) of concentration (note: the RSD index characterizes the uniformity of any mixture). In that report,30 it was argued that at intermediate rotation speeds, the maximum number of blade passes (maximum strain) is exerted on the powder, which leads to better mixing performance. Development of Statistical Predictive Models Using General Linear Modeling. On the basis of the ANOVA results in Table 3, statistical predictive models were developed via general linear modeling (GLM) method to link the critical formulation and process parameters (CPPs) with the derived response variables and to construct a powder blending process design space. As discussed previously, to illustrate the process

Table 5. GLM parameter estimate R2 values dependent

R2

CV

rootMSE

DepMean

Y1 Y2 Y3

0.3698 0.4511 0.4767

32.7853 83.6869 49.0052

11.8756 0.00169 0.00046

36.2222 0.00202 0.0009

Table 4 shows that at the significance level of α = 0.05, parameter estimates for β0,1, β2,1, and β3,1 are significant for Y1; β0,2, β1,2, β2,2 are significant for Y2, normalized; β0,3, β2,3, and β3,3 are significant for Y3; β1,3 is marginally significant for Y3. These results are consistent with process engineering analysis, as discussed below. Table 5 shows that, on the basis of screening results of statistically significant variables, a GLM model was able to describe the variability of response variables to the extent of R values of 0.608−0.690. For comparison, if a GLM model is developed on the basis of all of the independent variables (X1, X2, and X3) and their two-way interactions, it can describe the variability of the response variables to the extent of R values of 0.886−0.973 (data not shown). API Concentration (X1) and MCC Concentration (X2). ANOVA results show that X1 and X1*X2 have significant effect 223

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

(at significance level α = 0.05) on Y1. Furthermore, since parameter estimates for X1 and X2 are negative for Y1, the larger X1 or X2, the smaller Y1. This could be explained from a process engineering standpoint: the larger initial X1 or X2 in the powder bed, the larger the thermodynamic driving force characterized by eq 2, thus, the faster the blending process reaching macrohomogeneity level in the powder bed. In addition, since the parameter estimate shows that X1 is significant for Y2, normalized and X1 is marginally significant for Y3, it can be concluded that blend composition such as API concentration significantly impacts the second mixing stage. Impeller Rotation Speed (X3). The impeller rotation speed has a significant effect (at a significance level of α = 0.05) on both Y2,normalized and Y3. The parameter estimate on X3 is positive for Y2, normalized. This indicates that a higher X3 will result in a larger Y2, normalized. This is understandable from a process engineering standpoint, since a larger X3 will lead to a more dynamic powder bed; thus, a more fluctuated second stage of powder mixing will be captured by NIR real-time monitoring due to the dynamic light path the NIR radiation experiences in the dynamic powder bed. The effect of X3 on Y3 is a bit complicated on the basis of the GLM parameter estimate, since the parameter estimate on X3 is negative; however, the parameter estimate on X3*X3 is positive. Overall, intuitively the larger X3 will lead to a more fluctuated second stage of powder mixing. Confirmation from Wet Chemistry Assay of Blends and Chemometrics Modeling at 30 min and 60 min. A partial least squares (PLS1) calibration model was developed to correlate the UV absorbance of the powder blends with the ibuprofen mole concentration (mol/L), as shown in Figure 5.

According to the Complete-Random-Mixture (CRM) model,22,31 this lowest API concentration scenario presents the largest measurement uncertainty for API concentration. For the blending process of batch F2 (75 rpm), the PLS1 prediction error for 60 min (4.64%) was lower than that for 30 min (6.38%), which suggests that extended blending time may improve the blending homogeneity. However, for the blending process of batch F13 (325 rpm), a higher PLS1 prediction error was obtained for 60 min (−17.00%) than for 30 min (−6.23%), which suggests that back-mixing or particle segregation may take place in this case. The opposite effect of impeller rotation speed on the blending homogeneity for batch F2 and batch F13 was consistent with results obtained from a previous continuous powder blending study.30 Recently, characterization of continuous blending performance based on residence time distribution (RTD) modeling has shown that higher blade speed facilitates transverse mixing but leads to a highly variable RTD profile.32 Their results suggested that moderate RPM may facilitate mixing. In this work, all of the blends have the same residence time, which is equal to the batch blending time. For the batch mode mixing studied in this work, while higher impeller rotation speed facilitates transverse mixing as observed in continuous mixing, the phenomenon that higher blade speed leads to a highly variable RTD profile32 could be reflected as increased inhomogeneity within the dynamic powder bed. This may in part explain why comparing to the results of 30 min of blending time, an increased inhomogeneity was observed for 60 min of blending time. Therefore, this work suggests that both moderate impeller rotation speed and prolonged blending time are beneficial for achieving final blending homogeneity.



CONCLUSIONS In this work, a model pharmaceutical powder blending system consisting of ibuprofen (drug), MCC, and lactose anhydrous was monitored in real-time via inline near-infrared (NIR) spectroscopy. A 33 full factorial design was executed for powder blending experiments with various formulation compositions and impeller rotation speeds of the KG-5 high-shear mixer. A moving window average was applied to establish the evolution of stdev over the course of powder blending. Two distinct process segments were found to exist: an initial period during which the stdev rapidly decreased and a following fluctuation period during which both the mean and the stdev varied with the formulation and process parameters. Critical formulation and process parameters (CPPs) were identified via ANOVA: (1) the formulation compositions are primary factors dictating how fast the powder system can achieve macrohomogeneity (often within 1−2 min); and (2) both the impeller rotation speed and formulation composition are the primary factors dictating both the powder blending homogeneity at the microlevel and the measurement error associated with the real-time dynamic NIR monitoring environment. It was shown that appropriate impeller rotation speed range is critical to ensure optimal powder blending performance with practically acceptable dynamic noise. General Linear Modeling (GLM) was used to link the critical formulation and process parameters (CPPs) with the derived response variables and to construct a powder blending process design space. The results of parameter estimates from GLM were interpreted from a process engineering perspective. UV analysis of blending samples collected at two time points (30 and 60 min) from several representative blending batches suggests that both moderate impeller rotation speed and prolonged blending time

Figure 5. PLS1 calibration model to correlate the UV spectral data of a binary mixture of API and Eudragit L100 with the API concentration.

This calibration model was then applied to the UV absorbance spectral data for predicting the ibuprofen mole concentration of grabbed samples at three different locations within the powder bed at two different times (30 and 60 min). The prediction results are shown in Figure 6 and Figure 7. Figure 6 shows the results of the location-to-location variability within the powder bed at the same time. For 4 out of 5 blending batches (F2, F3, F4, F13), the location-to-location variability at 30 min is larger than that at 60 min, which indicates the location-to-location variability decreases over time. Blending batch F27 exempts an opposite behavior, which is probably due to the fact it represents the worst scenario, i.e., the combination of the lowest API concentration and the lowest impeller speed. 224

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

Figure 6. Evolution of location-to-location API concentration variability vs blending time for representative powder blending process batches.

Figure 7. Predicted API concentration for powder blending batches at blending times of 30 and 60 min.

are beneficial for achieving final blending homogeneity. Therefore, an integrated PAT approach that combines NIR real-time process monitoring, DOE, ANOVA, GLM, and UV analysis was established for evaluating the process dynamics and measurement error of a continuous and dynamic pharmaceutical powder blending bed. It provided a methodology to address practical powder blending challenges from both process engineering and regulatory science perspectives.



Notes

The views and opinions expressed in this work are only of the authors and do not necessarily reflect the policy or statement of the U.S. FDA. The authors declare no competing financial interest.



ACKNOWLEDGMENTS The authors wish to acknowledge Dr. Meiyu Shen, Senior Mathematical Statistician, Division of Biometrics VI, Office of Biostatistics, Office of Translational Science, Center for Drug Evaluation and Research, FDA, for strong statistical support. H. Wu wishes to acknowledge Dr. Robbe Lyon (retired from the FDA in 2009) for his mentorship. This project was partially

AUTHOR INFORMATION

Corresponding Author

*Tel: 301-796-0022. Fax: 301-796-9816. E-mail: Huiquan.wu@ fda.hhs.gov. 225

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226

Organic Process Research & Development

Article

(27) Vanarase, A. U.; Järvinen, M.; Passo, J.; Muzzio, F. J. Powder Technol. 2013, 241, 263−271. (28) Danckwerts, P. V. Appl. Sci. Res., Sect. A 1952, 3, 279−296. (29) Fan, L. T.; Chen, Y.-M.; Lai, F. S. Powder Technol. 1990, 61, 255− 287. (30) Vanarase, A. U.; Muzzio, F. J. Powder Technol. 2011, 208, 26−36. (31) Pan, T.; Barber, D.; Coffin-Beach, D.; Sun, Z.; Sevick-Muraca, E. M. J. Pharm. Sci. 2004, 93, 635−645. (32) Gao, Y.; Vanarase, A.; Mizzio, F. J.; Ierapetritou, M. Chem. Eng. Sci. 2011, 66, 417−425.

supported by FDA CDER Regulatory Science and Review Program (project code: RSR-04-16) for research focused on exploring the regulatory science utility of integrating DOE and PAT for the 21st Century CGMP Initiative. The OTR internship which allowed M. White to carry out some experiments for this work is acknowledged. The constructive suggestions from two unanimous reviewers of this manuscript are greatly appreciated.



REFERENCES

(1) FDA Guidance for Industry. Investigating Out-of-Specification (OOS) Test Results for Pharmaceutical Production, 2006. Available at: http:// www.fda.gov/downloads/Drugs/Guidance/Compliance/ RegulatoryInformation/Guidances/UCM070287.pdf (accessed on 01/ 09/2014) (2) Lacey, P. M. C. J. Appl. Chem. 1954, 4, 257−268. (3) Bridgwater, J. Powder Technol. 1976, 15, 215−236. (4) Williams, J. C. Powder Technol. 1968, 2, 13−20. (5) McCarthy, J. J.; Shinbrot, T.; Metcalfe; Wolf, J. E.; Ottino, J. M. AIChE J. 1996, 42 (12), 3351−3363. (6) Hogg, R. KONA Powder Part. J. 2009, 27, 3−17. (7) Bridgwater, J.; Broadbent, C. J.; Parker, D. J. Trans. IChemE. 1993, 71A, 675−681. (8) Forster, R. N.; Seville, J. P. K.; Parker, D. J.; Ding, Y. KONA 2000, 18, 139−148. (9) Knight, P. C.; Seville, J. P. K.; Wellm, A. B.; Instone, T. Chem. Eng. Sci. 2001, 56, 4457−4471. (10) Portillo, P. M.; Vanarase, A. U.; Ingram, A.; Seville, J. K.; Ierapetritou, M. G.; Muzzio, F. J. Chem. Eng. Sci. 2010, 65, 5658−5668. (11) Xie, L.; Wu, H.; Shen, M.; Augsburger, L.; Lyon, R. C.; Khan, M. A.; Hussain, A. S.; Hoag, S. W. J. Pharm. Sci. 2008, 97 (10), 4485−97. (12) Muzzio, F. J.; Robinson, P; Wightman, C.; Brone, D. Int. J. Pharm. 1997, 155, 153−178. (13) PAT - A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance, Guidance for Industry ; U.S. Department of Health and Human Services, Food and Drug Administration, Center for Drug Evaluation and Research (CDER): Rockville, MD, 2004. Available at http://www.fda.gov/downloads/ Drugs/Guidances/ucm070305.pdf (accessed on 01/09/2014) (14) FDA/ICH Guidance for Industry, Q8(R2) Pharmaceutical Development; 2009. Draft available at: http://www.fda.gov/ downloads/Drugs/Guidance/Compliance/Regulatory/Information/ Guidances/ucm073507.pdf (accessed on 01/09/2014) (15) FDA Advancing Regulatory Science at FDA: A Strategic Plan. August 2011. Available at: http://www.fda.gov/downloads/ ScienceResearch/SpecialTopics/RegulatoryScience/UCM268225.pdf (accessed on 01/09/2014) (16) Woodcock, J. Am. Pharm. Rev. 2004, No. Nov/Dec, 1−3. (17) Yu, L. Pharm. Res. 2008 Apr, 25 (4), 781−91. (18) Wu, H.; Hussain, A.; Khan, M. Chem. Eng. Commun. 2007, 194, 760−779. (19) Sekulic, S. S.; Ward, H. W.; Brannegan, D. R.; Stanley, E. D.; Evans, C. L.; Sciavolino, S. T.; Hailey, P. A.; Aldridge, P. K. Anal. Chem. 1996, 68, 509−513. (20) El-Hagrasy, A. S.; Morris, H. R.; D’Amico, F.; Lodder, R. A.; Drenen, J. K., III. J. Pharm. Sci. 2001, 90 (9), 1298−1307. (21) Li, W.; Johnson, M. C.; Bruce, R.; Rasmussen, H.; Worosila, G. D. J. Pharm. Biomed. Anal. 2007, 43, 711−717. (22) Wu, H.; Tawakkul, M.; White, M.; Khan, M. A. Int. J. Pharm. 2009, 372 (1−2), 39−48. (23) Wu, H.; Khan, M. A. J. Pharm. Sci. 2009, 98 (8), 2784−2798. (24) Sulub, Y.; Wabuyele, B.; Gargiulo, P.; Pazdan, J.; Cheney, J.; Berry, J.; Gupta, A.; Shah, R.; Wu, H.; Khan, M. A. J. Pharm. Biomed. Anal. 2009, 49, 48−54. (25) Scheibelhofer, O.; Balak, N.; Koller, D. M.; Khinast, J. G. Powder Technol. 2013, 243, 161−170. (26) Montgomery, D. C. Introduction to Statistical Quality Control, 4th ed.; John Wiley & Sons, Inc.: Hoboken, NJ, 2001. 226

dx.doi.org/10.1021/op500085m | Org. Process Res. Dev. 2015, 19, 215−226