Characterization of pharmaceutical powder blends by … · Characterization of Pharmaceutical Powder Blends by NIR Chemical Imaging ... Content uniformity of solid oral dosage

Characterization of PharmaceuticalPowder Blends by NIR Chemical Imaging

HUA MA,1 CARL A. ANDERSON1,2

1Graduate School of Pharmaceutical Sciences, Duquesne University, Pittsburgh, Pennsylvania 15282

2Duquesne University Center for Pharmaceutical Technology, Pittsburgh, Pennsylvania 15282

Received 8 July 2007; revised 26 August 2007; accepted 19 September 2007

Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/jps.21230

Corresponde5281; Fax; 216

Journal of Pharm

� 2007 Wiley-Liss

ABSTRACT: This study demonstrates the capabilities of NIR imaging as an effectivetool for characterization of pharmaceutical powder blends. The powder system used inthis small-scale powder blending study consists of acetaminophen (APAP, the modelAPI), microcrystalline cellulose (MCC) and lactose monohydrate. Mixtures of thesecomponents were blended for different times for a total of ten time points (ten blendingtrials). Images collected from multiple locations of the blends were used to generate aqualitative description of the components’ blending dynamics and a quantitative deter-mination of both the blending end point and the distribution variability of the compo-nents. Multivariate analyses, including pure-component PCA and discriminate PLS,were used to treat the imaging data. A good correlation was observed between imagingresults and a UV-Vis monitoring method for determining blend homogeneity. Scoreimages indicated general trends of the distribution of blending constituents for all tenblending trials. The API distribution pattern throughout blending was detected and theAPI domain size for different blending trials was compared. Blending insights obtainedfrom this study may be transferable to large scale powder blending. Blending processunderstanding obtained from this study has the potential to facilitate the optimization ofblending process control in the future. � 2007 Wiley-Liss, Inc. and the American Pharmacists

Association J Pharm Sci 97:3305–3320, 2008

Keywords: powder blending; NIR ch
emical imaging; blending characterization;PAT; pure component principle component analysis; discriminate partial least squareanalysis; multiple location blend image collection; blend compact
INTRODUCTION

Powder blending is a fundamental and importantunit operation in the manufacturing of pharma-ceutical products. The desired outcome of thisoperation is a homogeneous mixture of rawmaterials. Content uniformity of solid oral dosageforms is likely to be compromised without homo-geneity at unit-dose scale at the end of theblending stage.

nce to: Carl A. Anderson (Telephone: 216 844844 3106; E-mail: [email protected])

aceutical Sciences, Vol. 97, 3305–3320 (2008)

, Inc. and the American Pharmacists Association

JOURNAL OF PH

A common blend uniformity practice is toremove samples by thief probe sampling and ana-lyze samples using UV-Visible spectroscopy.1–5 Ablend is considered homogenous when the relativestandard deviation (RSD) values among multiplesamples collected at the same time are withinspecification. This method of monitoring blenduniformity requires invasive sampling proce-dures. Significant sampling bias is often intro-duced with the use of thief probes for removal ofsamples from powder blenders, as addressed in anumber of studies on thief sampling from bulkpowder bed.1–3,5,6 An alternate method for blendmonitoring is to determine the homogeneity ofblend components at a probe location from data

ARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008 3305

3306 MA AND ANDERSON

collected by near-infrared spectroscopy (NIR).NIR monitoring is a noninvasive technique withthe potential to minimize analysis times andsample preparation while eliminating the sam-pling errors associated with the disruption ofpowder beds and removal of non-representativesamples from the blender by a sample thief. Withthe aid of chemometric tools, single point NIRmonitoring methods are faster and more reliablealternatives to the traditional monitoring meth-ods.7,8 Studies using NIR for blending uniformityassessment have demonstrated qualitative andquantitative approaches for on-line and in-linepowder blending process monitoring;7–13 exploredmethods of estimating the effective sample sizewhen using NIR fiber-optic probe for spectralmeasurement of powder blends;14,15 and useddesign of experiments with varied parameters oncritical process variables for a process analyticaltechnology (PAT) approach to assessing powderblending uniformity.16–18 It is important to notethat most applications of NIR for blend monitor-ing typically rely upon a single sample pointmonitored over time as a means of demonstratinghomogeneity; this is in contrast to conventionalmethods which employ several sample locations ata single time point. The typical NIR approachassumes that the sample probed at the singlepoint at different times adequately representativeof the contents throughout the blend; thus, itimplies that the measurement location is repre-sentative of the system as a whole. It is thereforecritical that the measurement point is either themost representative or the most variable region(i.e., last to reach homogeneity). Therefore, thechoice of probe location is critical for the accurateblending process control by NIR.

Knowledge of the behavior of materials duringblending is critical to developing an adequateunderstanding of the blending process and sub-sequent process control. Despite the surge ofacademic interest in industrial applications usingsingle-point NIRS for powder blending evalua-tion, there is still much to be learned about thedynamic component behaviors during blending.Few models exist for prediction of powder mixingbehaviors and fewer empirical studies havebeen conducted utilizing common pharmaceuticalmaterials to justify these models.13,19 Processunderstanding gained by blend models andaccurate process control based on such under-standing are critical in the development of processanalytical technology (PAT). Therefore, informa-tion on spatial distribution of each component

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008

during blending is initially the key to gain blend-ing process understanding. To achieve visualiza-tion of the spatial distribution of components,and to further delineate behavior of materialsduring blending, NIR chemical imaging is pro-posed, since it provides chemical informationand spatial distribution of components within asample simultaneously.20,21 While imaging is notexpected to be used as a routine powder blendmonitoring tool, it is valuable in developing anunderstanding of the best use of single point NIRsystems.

The current surge in interest in NIR chemicalimaging has been spurred by the commercialavailability of liquid crystal tunable filters (LCTF)and Focal Plane Array (FPA) detectors operatingin the NIR spectral region. These are key compo-nents for imaging systems capable of collectinghigh signal to noise images in relatively shorttimes.22 These technological advances have enabl-ed NIR chemical imaging to meet the challenginganalytical needs of dosage form developmentlaboratories and to serve as a supplement oralternative to conventional, non-imaging NIRspectroscopy.23

The number of applications of NIR chemicalimaging in pharmaceutical analyses has increas-ed significantly in the past few years. Theversatility of this technique is indicated in manyapplications including: root cause investigationsat drug dosage forms with performance failureback to manufacturing processes,24–26 identifica-tion of the composition of multiple tablets throughblister packaging,27 visualization of chemicalcomposition and reaction kinetics,28 and, recently,determination of quality for drug products pur-chased from internet.29

However, among these applications, very littleresearch has been conducted using NIR chemicalimaging to understand and characterize pharma-ceutical powder blending processes. Lyon et al.30

evaluated the blend homogeneity by imagingtablets made from different grades of manuallyblended powder blends. El-Hagrasy et al.31 usedimaging as a supplementary tool to confirm theblending end point determination by the singlepoint NIR method. To date, no accounts of NIRimaging to gain critical knowledge of dynamicbehavior of components during blending havebeen reported.

NIR imaging is used in the present study tocharacterize blending process qualitatively andquantitatively via a novel sample preparationand image data collection method. Ten identical

DOI 10.1002/jps

NIR CHEMICAL IMAGING 3307

three-component mixtures, consisting of APAP(the model API), MCC, and lactose, were used torepresent ten blending trials blended for differentperiods of time. At each blending trial, aftermixing, images of the blend surface were collected;then, the mixture was compressed into a compactand the bottom and cross section of the compactswere imaged. By this means, the distribution ofcomponents in a significant fraction of a blendcould be assessed. Imaging of multiple locations,especially at internal cross sections of the blends,is crucial for an accurate characterization ofblending process. It has been reported previouslythat mixing patterns at internal cross-sectionsof mixtures vary significantly from patternsobserved from the exposed surfaces (top andbottom), indicating that blending patterns observ-ed at mixture surfaces are not representative ofmixing processes inside the powder bed.32

Multivariate analysis and chemometric toolswere applied to the collected images resulting in aqualitative description of the blending dynamicsacross the blending trials and quantitativedetermination of the blending pattern. The critic-al blending areas where the component distribu-tion changes most, or the areas that have thegreatest potential to remain non-uniform duringblending, were identified. A good correlationbetween imaging and a traditional UV monitoringmethod, in terms of end point determination, wasestablished, demonstrating the applicability ofthe NIR calibration. The API distribution patternat different locations was detected, and the APIdomain size distribution variation of differentblending trials points was demonstrated.

It is demonstrated from this study that NIRchemical imaging is an effective tool to betterunderstand blending characteristics and mechan-isms. Blending process understanding gainedthrough these characterizations by NIR imag-ing yields valuable knowledge for understandingbehavior of blending constituents during blending(i.e., formulation development). Additionally, thesedata are used to optimize single-point NIR blendmonitoring method, and develop blend monitoringapplications (i.e., sensors for PAT).

EXPERIMENTAL SECTION

Materials

A three-component powder mixture consisting of5% wt/wt Compap1 (Acetaminophen USP 90%,Mallinckrodt, Inc., Raleigh, NC) as the active

DOI 10.1002/jps JOUR

pharmaceutical ingredient (API), 31.7% wt/wtAvicel PH200 (MCC, FMC Biopolymer, Mechan-icsburgh, PA), and 63.3% wt/wt Fast-Flo1 lactose(Foremost Farm USA, Rothschild, WI) as the twoexcipients was used for all time points. Meanparticle sizes of Compap1, MCC, and lactose were�80, 180, and 100 mm, respectively. Methanol(Fisher Scientific, Pittsburgh, PA), which wasused in the UV assay, was optima grade.

Mixing

A 6 cm (tall) by 4 cm (diameter) cylinder-shapedaluminum mini-blender (made in-house) was usedas the blend vessel. A quartz window was designedfor sealing the top end and the bottom end was leftopen for loading blending materials. The blenderwas charged with lactose, MCC and API throughthe open bottom end for all ten time points. Onceloaded, the mini-blender bottom was sealed by acork stopper and the vessel was transferred andsecured into a lab-scale bin blender (L.B. Bohle,Maschinem þVerfahrem GmbH, Ennigerloh,Germany) rotated at 25 rpm. A total of tentrials (time-points) were conducted; each usedthe three-component powder mixture as pre-viously described. The blend time for each trialwas: 0.5, 1, 2, 5, 10, 15, 20, 25, 30, and 40 min. Ateach time point, one mixture was blended andsampled for both UV and imaging data collection.For each blending trail, the total weight ofmixture was 30 g (�70% capacity of the blendvessel volume).

Preparation of Compacts for Imaging

After surface images were collected, powder blendof each trial was compressed into a compact inthe mini-blender using a Carver Press (Fred S.Carver, Inc., Menomonee Falls, WI) at 1000 lbscompression force and a 30 s dwell time. The1000 lbs compression force was determined bytrial and error as the optimal compression forcerequired solidifying the blend, to make theresultant compact solid enough to endure sub-sequent cutting, but not to disturb the blendingpattern of the constituents.

Imaging and UV Data Collection

Near-Infrared Chemical Imaging Data Collection

The bin blender was stopped at every predeter-mined time point (according to each blendingtrial) and the mini-blender was removed.

NAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008

Figure 1. (a) Borders of concatenated images super-imposed upon the surface of the blend top (�60% ofsurface imaged). (b) Blend contents were compressedinto a compact allowing the bottom of the blend to beimaged. (c) Borders of concatenated images superim-posed upon the cross section of the compact after thecompact was cut into halves (�70% of surface imaged).


A MatrixNIRTM (Malvern, Inc., Olney, MD)imaging system was used for image data collec-tion. The field of view (FOV) of the present study,17.2� 21.5 mm, was selected based on a priorinvestigation.33 The wavelength range was select-ed from 1400 to 1675 nm with 5 nm increment.

Six images were collected for each blend. Thefirst two images were collected horizontally acrossthe quartz window of the mini-blender to covera strip view of the blend top, as illustrated inFigure 1a. Mapping of the blend top is achiev-ed via movement along the x-axis of a three-dimensional translation sample stage. The x- andy-direction stages control positioning of thesample in the FOV and the z-direction axis isused for optical focusing. Movement along y- andz- axes is controlled by AcuudexTM linear stages(Aerotech, Inc., Pittsburgh, PA). Based upon theFOV, the concatenated image of the blend topcovered about 60% of the blend top surface area.The next two images were collected from thebottom of the blend after the blend was com-pacted; the same percent of surface coverage wasachieved as images collected from the blend topsurface. The blend compact after compression isshown in Figure 1b. The last two images werecollected from the cross section of the blend afterthe compact was cut into halves; these imagescovered about 70% of the cross section surfacearea after concatenation, as illustrated inFigure 1c. Light and dark reference imageswere collected from 99% reflectance standard(Referencelight) (Labsphere1, Cranfield, UnitedKingdom) and polished stainless steel plate(Referencedark) for every blending trial, respec-tively. Pure component images of each of thecompacted three blending components afterwere collected. Although Compap1 was used asthe API, which contains 10% w/w starch, theimage of this constituent is referred to as a purecomponent image in this study. Overall, sixtyimages were collected over the ten blendsthroughout the blending process; additionally,ten light and ten dark images were collected forthe ten blending trials, and three pure compo-nents images were collected for reference andmodel development purposes. Each single imagehas dimension of 256� 320 pixels, with the totalnumber of pixels in one image of 81920.

UV Data Collection

Following the image data collection, three UVsamples were collected from three locations of


each compact. The first UV sample was collectedby scraping off powder from the cross section ofthe compact and was termed as center sample.The second UV sample was collected from one

DOI 10.1002/jps


edge of the compact and was termed as edgesample. The third UV sample was collected from arandom section of the blend compact. The mass ofeach UV sample was �1 g.

Image Pretreatment

Before image preprocessing tools were applied, allimages were converted to log (1/R) absorbanceunits using the light and dark images collected atthe corresponding blending trial (time point)according to the equation below:

ImageFinal

¼ � logImageSample � ReferenceDark

ReferenceLight � ReferenceDark

� �

Here, ImageFinal, ImageSample ReferenceDark, andReferenceLight are the processed sample image,raw data, dark and light reference images, res-pectively. Standard Normal Variate (SNV) andSavitsky-Golay smoothing (9-point window, noderivative, second order polynomial) were appliedin the spectral dimension as to enhance contrastamong different constituents in the images. Afterpreprocessing, single images from the same locat-ion were concatenated to one continuous image,that is, the two images collected from the blend topwere concatenated into one ‘top’ image. The sameconcatenation operation was performed on imagesof other locations (bottom and cross section). Allimage preprocessing and image analyses wereperformed using Matlab 7.0.1 software (Math-works, Inc., Natick, MA) with PLS toolbox(Eigenvector, Inc., Seattle, WA) and customprograms written in-house.

Figure 2. Relative standard deviation (RSD) profilescalculated from UV-Vis (dotted line) and imaging ana-lysis (broken line) of API concentrations of the 10 three-component blend trials. Data are plotted as time ofblending for each trial.

Acetaminophen Assay

At each time point, the UV samples were weighedand dissolved with methanol and diluted tovolume with distilled water. Acetaminophenconcentrations were detected using a HP 8543UV-Vis spectrometer (Agilent Technologies, Inc.,Wayne, PA) at 280 nm. The calibration curve foracetaminophen (absorbance versus concentra-tion) was linear over the range 0–100 mg/ml(r2> 0.999). The precision of the method wasexamined by preparing the calibration on 5different days (RSD< 2%). Recovery was calcu-lated to be between 95% and 101%.


RESULTS AND DISCUSSION

Ultraviolet Spectroscopy Analysis

UV spectroscopy is a commonly used method formonitoring of powder blending. Relative standarddeviation (RSD) among multiple samples is acommon means of determining the end point of ablend. If RSD value is within specification, theblend is considered adequately homogeneous.34

The RSD profile of each of the ten blend trials, asmeasured by UV-Vis analyses, is illustrated inFigure 2 (the RSD profile of the imaging data willbe discussed in the next section). The RSD iscalculated from the assay of the samples from thecenter, edge and a random part of the compact.The RSD of each trial decreases with increasingblending time with the exception of a spike duringtrial 5 (at 10 min). The spike indicates that thereexists a high variance among the three UVsamples collected at this particular time point.However, the cause of this high variance can notbe explained from the UV data. The initial RSDvalue changes are rapid with a local minimum atthe trial blended for 15 min.

The same overall blending trend is observed inFigure 3; here, API concentration is plotted foreach trial. While the minimum and maximum APIsamples varied in concert with the RSD of thesample, the mean API content reached thenominal level (�5%) with very little blendingtime. Figures 2 and 3 indicate that the system wasrelatively well blended at (and after) 15 min. APIconcentration between center sample and edgesample at each trial is also compared in Figure 3.It is readily observed that all of edge samples


Figure 3. API concentration change across blend trials from UV-Vis data, plottedagainst time of blending for each trial. The minimum (closed diamonds), mean (solidcircles), and maximum (closed squares) API concentrations are plotted along with theAPI concentration of the edge sample (cross) and API concentration of the centersample(asterisk).


corresponded to the maximum API%; while mostof the center samples were represented by theminimum API%. This indicates an initial ten-dency for API to be distributed at the periphery ofthe blend vessel. The data in Figure 3 describe, ina very general way, the pattern of API distributionin the different trials.

The UV monitoring method suggests an appro-priate blending time to reach a pseudo-steadystate. By segregating samples from the edge andthe middle, a general description of the APIdistribution is developed. However, this methoddoes not provide adequate detail with respect tospatial resolution to begin to understand thebehavior of the materials in this system andconsequently, the blending process itself. There-fore, a more detailed description of blendingtendencies was sought via NIR imaging.

Near-Infrared Chemical Imaging

Near-infrared chemical imaging was used todescribe the distribution of the model API andexcipients during the blending process. PCA and


discriminate PLS in the spectral dimension wereused to describe the spatial distribution of theblend components both qualitatively and quanti-tatively. These visualizations over the ten trialsillustrate the general blending behavior of thissystem.

PCA Score Images

Principal component analysis was applied toimage data. The components were calculatedusing a spectral matrix consisting of spectraextracted from the three preprocessed purecomponent images. A total of 2400 spectra wereused to calculate the basis vectors; this matrix wascomposed of 800 spectra from each pure compo-nent image. The first principal component (PC)explained 95.69% of the spectral variance; thesecond principle component explained an addi-tional 3.97% of the spectral variance. Thus, a PCAmodel containing these two PCs was developedand used to generate score images for all thepreprocessed blend images by projecting the blendimage spectra onto the model. The score images onthe first PC are used to display the dynamic

DOI 10.1002/jps


blending trend of the three blending constituentsfor the ten trials.

The loading spectrum on the first PC is plottedin Figure 4 along with the API spectrum from themean-centered pure components spectra libraryand the differential spectrum between API andMCC. It indicates that the loading spectrum onPC1 resembles the mean-centered API spectrum,but is most likely the differential spectrumbetween API and MCC (spectrum of API—spectrum of MCC), as the correlation coefficientbetween this loading spectrum and the differen-tial spectrum is 0.999 along the whole wavelengthrange used. This is because PCA calculates thelargest variance in the spectral matrix, and APIand MCC spectra are the least similar, it isexpected that their spectra difference will besignificantly represented in early PCs, especiallythe first PC. The loading spectrum on PC2, whichcaptured less than 4% of the total variance,resembles the differential spectrum betweenlactose and MCC (plot not shown), thus this PCcan distinguish between the two excipients, butnot the API. So, the second PC does not have thediscriminating power to differentiate among thethree components. Therefore, only score imageson PC1 were analyzed in this study.

By projecting the selected PC (PC 1) onto theimage data, pixels with high localized APIconcentration results in higher score values asthe mean centered API spectra are similar to thebasis vector (PC1, as demonstrated in Fig. 4);whereas pixels with high localized MCC concen-tration will have the highest negative score values

Figure 4. Comparison of loading spectrum on PC1with the API spectrum from mean-centered pure com-ponent spectra library and the differential spectrumbetween API and MCC.


as those spectra possess the greatest similarity tothe opposite of the basis vector. The projectedscore values for pixels with localized high lactoseconcentration are between the two extremes. As aresult, a pseudo-assignment for specific pixels inthe score images was established. In all the scoreimages, the white pixels (high positive score valuepixels) represent API, dark pixels (high negativescore value pixels) represent MCC, and graypixels (score values in the middle) representlactose.

Verification of the pseudo-assignment of blendconstituents to score ranges was accomplished bycomparing the spectra of the selected pixels (high,medium, and low scores) with the pure componentspectra, as shown in Figure 5. Spectra of pixelswith medium and low score values possess highfidelity to the pure excipients (lactose and MCC).Spectra of pixels with high score values showedhigh correlation with the API spectra, althoughspectral contribution from the excipients can beobserved in the wavelength region 1450–1550 nm,as demonstrated in Figure 5. The assignment ofconstituents is thereby justified and will form thebasis for subsequent descriptions of blendingbehavior.

Qualitative delineation of the blending dynamicsof each component is illustrated in the resultantscore images. Each image in Figure 6a–g consistsof three concatenated score images of the threelocations of the blend compact (top, cross section,and bottom) at blending time points 0.5, 1, 2, 5, 10,15, and 40 min, respectively. Score images after15 and before 40 min have similar distributions ofconstituents and are not displayed here.

The primary concern of most blend analyses isthe distribution of API; here, images of the blendand compact from each trial describe the changesin distribution of API during blending. The initialphase of blending was characterized by convectivemixing. The API was charged last into the vessel,and through the first two trials (0.5 and 1 min),Figure 6a and b; it remains primarily on thebottom and top of the blend vessel. Convectiondistributes the API through the next three trials(2, 5, and 10 min) and the API is distributed alongthe outer surfaces of the blend parallel. Distribu-tion of API along the outer portions of the blendvessel is illustrated in Figure 6c–e. Note that thecross section image in these figures highlights thepresence of API at the perimeter of the blendvessel and indicates less API content in the centerof the blend. This observation is consistent withthe earlier UV-Vis data. The steady state of API


Figure 5. Spectra of different intensity pixels in a sample score image (a) comparedwith blending constituents spectra in the powder mixture (b).


distribution after 15 min is demonstrated inFigure 6f and g, which provides the cause of thestable trend found at the same period of time inFigures 2 and 3.

The PCA images also demonstrate the distribu-tion of excipients for different blending times. Inthese experiments, MCC is charged after lactoseand before API. Isolation of MCC is expected andobserved in the cross section from the beginning ofblending through 10 min (trials 1 through 5). Thevolume of the isolated MCC mass is graduallyreduced as blending proceeds, and eventuallydissipates at 15 min. While API distribution is


frequently the only consideration during blendinganalysis, the appropriate concentration of exci-pients is important to the manufacturability andperformance of the dosage form. As a result,segregation of excipients should be treated withgreat caution. The API concentration profiles inFigure 3 give no indication of the poor blending ofexcipients. Thus, insufficient mixing of excipientsobserved in the images can not be detected fromthe UV data. This practice is not atypical for blendmonitoring in drug products.

Particle properties, such as size, density, shape,friction coefficient and cohesivity play a role in

DOI 10.1002/jps

Figure 6. Concatenated PCA score images of selected blending trials of three locations(T¼ blend top surface, C¼ cross section surface of blend compact, B¼ bottom surface ofblend compact). (a) trial 1(t¼ 0.5 min); (b) trial 2 (t¼ 1 min); (c) trial 3 (t¼ 2 min); (d) trial4 (t¼ 5 min); (e) trial 5 (t¼ 10 min); (f) trial 6 (t¼ 15 min); (g) trial 10 (t¼ 40 min). Pseudoassignments of pixel intensity indicate that white pixels (corresponding to high value inthe color scale) represent the API; dark pixels (corresponding to low value on the colorscale) represent MCC; and gray pixels (corresponding to medium values on the colorscale) represent lactose, or blends of API and MCC. Units of on the x and y axes arenumber of pixels. Units of the color bar are in arbitrary units.

DOI 10.1002/jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008



segregation.35 However, in this system the sug-gested cause for MCC segregation is the mechan-ism of radial mixing. It has been reported thatradial mixing occurs as the mixture winds arounda central point at the interface between thecascading layer and material undergoing solid-body rotation.36 Since MCC was loaded in themiddle, the bulk of MCC powder served as therotating solid-body due to its location in the blendmixture at early blending times. With increasingrotations, the MCC segregation is reduced dueto increased movement-induced diffusion of theMCC particles into the bulk powder body, leadingto decreased MCC content stagnant in the centerregion.

As further example of the utility of NIRimaging, the trial at 5 min demonstrated a highMCC concentration and a small accumulation ofAPI next to it in the cross section (Fig. 6d). Inthe particular state of blending illustrated byFigure 6d, a sample collected for UV-Vis analysiswhich included the high concentration of MCCadjacent to a local aggregation of API may containthe correct concentration of API; however, adosage form resulting from this particular samplewould be quite different from a dosage formcontaining nominal concentrations of all consti-tuents. The UV-Vis analysis for this samplecannot indicate or imply the degree of homo-geneity of the API distributed within the sample.This image suggests that a variety of samplesshould be collected to reflect the overall blending

Figure 7. Predicted APAP concentration (section(open circles), and bottom(asterisks) oplotted against blending time. Horizontal darkAPAP concentration in the mixture.


status of the powder bed in order to accuratelydetermine the blending end point when conven-tional blend monitoring method is used, whichsupports the sampling protocol in the FDAguidance that extensive sampling of the mixtureshould be performed when assessing the powdermixture uniformity.37

Root causes of abnormal values from the UVanalysis can also be sought through imaging. Forexample, images of the blending trial at 10 min(trial 5, Fig. 6e) demonstrate that more API wasdistributed at the area near one of the edges in thetop and bottom locations, which suggests thatone of the UV edge samples (termed ‘sample fromthe edge’) might have a greater probability ofcontaining excess concentration of API than theother two UV samples. It would be the direct causeof the high variance among the UV samples, thuscausing the spike in the RSD profile, at this timepoint (Fig. 2).

The score images in Figure 6 not only demon-strate the API distribution pattern but alsoprovide interpretations of the cause of the profilesobtained from the UV-Vis data. This qualitativeexample highlights the importance of utilizingNIR to understand the blending pattern ofcomponents. Based upon the PCA score images,a suggested end point for this system is �15 min.This is the first image in which no visiblesegregation of any component is apparent. Theanalysis of images agrees with the local minimumobserved in the UV-Vis data.

%) from images of top (crosses), crossf the blend compact at each blend trial,

blue solid line indicates the 5% nominal

DOI 10.1002/jps

Figure 8. Predicted APAP concentration variability within locations during blendingprocess for ten blending trials calculated from imaging data.

DOI 10.1002/jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 8, AUGUST 2008


Figure 9. Relative standard deviation (RSD) profilesfor ten blending trials based on APAP concentrationvariability within locations of top (crosses), cross section(open circle) and bottom (asterisks) of blend compact.


Quantitative Analyses of PLS Predicted Images

Predicted API Concentration Profiles. A discriminatecalibration model using non-linear iterative partialleast squares (NIPLS) algorithm was calculated fromthe same pure component spectral matrix used forPCA. Using this algorithm, spectral information ofeach pixel is converted into a quantitative measure-ment of concentration. The coding regimen used inthis NIPLS discriminate analysis (DA) is 0 and 1, sothat the component concentration in a predictedimage is proportional to 1 depending on the spectralinformation of the pixels.

In this study, quantitative analyses werefocused on the predicted API images of the threelocations (top, cross section, and bottom) along theblending time. The profile of predicted mean APIconcentration at the three locations for each trialis illustrated in Figure 7. It is evident that the APIconcentration was persistently higher in thebottom and top of the vessel, whereas less APIwas constantly present in the cross section. Theprofile suggests that higher API concentrations atthe two ends of the powder bed were induced bythe initial convective mixing, whereas the dis-tribution of API into the cross section can only beachieved through diffusion mixing mechanism.38

In this study, the diffusion of API into the cross-section took about 15 min, as indicated by the timerequired for the difference of API content ofthe three locations to reach a minimum (similarto qualitative observations in Fig. 6 and UV-Vis results in Figs. 2 and 3). This quantitativeanalysis of the data provides further evidence thatthe blending material loading order is a contribut-ing factor to blending efficiency and a potentialrate-limiting factor in achieving blend homoge-neity. Further, NIR imaging investigations on thespecific effects of the loading order on the materialdistribution pattern and the blending efficiencyare ongoing.

Localized Heterogeneity of Predicted API Concen-tration. API distribution variability across dif-ferent blender locations was studied to betterunderstand blending behavior. The API distribu-tion variability within locations was depicted bycomparing the predicted API concentrations offour areas within each location (represented byone concatenated image) across the ten trials. Forexample, the image of the top of the blend wasdivided along the longest axis into four equalsections (left edge, left-middle, right-middle, rightedge). This division allows the comparison of


homogeneity at the two edges and the center areasof each blend compact.

The resultant comparison of the four predictedAPI concentrations and the mean concentration ateach location along blending time is demonstratedin Figure 8. The white bar and the bar withupward diagonal pattern represent the two edgeareas of the blend (left and right, respectively); thebars with horizontal and downward diagonalpatterns represent largely the center areas ofthe blend. Figure 8 illustrates that API concen-tration is consistently higher at the edges (leftand right) at trials blended less than 15 min. After15 min, the concentration disparity decreasedat each of the three locations. This observationis consistent with phenomena illustrated inFigures 3, 6, and 7, which depict higher APIconcentration at the edges of the vessel when themixtures are blended for less than 15 min.

RSD Profiles Based on Predicted API Concentration.The RSD profiles of the predicted API concentrationin the three locations (top, center, and bottom) areshown in Figure 9. The RSD values were calculatedfrom the API content of all pixels within a location,and are indicative of the intra-region heterogeneity.The three profiles demonstrated similar trends toUV-Vis analysis over the ten trials, including a spikeof high RSD value at 10 min.

Comparison of UV-Vis analysis with imageanalysis demonstrates the similarity of thetechniques when viewed at similar scales. ARSD value for the image data was calculatedfrom all pixels in the top, middle and bottom(N¼�480000). These values were then comparedto the RSD values calculated from UV-Vis

DOI 10.1002/jps


analysis in Figure 2. Similar RSD profiles areobserved for both measurements; notably, bothprofiles reach their respective lowest value at15 min. The relatively larger magnitude of theRSD profile obtained from the imaging data is dueto the large number of data points included in theRSD calculation.

API Domain Distributions. API domain sizedistribution profiles were calculated from thepredicted API images using customized programswritten in-house, which incorporated morpho-

Figure 10. APAP domain size distribution ablend compact for trials 1 and 6 (0.5 and 15 m


logical analysis functions from Matlab1 ImageProcessing Toolbox. The term ‘API domain size’refers to the sizes of the areas containing localizedhigh concentration of API. The API domain sizedistributions at the top, cross section and bottomof the blend for trials 1 and 6 (blended for 0.5 and15 min, respectively) are illustrated in Figure 10.As evident for trial 1, the API domain distribu-tions at the three locations are different. Note theAPI accumulation at the cross section of trial 1(0.5 min) which results in a single large domainof approximately 1000 mm. In contrast, in trial 6

t the top, cross section and bottom of thein).



(15 min), the API domain size distributions at thethree locations appear similar. The similar APIdomain size distributions throughout the blendat 15 min are consistent with other imaging andUV-Vis analyses.

High Variability Regions Within the Blending

Vessel. Variability at specific locations within theblend vessel was characterized by examining fixedlocations through the ten trials. Images wereconstructed by calculating the standard deviationof each pixel across all ten trials in an effort tovisualize the variability in blending homogeneityof regions within the blending vessel. The result-

Figure 11. Images of pixel standard deviatithe top (a) cross section (b) and bottom (c) of theare number of pixels. Units in the color bar a


ing images are shown in Figure 11. The higher SDvalue a pixel has, the whiter (or brighter) itappears in the constructed image. Figure 11indicates that the edges (top, cross-section, andbottom) demonstrated higher variability thanother locations in the blend vessel across thetrials (different blend times). The center of thecross section highlights an MCC domain as aregion of relatively intense change (increased SD)due to the movement of this domain during theblending process. Data in Figure 11 indicate thatthe two edges and the center of the blend areregions critical to assessing blend homogeneity asa function of time.

on calculated through blending trials forblend compact. Units on the x and y axes

re in arbitrary units.

DOI 10.1002/jps


CONCLUSIONS

In this study, NIR chemical imaging was used todemonstrate distribution of blend componentsand chemical concentration during trials simulat-ing a small scale blending process. Imaginganalyses demonstrate the distribution of blendcomponents and chemical abundance simulta-neously, thus providing previously unavailabledetail about the blend constituents. The studyalso suggests that for on-line blend monitoring, asingle sensor placed at the edges of the blendvessel would be most appropriate, since high APIdistribution variability was found in these areas,as demonstrated in Figures 8 and 11. Addition-ally, the imaging of the cross-section of each trialillustrated distribution of MCC from the center ofthe blend, highlighting the importance of mon-itoring non-API constituents during blending.These characterizations of blending are not achiev-ed by traditional blend monitoring methods.

In this study, ten independent powder mixtureswere used for both UV and imaging investigation.For the ten individual blends, consistency ofcertain blending behaviors was observed, whichwas strongly indicated by the MCC segregation inthe cross section and the declined RSD profilesover time from both UV and imaging data. Theblending mechanism observed throughout the tenblending trials in the mini-blender is alsoconsistent with what has been reported in theliterature using large scale blending. Hence, itsuggests that the blending insights obtained fromthis study may be transferable when studyinglarger scale blending process using powdersystems with similar particle properties (particlesize range, shape, flowability, cohesivity, etc.)blended in a device of similar shape as the blenderused in this small-scale blending.

The results from this study underscore thevaluable role of NIR chemical imaging in achiev-ing a fundamental understanding of pharmaceu-tical processes and optimization of processmonitoring, both of which are critical to processanalytical technology (PAT).

ACKNOWLEDGMENTS

The authors would like to take this opportunity tothank Dr. James K. Drennen for his valuablediscussion on the blending experiment part ofthe present study. We would also like to thank


Dr. Robert P. Cogdill for his unreserved sugges-tion on the data analysis.

REFERENCES

1. Berman J. 1995. Blend uniformity and unit dosesampling. Drug Dev Ind Pharm 21:1257–1283.

2. Berman J, Shoeneman A, Shelton J. 1996. Unit dosesampling: A tale of two thieves. Drug Dev IndPharm 22:1121–1132.

3. Muzzio FJ, Robinson P, Woghtman C, Brone D.1997. Sampling practices in powder blending. IntJ Pharm 155:153–178.

4. Muzzio FJ, Goodridge CL, Alexander A, Arratia P,Yang H, Sudah O, Mergen G. 2003. Sampling andcharacterization of pharmaceutical powders andgranular blends. Int J Pharm 250:51–64.

5. Chang R, Shukla J, Buehler J. 1996. An evaluationof a unit-dose compacting sample thief and a dis-cussion of content uniformity testing and blendingvalidation issues. Drug Dev Ind Pharm 22:1031–1035.

6. Harwood C, Ripley T. 1977. Errors associated withthe thief probe for bulk powder sampling. J PowderBulk Solids Technol 11:20–29.

7. Sekulic SS, Ward HW, Brannegan DR, Stanley ED,Evans CL, Sciavolino ST, Hailey PA, Aldridge PK.1996. On-line monitoring of powder blend homoge-neity by near-infrared spectroscopy. Anal Chem68:509–513.

8. Maesschalck RD, Sanchez FC, Massart DL, Doh-erty P, Hailey P. 1998. On-line monitoring of pow-der blending with near-infrared spectroscopy. ApplSpectrosc 52:725–731.

9. Sekulic SS, Wakeman J, Doherty P, Hailey PA.1998. Automated system for the on-line monitoringof powder blending processes using near-infraredspectroscopy. Part II. Qualitative approaches toblend evaluation. J Pharm Biomed Anal 17:1285–1309.

10. Hailey PA, Doherty P, Tapsell P, Oliver T, AldridgePK. 1996. Automated system for the on-line mon-itoring of powder blending processes using near-infrared spectroscopy. Part I. System developmentand control. J Pharm Biomed Anal 14:551–559.

11. Berntsson O, Danielsson LG, Johansson MO, Foles-tad S. 2000. Quantitative determination of contentin binary powder mixtures using diffuse reflectancenear infrared spectrometry and multivariate ana-lysis. Anal Chim Acta 419:45–54.

12. Berntsson O, Danielsson LG, Lagerholm B, Foles-tad S. 2002. Quantitative in-line monitoring ofpowder blending by near infrared reflection spec-troscopy. Powder Technol 123:185–193.

13. Ufret C, Morris K. 2001. Modeling of powder blend-ing using on-line near-infrared measurements.Drug Dev Ind Pharm 27:719–729.



14. Berntsson O, Danielsson LG, Folestad S. 1998.Estimation of effective sample size when analyzingpowders with diffuse reflectance near-infraredspectrometry. Anal Chim Acta 364:243–251.

15. Cho JH, Gemperline PJ, Aldridge PK, Sekulic SS.1997. Effective mass sampled by NIR fiber-opticreflectance probes in blending processes. Anal ChimActa 348:303–310.

16. El-Hagrasy AS, Delgado-Lopez M, Drennen JK.2006. A process analytical technology approach tonear-infrared process control of pharmaceuticalpowder blending. Part II. Qualitative near-infraredmodels for prediction of blend homogeneity.J Pharm Sci 95:407–421.

17. El-Hagrasy AS, D’Amico F, Drennen JK. 2006.A process analytical technology approach tonear-infrared process control of pharmaceuticalpowder blending. Part I. D-optimal design for char-acterization of powder mixing and preliminaryspectral data evaluation. J Pharm Sci 95:392–406.

18. El-Hagrasy AS, Drennen JK. 2006. A process ana-lytical technology approach to near-infrared pro-cess control of pharmaceutical powder blending.Part III. Quantitative near-infrared calibrationfor prediction of blend homogeneity and character-ization of powder mixing kinetics. J Pharm Sci95:422–434.

19. Portillo P, Muzzio FJ, Ierapetritou MG. 2006. Char-acterizing powder mixing processes utilizing com-partment models. Int J Pharm 320:14–22.

20. Lewis EN, Carroll JE, Clarke F. 2001. NIR imaging:A near infrared view of pharmaceutical formulationanalysis. NIR News 12:16–18.

21. Lewis EN, Schoppelrei J, Lee E. 2004. Near-infrared chemical imaging and the PAT initia-tive-NIR-CI adds a completely new dimension toconventional NIR spectroscopy. Spectroscopy 19:26–36.

22. Tran CD. 2005. Principles, instrumentation, andapplications of infrared multispectral imaging, anoverview. Anal Lett 38:735–752.

23. Reich G. 2005. Near-infrared spectroscopy and ima-ging: Basic principles and pharmaceutical applica-tions. Adv Drug Deliv Rev 57:1109–1143.

24. Clarke F. 2004. Extracting process-related informa-tion from pharmaceutical dosage forms usingnear infrared microscopy. Vibrational Spectrosc34:25–35.

25. Roggo Y, Jent N, Edmond A, Chalus P, Ulmschnei-der M. 2005. Characterizing process effects onpharmaceutical solid forms using near-infraredspectroscopy and infrared imaging. Eur J PharmBiopharm 61:100–110.


26. Roggo Y, Edmond A, Chalus P, Ulmschneider M.2005. Infrared hypersepctral imaging for qualita-tive analysis of pharmaceutical solid forms. AnalChim Acta 535:79–87.

27. Malik I, Poonacha M, Moses J, Lodder RA. 2001.Multispectral imaging of tablets in blister packa-ging. AAPS Pharm Sci Tech 2:1–7.

28. Tran CD. 2000. Visualising chemical compositionand reaction kinetics by the near infrared multi-spectral imaging technique. J Near Infrared Spec-trosc 8:87–99.

29. Westenberger BJ, Ellison CD, Fussner AS, JenneyS, Kolinski RE, Lipe TG, Ylyon RC, Moore TW,Revelle LK, Smith AP, Spencer JA, Story KD, TolerDY, Wokovich AM, Buhse LF. 2005. Quality assess-ment of internet pharmaceutical products usingtraditional and non-traditional analytical techni-ques. Int J Pharm 306:56–70.

30. Lyon RC, Lester DS, Lewis EN, Lee E, Yu LX,Jefferson EH, Hussain AS. 2002. Near-infraredspectral imaging for quality assurance of pharma-ceutical products: Analysis of tablets to assess pow-der blend homogeneity. AAPS Pharm Sci Tech 3:1–15.

31. El-Hagrasy AS, Morris HR, D’amico F, Lodder RA,Drennen JK. 2001. Near-infrared spectroscopy andimaging for the monitoring of powder blend homo-geneity. J Pharm Sci 90:1298–1307.

32. Wightman C, Mort PR, Muzzio FJ, Riman RE,Gleason EK. 1995. The structure of mixtures ofparticles generatedby time-dependent flows. Pow-der Technol 84:231–240.

33. Ma H, Anderson CA. 2007. Optimization of magni-fication levels for near infrared chemical imaging ofblending of pharmaceutical powders. J Near Infra-red Spectrosc 15:137–151.

34. Berman J, Elinsiki D, Gonzales C, Hofer J, JimenezP, Planchard J, Tlachac R, Vogel P. 1997. Blenduniformity analysis: Validation and in-process test-ing. Technical Report No25 PDA 51:S1–S99.

35. Alexander AW, Shinbrot T, Muzzio FJ. 2001. Gran-ular segregation in the double-cone blender: Tran-sitions and mechanisms. Phys Fluids 13:578–587.

36. Alexander A, Arratia P, Goodridge C, Sudah O,Brone D, Muzzio F. 2004. Characerization of theperformance of bin blenders part 1of 3: Methodol-ogy. Pharm Technol 28:70–86.

37. Food and Drug Administration. 2003. Powder blendsand finished dosage units- Stratified in-processdosage unit sampling and assessment. Guidancefor industry:4.

38. Hogg R, Cahn DS, Healy TW, Fuerstenau DW.1966. Diffusional mixing in an ideal system. ChemEng Sci 21:1025–1038.

DOI 10.1002/jps

Documents

Characterization of pharmaceutical powder blends by … · Characterization of Pharmaceutical Powder Blends by NIR Chemical Imaging ... Content uniformity of solid oral dosage