PHARMACEUTICAL TECHNOLOGY Near Infrared Spectroscopy of Magnesium Stearate Hydrates and Multivariate Calibration of Pseudopolymorph Composition JOHN F. KAUFFMAN,1 VENKAT TUMULURI,1 CHANGNING GUO,1 JOHN A. SPENCER,1 WILLIAM H. DOUB,1 GARY A. NICHOLS,2 SCOT R. RANDLE,2 STEPHEN WU2 1 FDA, CDER, Division of Pharmaceutical Analysis, 1114 Market St., St. Louis, Missouri 63101 2 Pharmaceutical R&D, Tyco Healthcare Group, Mallinckrodt Division, 385 Marshall Ave., St. Louis, Missouri 63119-1831 Received 5 December 2006; revised 2 February 2007; accepted 8 March 2007 Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/jps.21006 ABSTRACT: Samples of magnesium stearate monohydrate and dihydrate were used to prepare standard mixtures of known pseudopolymorphic composition. Near infrared spectra (NIR) of the standard mixtures were measured to develop multivariate calibration models for the pseudopolymorphic composition of magnesium stearate by partial least squares (PLS) regression. Magnesium stearate hydrate compositions of the standard mixtures were compared against the hydrate composition based on thermogravimetric analysis (TGA). The mixture compositions determined from TGA mass loss on drying (LOD) measurements were found to be inaccurate. PLS regression was applied to the TGA thermograms of the standard mixtures to generate more accurate reference values, and this model was then applied to a set of validation samples. Application of the NIR PLS model to the validation sample set resulted in precise estimates of sample pseudopolymorphic composition when compared to the TGA PLS reference values. The NIR PLS model was found to be more sensitive than TGA LOD to small quantities of hydrates, and the TGA PLS model was also found to be more sensitive that TGA LOD. The results demonstrate the challenges and opportunities that arise when rapid, nondestructive spectroscopic methods depend on insensitive or inaccurate reference methods for development of multivariate calibration models. ß 2007 Wiley-Liss, Inc. and the American Pharmacists Association J Pharm Sci 97:2757–2767, 2008 Keywords: hydrate; near infrared spectroscopy; partial least squares regression; multivariate calibration; pseudopolymorphs; thermogravimetric analysis INTRODUCTION Though magnesium stearate (MgSt) has been widely used as a solid lubricant in pharmaceutical Correspondence to: John F. Kauffman (Telephone: 314-5392168; Fax: 314-539-2113; E-mail: John.Kauffman@fda.hhs.gov) Journal of Pharmaceutical Sciences, Vol. 97, 2757–2767 (2008) ß 2007 Wiley-Liss, Inc. and the American Pharmacists Association tableting for decades, it continues to pose problems for formulators. Several studies have shown that the physical properties and lubricity of MgSt vary from manufacturer to manufacturer, and from lot to lot.1,2 MgSt has also been identified as a potential dissolution retardant, due to its propensity to coat the surface of polar particles of active pharmaceutical ingredients (APIs) and render their surfaces hydrophobic.3,4 To complicate matters, MgSt has been shown to reduce the JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 2757 2758 KAUFFMAN ET AL. tensile strength of tablets, and the magnitude of this effect depends on the mode of deformation (plastic or elastic) during compaction.5–9 Recently, Rao et al.2 have reviewed the influence of the bulk, particulate and molecular properties of MgSt on lubrication efficiency, and they argue that solid state properties at all three length scales must be controlled in order to predict the behavior of the lubricant. Both bulk properties (e.g., density, porosity and accessible surface area) and particulate properties (e.g., crystal habit, particle size distribution and aggregation state) influence lubrication efficiency by mediating the ability of the lubricant to disperse throughout the formulation, and interact with the surfaces of the tableting equipment. Some of the bulk and particulate properties (e.g., density and particle size distribution) can be adjusted by processing MgSt after manufacture, however the molecular level properties (e.g., fatty acid composition and hydration state) are generally established at the time of manufacture, and these in turn affect some of the bulk and particulate properties such as crystal habit and porosity. Thus the molecular level properties are expected to be critical quality attributes with respect to lubrication efficiency. It has been hypothesized that when MgSt has a high dihydrate content, the lubricity of the formulation is increased.1,2,10 The common physical explanation for this phenomenon is that the water of crystallization of magnesium stearate dihydrate (MgSt-D) increases the intermolecular distances between lamellar layers in comparison to magnesium stearate monohydrate (MgSt-M), and thereby reduces the resistance of the MgSt particles to shearing forces.1,10 However, the evidence supporting this hypothesis is primarily anecdotal. Though several studies have been performed that show a correlation between efficient lubrication and an increase in magnesium stearate dihydrate (MgSt-D) fraction, few studies are available in which the hydration state of MgSt is well controlled.1,2,10,11 One issue preventing the systematic investigation of the hydrate dependence of MgSt lubrication efficiency is the lack of a simple, robust method for making quantitative measurement of the pseudopolymorphic composition of commercially available magnesium stearates. Typically, powder X-ray diffraction (PXRD), differential scanning calorimetry (DSC), and thermogravimetric analysis (TGA) are used to establish the composition of mixed hydrates of MgSt.2,11 Though these methods are adequate to characterize the MgSt JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 composition, they are time consuming, insensitive to small variations in hydration state, and typically utilize very small amounts of material, leading to potential sampling problems. The purpose of this article is to develop near infrared (NIR) spectroscopy as a tool for quantitative assessment of the pseudopolymorphic composition of pharmaceutical grade MgSt. Samples of MgSt-M and MgSt-D were manufactured, and mixtures of these compounds were prepared as composition standards. NIR spectra of the mixtures were measured, and the spectra were calibrated to the known composition of the mixtures using partial least squares regression. The accuracy of the model developed with the standard samples was evaluated by applying it to the prediction of mixture compositions of several validation samples. The results indicate that NIR spectroscopy with chemometric calibration is an accurate and robust method for determination of the pseudopolymorphic composition of MgSt mixtures composed of the monohydrate and dihydrate. METHODS In this article, we adhere to the USP/NF definition of magnesium stearate, in which the fatty acid content of the starting material is no less than 40% stearic acid, and the sum of stearic and palmitic acids comprise no less than 90% of the material. The terms MgSt, MgSt-M, and MgSt-D denote the magnesium salts of these fatty acid mixtures. The reported hydrate compositions are computed from the average Molar Mass of the material based on the measured stearate and palmitate composition of each sample. Thus, 100% MgSt-M and MgSt-D denote samples with 1 and 2 water molecules per Mg ion, respectively, and the number of Mg ions is computed from the average Molar Mass of the fatty acid Mg salt mixture. Hereafter, the term mixture will denote mixtures of MgSt-M and MgSt-D, unless fatty acid mixture is stated explicitly. Preparation and Characterization of Hydrates MgSt monohydrate and dihydrate were prepared from stearic acid.11,12 The fatty acid composition of the MgSt was determined by gas chromatography as the fatty acid methyl esters as described in the USP/NF monograph for magnesium stearate. Samples of MgSt-M and MgSt-D were prepared, and characterized by PXRD, DSC, and TGA. DSC DOI 10.1002/jps NEAR INFRARED SPECTROSCOPY OF MAGNESIUM STEARATE measurements were performed with a TA Instruments Q100 DSC, over the 25–1508C range with a scan rate of 28C per minute. Aluminum crimp pans with pinhole lids were used for all measurements, and the instrument was calibrated with indium prior to the measurements. TGA measurements were made with a TA Instruments Q50 TGA using a scan rate of 58C per minute. The crystal structure of the MgSt-M and MgSt-D samples were confirmed by PXRD, measured with a Siemens D500 X-ray diffractometer over the range of 2u ¼ 28 to 408 using a 0.028 step size and 1 s per step integration time. The diffractograms were consistent with the published powder patterns of MgSt-M and MgSt-D.11 Preparation of Standards The complete sample set examined in this article consisted of 10 standard mixtures and 14 validation samples. After correcting for nonassociated water, the MgSt-M and MgSt-D stock materials used for standard mixture preparation were found by TGA to contain 92.0% monohydrate and 0% dihydrate, and 95.4% dihydrate and 0% monohydrate, respectively. Approximately 250 mg of each standard mixture was prepared by weighing the appropriate amount of MgSt-M and MgSt-D stock materials into a plastic mixing vial. The mixtures were then homogenized for 2 min by random mixing using a SPEX Mixer/Mill. Milling balls were not used during this operation to avoid altering the particle size distribution of the mixtures. Subsequent particle size measurements verified that the particle size was not altered by the mixing process. The homogenized mixtures were then delivered to clear glass vials for NIR analysis. Monohydrate/dihydrate standard mixtures were prepared from the 92.0% MgSt-M and 95.4% MgSt-D samples described above, with compositions of 0, 5.3, 10.2, 24.7, 39.8, 60.8, 75.0, 90.8, 95.0, and 100 percent MgSt-D by mass. The compositions of the resulting mixtures ranged from 92% to 0% MgSt-M and 0% to 95.4% MgSt-D, and are specified as ‘‘Prepared Composition’’ in Table 1. Standard mixture samples are labeled S1 through S10 in order of increasing MgSt-D composition, and validation samples are labeled V1 through V14 in Tables 2 and 3. All samples were prepared from fatty acid mixtures containing 66% stearic acid and 30% palmitic acid, except validation samples V1, V2, and V3, which were prepared from fatty acid mixtures containing 90% stearic acid and 5% palmitic acid. DOI 10.1002/jps 2759 The validation sample set includes three mixed hydrate samples that were coprecipitated in the manufacturing process (samples V1, V2, and V3), eight MgSt-M samples with median particle sizes varying from 3 to 10 microns (samples V4 through V11), and three MgSt-D samples with particle sizes varying from 2 to 3 microns (samples V12 through V14). Near Infrared Spectroscopy NIR spectra were measured with a ThermoNicolet Antaris Near Infrared Analyzer at 8 cm1 resolution. Each spectrum was generated from 16 scans in diffuse reflectance mode measured through the bottom of a clear glass vial. Near infrared spectra of the standards and coprecipitated validation samples were replicated four times. Two sets of replicate measurements were collected in a systematic order from one extreme in dihydrate concentration to the other, and two sets were collected in random order. One of the sets of randomly selected samples was measured after first shaking the glass vial, and then tapping it to compact it into the bottom of the vial. Replicate measurements were used to evaluate the impact of spectral variation on the calibration model, as well as the ability of spectral preprocessing to suppress spectral variation. Multivariate Calibration The standard mixture samples were used as the training set for calibration of NIR spectra and TGA thermograms to the pseudopolymorphic content of MgSt. Partial least squares (PLS) regression analysis was performed using Pirouette1 3.1(Infometrix, Inc., Bothell, WA). The PLS models were then used to predict the hydrate compositions of the validation samples. Details of the chemometric models and their application for prediction of MgSt hydrate composition are presented in ‘‘Results and Discussion.’’ Particle Size Determination The particle size distributions of all 24 samples were determined using a Malvern series 2600 laser diffraction system (Malvern Instruments Ltd, Malvern, Worcestershire, UK) equipped with a 63-mm lens (size range from 1.2 to 118 mm) and PS1 stirred cell. Isopropanol was used as dispersant agent. A small amount of a MgSt/ isopropanol slurry was gradually added to JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 2760 KAUFFMAN ET AL. Table 1. Mg Stearate Hydrate Compositions of the Standard Mixtures TGA LODb,c Prepared Composition Sample Namea S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 TGA PLS Validationd Median Particle Size (microns) Mass% Monohydrate Mass% Dihydrate Mass% Monohydrate Mass% Dihydrate Mass% Monohydrate Mass% Dihydrate 10.6 10.9 10.9 11.9 12.7 13.3 13.3 14.9 14.1 14.3 92.0 87.1 82.6 69.3 55.4 36.1 23.0 8.5 4.6 0.0 0.0 5.1 9.7 23.6 38.0 58.0 71.6 86.6 90.7 95.4 92.0 79.9 74.1 63.0 51.6 35.3 23.6 11.4 8.3 0.0 0.0 9.7 14.5 27.3 40.3 58.4 70.1 84.5 88.3 95.4 92.7 86.7 82.0 69.4 55.3 35.5 23.1 8.3 5.0 0.75 0.67 5.4 10.4 23.5 38.0 58.6 71.5 86.8 90.3 94.6 a Sample names are described in Section ‘‘Methods.’’ The fatty acid composition of these samples is 66% stearic acid and 30% palmitic acid. b Measured hydrate fractions were corrected for the mass of free water by dividing by one minus the fractional loss on drying due to unassociated water. c The theoretical loss on drying of pure monohydrate and dihydrate samples is 3.04% and 5.9%, respectively. d Predicted MgSt hydrate compositions from the leave-one-out cross validation computation. The extreme values of each hydrate show the largest residual error, 0.7%. This is expected, because the leave-on-out validation extrapolates the extreme values when they are left out of the calibration set. the stirred cell filled with dispersant until the obscuration was in the range of 0.2–0.3. The sample was stirred for 2 min before taking the measurement. The median particle sizes of the standard mixtures and validation samples are shown in Tables 1 and 2, respectively. RESULTS AND DISCUSSION The TGA traces for the standard mixtures and the coprecipitated validation samples are shown in Figure 1. The validation sample traces are offset by 2% for clarity. The MgSt-M sample loses Table 2. Mg Stearate Hydrate Compositions of the Validation Samples TGA LODb,c Sample Namea V13 V23 V33 V4 V5 V6 V7 V8 V9 V10 V11 V12 V13 V14 TGA PLS Median Particle Size (microns) Mass% Monohydrate Mass% Dihydrate Mass% Monohydrate Mass% Dihydrate 21.1 21.4 20.0 7.0 9.9 9.0 9.6 8.3 2.8 3.1 3.0 2.1 3.1 2.1 98.9 52.6 50.2 92.2 94.0 93.8 96.2 92.7 92.7 93.5 92.6 0.0 0.0 0.0 0.0 33.3 39.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 93.1 94.1 94.6 85.6 68.1 60.1 88.6 89.0 89.1 87.3 88.6 90.0 90.9 91.3 4.0 5.6 2.0 6.6 24.7 33.1 3.3 3.0 2.8 4.8 3.4 1.9 1.1 0.6 91.3 89.6 93.4 a Sample names are described in Section ‘‘Methods.’’ The fatty acid composition of these samples is 66% stearic acid and 30% palmitic acid, except samples V1, V2, and V3, which are 90% stearic acid and 5% plamitic acid. b Measured hydrate fractions were corrected for the mass of free water by dividing by one minus the fractional loss on drying due to unassociated water. c When the fatty acid composition of the samples is 90% stearic acid and 5% palmitic acid, the theoretical loss on drying of pure monohydrate and dihydrate samples is 2.97% and 5.77%, respectively. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 DOI 10.1002/jps NEAR INFRARED SPECTROSCOPY OF MAGNESIUM STEARATE 2761 Table 3. Predicted Mg Stearate Hydrate Compositions of Validation Samples Based on NIR PLS Model Monohydrate Predictionsb Sample Namea V1 V1 V1 V1 V2 V2 V2 V2 V3 V3 V3 V3 V4 V5 V6 V7 V8 V9 V10 V11 V12 V13 V14 RMSEC RMSECV RMSEP RMSEP Dihydrate Predictionsb Replicate None Derivative MSC SNV None 1 2 3 4 1 2 3 4 1 2 3 4 1 1 1 1 1 1 1 1 1 1 1 90% 90% 90% 90% 68% 68% 68% 67% 62% 62% 62% 61% 89% 89% 90% 93% 91% 85% 86% 87% 9% 5% 4% 1% 1% 9.6% 3.2% 89% 89% 89% 89% 63% 63% 64% 65% 58% 58% 58% 59% 89% 88% 91% 90% 89% 85% 86% 85% 10% 9% 8% 1% 1% 8.6% 3.7% 86% 86% 86% 86% 61% 61% 61% 60% 56% 56% 56% 55% 92% 91% 92% 92% 92% 89% 90% 90% 4% 2% 2% 0.36% 0.44% 7.1% 3.8% 85% 86% 86% 86% 65% 65% 65% 66% 60% 60% 60% 60% 92% 92% 92% 95% 94% 98% 99% 100% 4% 1% 2% 0.42% 1.16% 9.0% 4.1% 2% 2% 2% 3% 25% 25% 25% 26% 31% 31% 31% 32% 3% 4% 2% 1% 1% 7% 7% 5% 87% 90% 91% 0.85% 1.18% 5.7% 3.2% TGA LOD TGA PLS Derivative 4% 4% 4% 3% 30% 30% 30% 28% 35% 35% 35% 34% 3% 4% 1% 2% 3% 7% 6% 7% 85% 86% 87% 0.82% 1.30% 4.9% 3.8% MSC SNV 6% 6% 6% 6% 32% 32% 32% 33% 37% 37% 37% 38% 0% 1% 0% 0% 0% 3% 2% 2% 92% 93% 93% 0.38% 0.46% 2.9% 4.0% 7% 7% 7% 6% 28% 28% 28% 27% 34% 34% 33% 33% 0% 0% 1% 3% 2% 7% 7% 8% 92% 94% 94% 0.86% 1.16% 4.3% 2.3% a Sample names are described in Section ‘‘Methods.’’ Predictions are from PLS models of NIR spectra following different preprocessing procedures. The same procedure is applied to MgSt-M and MgSt-D for each model. b unassociated water over the 25–508C range, whereas its water of crystallization is eliminated over the 90–1108C range. MgSt-D loses water over the 70–908C range, and the TGA thermogram shows no evidence for loss of unassociated water. These observations are consistent with the recent work of Sharpe, et al.12 Onset of the elimination of water from the dihydrate of coprecipitated validation samples is shifted by about 58C relative to the mixture standard thermograms, whereas the MgSt-M water loss from the validation samples is nearly identical to that of the standard mixtures. Particle size differences between the validation samples and standard mixtures may be responsible for this effect. Analysis of individual TGA thermograms was conducted by attributing mass loss on drying (LOD) to either the monohydrate or dihydrate depending upon the temperature at which the loss occurred. The inflection DOI 10.1002/jps point between the features due to the two hydrates (as determined by the first derivative of the TGA curve) was used to divide the monohydrate and dihydrate contributions to the loss on drying. The known compositions determined from the masses of stock materials added to each standard mixture (Prepared Composition) and the compositions determined from LOD measurements of individual TGA thermograms (TGA LOD) are presented in Table 1. One of the central issues in developing predictive chemometric models is the determination of reference values for the training set standards and validation samples. The reference values are necessary for the development of accurate calibration models and the evaluation of the accuracy of the predictive chemometric model. Table 1 indicates that the hydrate compositions determined directly from TGA LOD are not accurate. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 2762 KAUFFMAN ET AL. Figure 1. TGA thermograms of standard mixtures of MgSt-M and MgSt-D of known composition. Nominally pure components exhibit sigmoidal curves; whereas, mixtures exhibit curves that appear as the sum of two sigmoids. An offset of 2% has been added to the thermograms of the validation samples for clarity. TGA LOD compositions tend to underestimate the amount of the major component and overestimate the amount of the minor component in the mixtures. Comparison of TGA LOD to ‘‘Prepared Compositions’’ results in a root mean square error of prediction (RMSEP) of 5% and 3% for MgSt-M and MgSt-D, respectively, for the TGA LOD method, and the maximum errors observed are 9% and 5% for MgSt-M and MgSt-D, respectively. Therefore a method for correcting the TGA values must be developed in order to derive accurate reference values for the validation samples. TGA thermograms of the standard mixtures were calibrated to their known hydrate compositions using partial least squares (PLS) regression. No preprocessing was applied, and the data was not mean centered prior to analysis. This treatment resulted in the most accurate predictions. The optimum model included four factors, and the root mean squared error of calibration (RMSEC) of the four-factor model was 0.2% for both MgSt-M and MgSt-D. Hydrate compositions of standard mixtures predicted by leave-one-out cross validation (LOOCV) are given in Table 1, and the RMS error of cross validation (RMSECV) was 0.5% for both MgSt-M and MgSt-D. The largest residual errors in the LOOCV are approximately 0.7%, JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 which are observed at the extreme values, where the LOOCV procedure must extrapolate to compute the hydrate compositions. All measures of the error between the ‘‘Prepared Compositions’’ and the calibrated compositions are reduced substantially when the TGA PLS model is applied to the TGA thermograms of the standard mixtures. The TGA PLS model was applied to the TGA thermograms of the validation samples in order to assign reference values to the hydrate compositions of these samples, and the results are displayed in Table 2. Also shown in Table 2 are the hydrate compositions determined from TGA LOD calculations of individual TGA thermograms of the validation samples. Sample V1 is particularly intriguing, because the LOD measurement indicates that it is 99% MgSt-M, but the TGA PLS model predicts that it contains nearly 7% dihydrate. Figure 2 shows the DSC thermogram for sample V1. The small peak at 1208C has been assigned to the melting transition of the phase formed by dehydration of MgSt-D, whereas the large peak at 1308C is the melting transition of the phase formed by dehydration of MgSt-M. The inset in Figure 2 shows a comparison of the TGA thermograms of the 0% dihydrate validation DOI 10.1002/jps NEAR INFRARED SPECTROSCOPY OF MAGNESIUM STEARATE 2763 Figure 2. DSC thermogram of the 0% dihydrate validation sample. Inset: Comparison of the TGA thermograms of the 0% dihydrate validation sample and the 100% monohydrate standard mixture. sample (sample V1) and the 100% monohydrate standard mixture sample (sample S1), and highlights subtle differences between the thermograms including earlier onset of mass loss, shallower slope and larger total mass loss in the thermogram of sample V1. The DSC mass loss feature at 958C is also broader than that observed in pure MgSt-M, and is shifted to lower temperature, indicating that mass loss occurs from both MgSt-M and MgSt-D in this sample. These observations are consistent with the TGA PLS prediction that validation sample V1 contains MgSt-D, indicating that chemometric modeling of TGA thermograms can increase the sensitivity of this method to the presence of small amounts of the minor hydrate. Figure 3 displays NIR spectra of the complete set of standard mixtures. The data have been preprocessed by multiplicative scatter correction (MSC),13–15 which suppresses baseline variation. The NIR spectrum of MgSt-M has a sharp feature at 7045 cm1, and a broader feature at 5185 cm1. As the MgSt-D fraction in the mixture increases, the 7045 cm1 feature diminishes in intensity, and the 5185 cm1 feature shifts toward 5100 cm1, which is the peak of a broad MgSt-D feature. These are the regions where the mixture spectra show the strongest systematic variation with increasing MgSt-D composition. The broad DOI 10.1002/jps features in the 8200 cm1 range and the 6000 cm1 to 5300 cm1 range have been attributed to the glass vials that contain the sample. Principal component analysis (PCA) was performed on the entire set of 40 calibration mixture NIR spectra before and after preprocessing, and scores on the first two principal components are plotted against one another in Figure 4. Before preprocessing (Fig. 4A), the scores of three of the four spectra fall within a relatively narrow range of one another for each of the mixtures, however the scores for spectra measured after shaking and tapping are displaced along the Factor 1 axis. This indicates that density changes due to shaking and tapping influence the NIR spectra in a systematic manner. Figure 4B displays the scores for the identical set of spectra after MSC. Scores for all four spectra within each mixture category now reside in a very narrow region of the scores plot. These results indicate that the spectral variation observed within each composition category can be suppressed with MSC, which primarily influences the baseline of NIR spectra, and was designed to suppress features associated with particle scattering. Systematic spectral variation associated with hydrate composition is preserved after MSC preprocessing. Standard normal variate (SNV) scaling of the mixture spectra was found JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 2764 KAUFFMAN ET AL. Figure 3. NIR spectra of standard mixtures used for calibration model development. to be equally effective at suppressing unwanted spectral variation. Partial least squares (PLS) regression was used to calibrate NIR spectral variations to the pseudopolymorphic content of the MgSt hydrate standard mixtures. The hydrate composition reference values used to develop the NIR PLS model are the ‘‘Prepared Composition’’ values given in Table 1. The procedure was applied to both the full 9000–4000 cm1 spectral range and to a subset of the full spectral range including only the 7500–5900 cm1 and 5700–4400 cm1 ranges. Both models gave similar results, indicating that the method presented below is robust with respect to spectral range. The average RMSEP of the model based on the spectral subset was 0.7% lower than that of the full spectrum model, and for simplicity, only the results of the spectral subset model will be examine here. Four preprocessing procedures were considered including no preprocessing, first derivative computation, multiplicative scatter correction and standard normal variate scaling, and each of these was followed by mean centering of the data set. In all cases, linear models with correlation coefficients of 0.99 or greater were observed. With no preprocessing, eight factors were required to achieve the model, and with first derivative computation, three factors were required. Models utilizing MSC or SNV preprocessing required five factors. Root JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 mean squared errors of calibration (RMSEC) and leave-one-out cross validation (RMSECV) were found to be similar for all preprocessing procedures (see Tab. 3). Multiplicative scatter correction results in slightly better RMSEC and RMSECV than the other methods, but all preprocessing procedures appear to give adequate standard errors of calibration and cross validation. The NIR PLS models were then applied to the prediction of the monohydrate and dihydrate compositions of 14 validation samples, and the results are reported in Table 3. Prediction of less than 0% are observed for some of the dihydrate compositions, and are shown in Table 3 to convey the uncertainties in the predictions. However, since these compositions are physically forbidden, all negative compositions were set equal to zero for the purpose of computing RMSEP values. The predicted values of MgSt-M and MgSt-D composition of the validation samples are similar across all preprocessing procedures, as displayed in Table 3. The average standard deviation across preprocessing treatments over all samples is 3%. Two RMSEPs are shown in Table 3 for each preprocessing procedure, one based on TGA LOD reference values and the other based on the TGA PLS reference values. Table 3 indicates that the RMSEP is reduced significantly for all NIR PLS models when the TGA PLS reference values are used, and the overall RMSEP for both hydrates DOI 10.1002/jps NEAR INFRARED SPECTROSCOPY OF MAGNESIUM STEARATE 2765 Figure 4. Scores plots following principal component analysis (PCA) of the NIR spectra of standard mixtures shown in Figure 3. (A) Scores plot after PCA of raw spectra. (B) Scores plot after PCA of spectra following multiplicative scatter correction (MSC). Baseline suppression due to MSC significantly reduces variation between spectra of the same sample due to particle packing and instrumental effects. Similar reduction in spectral variation is observed after standard normal variate (SNV) scaling. The data shown in (B) appear as a vertical line when the x-axis scale is identical to that shown in (A). within each model is reduced to about ½ of the RMSEP using the TGA LOD reference values. The hydrate compositions of the validation samples predicted by the NIR PLS models are DOI 10.1002/jps also similar across NIR replicates when replicates were measured. Replicate 4 of each of the first three validation samples (V1, V2, and V3) was shaken prior to NIR measurement, and resulted JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 2766 KAUFFMAN ET AL. in the spectral variations characterized in Figure 4A. The predicted composition of each validation sample based on replicate #4 deviates from the other replicates slightly when no preprocessing is applied, but this variation is suppressed when any of the three preprocessing procedures is applied. The observed deviation in replicate #4 was greater for models based on the entire spectrum, but preprocessing also suppressed the effect in the full spectrum models. These observations indicate that both spectral selection and preprocessing are able to reduce the influence of unwanted spectral variation on the NIR PLS model of MgSt hydrate composition. Thus PLS calibration of NIR spectra to MgSt-M and MgSt-D compositions has been found to be an excellent method for the rapid, nondestructive analysis of MgSt hydrates. Application of this method to validation samples indicates that the predicted MgSt hydrate compositions are within 3% of the true values. CONCLUDING REMARKS Though the work presented here succeeded in developing an accurate NIR calibration for MgSt hydrate composition, several issues that are implicitly encountered in every multivariate calibration project have been brought to the surface. The first issue is the need for accurate reference values for standards and validation samples. This issue is particularly challenging when developing models for hydrates, because the reference methods are not very sensitive. Karl Fischer titration is capable of determining total water, but cannot assign the water to different crystal hydrates. DSC is capable of establishing the presence of different hydrates under favorable experimental circumstances, but cannot provide quantitative information without calibration. TGA LOD is one of the few methods that is capable of quantifying hydrate composition with adequate selectivity and sensitivity, and is the best method available for determination of reference values for MgSt hydrates. Nevertheless, there are several potential sources of error that contribute to the uncertainty of the reference values. The analysis of hydrate composition in MgSt makes the assumption that all molecules in the sample are in crystalline form. However, the MgSt crystal lattice is very soft, and it is likely that some of the material exists as noncrystalline JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008 ‘‘impurities’’ within the crystalline sample. This may be especially important in samples composed of fatty acid mixtures, where 4% of the sample mass is composed of molecules distinct from the two major molecular species in the sample. Powder x-ray diffraction is capable of distinguishing hydrates as well as dehydrated crystalline material, but lack of sensitivity has prevented us from identifying the form of the dehydrated material for the samples used in this study. Therefore, it is not surprising that mass closure in never achieved for any measure of MgSt hydrate composition in this work. Of course, it is also possible that the unaccounted mass is due to amorphous MgSt, but in the absence of evidence to the contrary, the only viable assumption in the present analysis is that all molecules are part of a crystal lattice. The second issue that requires attention is the fact that though the TGA PLS model appears to accurately correct the standard mixture compositions, the reference values for this calibration are also dependent on the TGA LOD of the MgSt-M and MgSt-D stock materials, which are only measured to be 92% and 95.4% hydrated, respectively. Without pure MgSt-M and MgSt-D standard samples, we have no way to assess the accuracy of the predictions that are outside of the range of the reference samples. Nevertheless, our goal is to develop a rapid, nondestructive method for determination of pseudopolymorphic content in pharmaceutical grade magnesium stearates, and for this purpose the standard materials used in this study are the best materials that are currently available. Finally, the most intriguing question raised by this study is based on the suspicion that NIR spectroscopy offers a more sensitive method (and therefore a potentially more accurate method) for analysis of MgSt hydrates than the TGA LOD reference method. This suspicion is corroborated by the DSC profile of sample V1, which verifies the existence of MgSt-D in a sample that was found to contain none by TGA LOD. Interestingly, the NIR PLS model based on reference values determined from TGA LOD also predicted a finite MgSt-D composition in sample V1, a fact which motivated the development of the TGA PLS model. However, it is not possible to establish the accuracy of the NIR PLS model independent of the reference method, and therefore NIR spectroscopy will never be considered more accurate than TGA LOD until a more accurate primary reference method is found. DOI 10.1002/jps NEAR INFRARED SPECTROSCOPY OF MAGNESIUM STEARATE REFERENCES 1. Wada Y, Matsumura T. 1994. Pseudopolymorphism and lubricating properties of magnesium stearate. Powder Tech 78:109–114. 2. Rao KP, Chawla G, Kaushal AM, Bansal AK. 2005. Impact of solid-state properties on lubrication efficacy of magnesium stearate. Pharm Dev Tech 10:423–437. 3. Ahmed M, Enever RP. 1976. Influence of magnesium stearate on the dissolution and biological availability of sulphadiazine tablet formulations. J Pharm Pharmacol 28:5. 4. Durig T, Fassihi R. 1997. Mechanistic evaluation of binary effects of magnesium stearate and talc as dissolution retardants at 85% drug loading in an experimental extended-release formulation. J Pharm Sci 86:1092–1098. 5. Jannin V, Berard V, N’Diaye A, Andres C, Pourcelot Y. 2003. Comparative study of the lubricant performance of Compritol(registered trademark) 888 ATO either used by blending or by hot melt coating. Int J Pharm 262:39–45. 6. Sheikh-Salem M, Alkaysi H, Fell JT. 1988. The tensile strength of tablets of binary mixtures lubricated with magnesium stearate. Drug Dev Ind Pharm 14:895–903. 7. Sheikh SM, Fell JT. 1981. The influence of magnesium stearate on the dependent strength changes in tablets. Drug Dev Ind Pharm 7:669–674. 8. Wurster DE, Likitlersuang S, Chen Y. 2005. The influence of magnesium stearate on the Hiestand Tableting Indices and other related mechanical DOI 10.1002/jps 9. 10. 11. 12. 13. 14. 15. 2767 properties of maltodextrins. Pharm Dev Tech 10: 461–466. Zuurman K, Van D V, Bolhuis GK. 1999. Effect of magnesium stearate on bonding and porosity expansion of tablets produced from materials with different consolidation properties. Int J Pharm 179: 107–115. Ertel KD, Carstensen JT. 1988. Chemical, physical, and lubricant properties of magnesium stearate. J Pharm Sci 77:625–629. Swaminathan V, Kildsig DO. 2001. An examination of the moisture sorption characteristics of commercial magnesium stearate. AAPS PharmSciTech 2:28. Stefan AS, Metin C, Ann WN, Harry GB. 1997. Physical characterization of the polymorphic variations of magnesium stearate and magnesium palmitate hydrate species. Struct Chem 8:73– 84. Gelaldi P, MacDougal Dl, Martens H. 1985. Linearization and scatter-correction for near-infrared reflectance spectra of meat. Appl Spectrosc 39:491– 500. Martens H, Nielsen JP, Engelsen SB. 2003. Light scattering and light absorbance separated by extended multiplicative signal correction. Application to near-infrared transmission analysis of powder mixtures. Anal Chem 75:394– 404. Isaksson T, Naes T. 1988. The effect of multiplicative scatter correction (MSC) and linearity improvement in NIR spectroscopy. Appl Spectrosc 42:1273– 1284. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 97, NO. 7, JULY 2008