A comparison of near infrared method development approaches using a drug product on different spectrophotometers and chemometric software algorithms

Transcription

1 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) Received: 9 June 2009 Revised: 30 September 2009 Accepted: 11 October 2009 Publication: JOURNAL OF NEAR INFRARED SPECTROSCOPY A comparison of near infrared method development approaches using a drug product on different spectrophotometers and chemometric software algorithms Assad Kazeminy, a Saeed Hashemi, a Roger L. Williams, b Gary E. Ritchie, c Ronald Rubinovitz d and Sumit Sen e,* a Irvine Pharmaceutical Services, Inc., 10 Vanderbilt, Irvine, CA 92618, USA b United States Pharmacopeial Convention, Twinbrook Parkway, Rockville, Maryland 20852, USA c Former United States Pharmacopeial Convention, Twinbrook Parkway, Rockville, Maryland 20852, USA d Büchi Corporation, 19 Lukens Drive, New Castle, DE 19720, USA e United States Food and Drug Administration, Fairchild, Irvine, CA 92612, USA. sumit.sen@hotmail.com It is well known that spectral variability in near infrared (NIR) spectroscopy can be attributed to the analyst, sample, sample positioning, instrument configuration and software (in both algorithm formats and structures used as well as in the execution of data pre-treatment and analysis). It is often acknowledged that the single largest factor impacting NIR results is sample presentation. However, what is obvious but not often acknowledged is that there are instrumental and software differences as well. These differences, evident in the quality of the spectra, may impact the chemometrics that are subsequently performed and, possibly, the results obtained from the multivariate statistical models. In order to investigate just what are these sources of variability, and just how much these variations may impact the results of the multivariate models for predicting the identification of pharmaceutical dosage forms, a study has been conducted. To the authors knowledge, no other systematic study of this kind has been published. In this study, we are interested in learning what variability, if any, the choices for instrument and software have on the development of a NIR method for the identification of pharmaceutical dosage forms. Furthermore, we would like to learn what and how do the choices made early on in the experimental design impact the final quality of the spectra and the resulting multivariate models obtained from these spectra. A study protocol was designed, using a common data set consisting of four formulations of Ibuprofen, involving three investigating parties, namely, US FDA, USP and Irvine Pharmaceutical Services and using three NIR instruments, namely (listed in alphabetical order), a Bruker spectrometer, a Büchi spectrometer and a Foss spectrometer. Based on the results and despite differences in instrument configuration [dispersive or Fourier Transform (FT)], number of spectral data points, principal components analysis (PCA) or factorisation algorithms, and validation modelling approach, exact and accurate spectroscopic results can be achieved using NIR spectroscopy for discriminate analysis. More importantly, this study shows that the same NIR method spectral range and pre-treatment parameters can be used, and that nearly the same multivariate models can be obtained, despite instrumental and software differences, to accurately predict the identity of pharmaceutical dosage forms. Keywords: near infrared (NIR) spectroscopy, instrument variability, chemometric software algorithms, multivariate discriminant analysis, PCA ISSN: IM Publications LLP 2009 doi: /jnirs.xxx All rights reserved

2 2 An NIR Comparison of Method Development Approaches Using A Drug Product Introduction It has been successfully demonstrated that near infrared (NIR) spectroscopy can be used for the analysis of pharmaceutical products using a single instrument and software configuration. 1 6 The focus of this study was to determine if the variability observed across different NIR instruments and chemometric software packages can be an impediment for the deployment of this technique if different instrument and software types are to be used for the development of NIR models. The United States Food and Drug Administration (US FDA), the United States Pharmacopeia (USP) and Irvine Pharmaceutical Services, have undertaken a feasibility study using NIR spectroscopy to identify, by the chemometric method of multivariate discriminate analysis using principal component analysis (PCA) and factorisation, four formulations of Ibuprofen (200 mg): two branded (Advil and Motrin) and two store-branded (CVS and Rite Aid) Ibuprofen (200 mg) immediate-release tablets, using three different instrument/ software combinations. The instruments used included (listed in alphabetical order), a Bruker Vector 22N FT-NIR spectrometer, Bruker Optik GmbH, Ettlingen, Germany, a Büchi NIRFlex Solids, Flawil, Switzerland and a Foss XDS Rapid Content Analyser, Laurel, USA. The software used from each instrument, respectively, were OPUS 5.5, NIRCal 5.2 and Vision 3.4. The Unscrambler 9.7, a stand-alone multivariate analysis and experimental design software package, was used as referee software to assist in developing a common model. The Bruker Vector 22N FTNIR spectrometer and the Foss XDS Rapid Content Analyser used for this study are both located at the USP, and the Büchi NIRFlex Solids spectrometer was located at the Irvine Pharmaceutical Services. A study protocol was designed, using a common data set consisting of these four formulations of Ibuprofen, involving the three investigating parties, using the above-mentioned different NIR instrument/software combinations. To our knowledge, this is the first study of its kind investigating the variability of NIR instrument/software combinations and their impact on chemometric models for discriminate drug analysis. There were four specific objectives of the study: (1) Demonstrate that by the use of the computerised algorithms of PCA and factorisation, three identical, or nearly identical, spectral libraries can be obtained on three different instruments. These spectral libraries will be comprised of the same samples drawn from the same manufactured lots, scanned the same way and modelled using the same software settings. (2) Compare four formulations of Ibuprofen: (200 mg) two branded and two store-branded Ibuprofen (200 mg). Immediate-release tablets will result in four distinct clusters in multi-dimensional space. (3) Demonstrate that the calibration models are able to analyse unknown tablets (validation set) and determine if they are ibuprofen (200 mg) immediate-release tablets from one of the four known sources of origin that constitute the libraries or are not from any of these known sources. (4) The results from each analysis of unknown samples should give the same predicted results, indifferent of the model used, and despite instrument, spectra, algorithm, or number of calibration samples, demonstrating that a protocol can be used in each laboratory, independent of location, device, analyst and software used to perform the identification. Experimental Instrumentation and software The instruments that were used included a Bruker Vector 22N FT-NIR spectrometer with a spectral range of cm 1 and spectral resolution set at 8 cm 1 (Bruker Optik GmbH, Ettlingen, Germany), Büchi NIRFlex Solids with a spectral range nm (10, cm 1 ) and spectral resolution set at 8 cm 1 (Büchi, Flawil, Switzerland) and Foss XDS Rapid Content Analyser with a spectral range of nm (25, cm 1 ) and a spectral resolution set by the specified bandpass of 9 nm (Foss NIRSystems, Inc., Laurel, MD, USA). The software used for each instrument were, respectively, OPUS 5.5, NIRCal 5.2, and Vision 3.4. The Unscrambler 9.7, a standalone multivariate analysis and experimental design software package (CAMO Technologies Inc., Trondheim Norway), was used as referee software to assist in developing a common model. Validation of instrument operation Tests for wavelength accuracy, photometric precision, accuracy and noise were performed on all of the instruments prior to tablet measurements as recommended by the USP general information chapter <1119> Near-Infrared Spectroscopy ( 7 (Graeme has asked me to reference this as Reference 7 and renumber the remaining references. Is this okay?) Vendor recommended performance tests for each instrument were run as well. The tests are routine quality control tests of the instrument performance and as such were performed at the prescribed intervals to verify correct instrument performance. Sample sets Table 1 lists the sources for the samples used for the study. Ten lots of four formulations of Ibuprofen (200 mg) were obtained: two branded (Advil, Motrin) and two store-branded (CVS, Rite Aid) Ibuprofen (200 mg) immediate release tablets. The formulations for each are shown in Table 2. Three sets of samples were prepared for each instrument. These sets comprise the calibration set, test set and validation set. For this study, manufacturer lots 1 8 were used for the calibration set and test set and manufacturer lots 9 and 10 were set aside and used for the validation sets. While one of the purposes of the protocol was

3 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) 3 Table 1. Sources of Ibuprofen (200 mg) samples used for this study. Lot Brand Advil Brand Motrin Generic CVS Generic Rite Aid 1 B PCA189 6EE0102 P = 400 tablets = 300 tablets = 500 tablets = 500 tablets 2 B87154 LLA103 6HE0515 P = 400 tablets = 300 tablets = 400 tablets = 500 tablets 3 B94669 PCA112 7BE0119 P = 300 tablets = 300 tablets = 750 tablets = 250 tablets 4 B27624 PCA226 7AE0039 P = 300 tablets = 240 tablets = 750 tablets = 250 tablets 5 B91364 PBA123 7CE0268 P = = 300 tablets = 750 tablets = 300 tablets 6 B73322 PBA194 7AE0699 P = 300 tablets = 300 tablets = 500 tablets = 300 tablets 7 B33863 PAA016 6LE0478 P = 300 tablets 5 50 = 250 tablets = 500 tablets = 240 tablets 8 B91414 PBA186 7AE0270 P = 400 tablets 5 50 = 250 tablets = 300 tablets 5 50 = 250 tablets 9 B98483 PEA106 6GE0118 P = 216 tablets = 200 tablets = 500 tablets 5 50 = 250 tablets 10 B91386 LLA329 7BE0606 P = 400 tablets = 200 tablets = 500 tablets = 200 tablets to control the preparation of the calibration, test, and validation sample sets, it was found later that the calibration set prepared for the Foss instrument was missing one tablet (see Table 6 footnote) and the validation set for the Foss instrument and the Bruker instrument were different from the validation set prepared for the Büchi instrument, containing only 28 tablets from the Motrin validation set instead of 40 tablets originally intended, all due to packaging errors. This error had no effect on the outcome of the study (model prediction accuracy of validation samples) as will be shown later in this paper. The sample sets were prepared as shown in Table 3. Sample collection and analyses The study protocol was modelled on that described by Scafi and Pasquini. 8 The multivariate chemometric approach to calibrate, validate and maintain the spectral libraries was based on work proposed by Workman and Brown. 9,10 Log 1/R (sometimes referred to as Absorbance) spectra from the Bruker and Foss instrument, and reflectance spectra from the Büchi instrument, were collected from each tablet for the calibration and test sets by averaging 32 scans in diffuse reflectance mode against a 99% reflectance background reference. The spectra were collected across the following ranges from each instrument: Bruker samples were scanned from 11,999 cm 1 to 3999 cm 1 (833 nm to 2500 nm), Büchi samples were scanned from 10,000 cm 1 to 4000 cm 1 (1000 nm to 2500 nm) and using the Foss instrument, each tablet was scanned from 400 nm to 2500 nm, Photographs of the tablets are shown in Figure 1. The Unscrambler 9.7 was used to investigate the Foss calibration set spectral properties that could be used to distinguish each formulation. Tables 4 and 5 show the series of spectral pre-treatments that were tested for use in order to achieve a PCA that could be used to determine the identity of the samples from the validation sets. The test sets were used to optimise the prediction of the final models before being used to predict the validation sets. The Bruker software, OPUS 5.5, uses the factorisation method to model and so the calibration set from this instrument, while subjected to PCA in The Unscrambler 9.7, required a different approach when constructed in the OPUS 5.5 software. Table 6 lists the model values for the calibration modelling approach. Development of NIR calibrations Based on observations from the Unscrambler 9.7 exploration, a Savitzky Golay, first derivative, 21 point smoothing, 3rd order polynomial, from 1000 nm to 2500 nm and a PCA model, using Mahalanobis distance in the wavelength range from nm, was postulated as the optimal model to apply to the Bruker, Büchi and Foss calibration data sets. Figures 2(a), 2(b)

4 4 An NIR Comparison of Method Development Approaches Using A Drug Product Table 2. Formulations of Ibuprofen (200 mg) samples. Advil Motrin CVS-Ibuprofen Rite Aid-Ibuprofen Acelytated monoglycerides Beeswax Propylene glycol Cellulose Lactose Sodium starch glycolate Ethoxyethanol Lecithin Pharmaceutical glaze Povidone Semithicone Sodium benzoate Sodium lauryl sulphate Sucrose Croscarmellose sodium Microcrystalline cellulose Pharmaceutical shellac Pregelatinised starch Carnauba wax Iron oxides Silicon dioxide Corn starch Ibuprofen Stearic acid Titanium dioxide Fd&C yellow #6 Magnesium stearate Polydextrose Polyethylene glycol Shellac Ppregelatinised starch Carnauba wax Hypromellose Iron oxides Colloidal silicon dioxide Corn starch Ibuprofen Stearic acid Titanium dioxide FD&C yellow #6 Magnesium stearate Polydextrose Polyethylene glycol Carnauba wax Hypromellose Fumed silica gel Corn starch Ibuprofen Stearic acid Titanium dioxide Croscarmellose sodium Microcrystalline cellulose Hypromellose Iron oxides Colloidal silicon dioxide Corn starch Ibuprofen Stearic acid Titanium dioxide and 2(c) show that four clusters can easily be separated from each other in all three data sets. The following discussions track the model development in the OPUS 5.5, NIRCal 5.2 and Vision 3.4 software packages. Pre-treatment I: baseline correction Based upon the exploration of data pre-treatments shown in Tables 4 and 5, a baseline correction using the function: f(x) = x min(x), where x is a variable and X denotes all selected variables for this sample, was first applied in the Unscrambler 9.7. A baseline correction was shown to be the simplest pre-treatment applied to the calibration data sets to effect separation of all four formulations, although barely. Furthermore, it was determined that while it is possible to apply a baseline transform function to the Foss calibration set using the Vision 3.4 software, the same mathematical transformation was not found to be available in the Bruker or the Büchi software. Using the Bruker OPUS 5.5 software, the closest approximation to the baseline offset function used in the Unscrambler 9.7 would be a normalisation function. This would then be followed by the Savitzky Golay, first derivative, 21 point smoothing, 3rd order polynomial, from 1000 nm to 2500 nm, followed by factorisation (which is a Bruker PCA algorithm using Euclidian distance within the Opus IDENT set-up module where the calibration set resides. It was further noted that the default first derivative transformation for the OPUS 5.5 software is a cubic polynomial equation. The Büchi NIRCal 5.2 software required the use of several spectra manipulations; to reverse the x-axis, orient the spectra in the same direction as they are for the Foss and Bruker data, prior to performing the PCA to the data [Figures 3(a), 3(b) and 3(c)]. The Savitzky Golay first derivative transformation for the NIRCal 5.2 software is also a cubic polynomial equation. In addition to other derivative functions, the NIRCal software allows the use of customisable linear filters, which allowed matching to the same derivative treatment used in the other

5 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) 5 Table 3. Study design. Laboratory/instrument Experimental Innovator Innovator Generic Generic Total samples manufacturer Advil manufacturer Motrin manufacturer CVS manufacturer Rite Aid United States Calibration Set 6 tabs 8 lots = 48 6 tabs 8 lots = 48 6 tabs 8 lots = 48 6 tabs 8 lots = Pharmacopeia/ Test Set 2 tabs 8 lots = 16 2 tabs 8 lots = 16 2 tabs 8 lots = 16 2 tabs 8 lots = Foss Validation set 20 tabs 2 lots = tabs 2 lots = tabs 2 lots = tabs 2 lots = United States Calibration Set 6 tabs 8 lots = 48 6 tabs 8 lots = 48 6 tabs 8 lots = 48 6 tabs 8 lots = Pharmacopeia/ Test Set 2 tabs 8 lots = 16 2 tabs 8 lots = 16 2 tabs 8 lots = 16 2 tabs 8 lots = Bruker Validation set 20 tabs 2 lots = tabs 2 lots = tabs 2 lots = tabs 2 lots = FDA/Irvine Pharmaceutical Calibration Set 6 tabs 8 lots = 48 6 tabs 8 lots = 48 6 tabs 8 lots = 48 6 tabs 8 lots = Services, Inc./ Test Set 2 tabs 8 lots = 16 2 tabs 8 lots = 16 2 tabs 8 lots = 16 2 tabs 8 lots = Büchi Validation set 20 tabs 2 lots = tabs 2 lots = tabs 2 lots = tabs 2 lots = Total software packages. The resulting filters for the Büchi calibration set are listed in Table 6. Pre-treatment II: Savitzky Golay first derivative, 21 point smoothing, cubic polynomial The smoothing and differentiation of paired data by the procedure of simplifi ed least squares, now called the Savitzky Golay Filter, 11 was developed for the removal of the random noise from paired data sets, such as that obtained for NIR, Log 1/R versus wavelength. For the data sets from this study, it can be shown that the calibration by PCA or factorisation is most affected by sampling variation manifested as baseline offsets. The offsets result in significant overlap of spectra with closely resembling spectral features, which cannot be completely resolved in spectral space. Tables 4 and 5 show a systematic approach of pre-treating the calibration sample spectra. The pre- treatments were applied and investigated Figure 1. Ibuprofen (200 mg) tablets. Upper left, Advil, upper right, Motrin, lower left CVS, and lower right, Rite Aid. for their ability to remove sampling variation from the spectra. The plus sign indicates that the samples from these formulations can be separated into distinct clusters from the other samples of a different formulation using the associated pretreatment, whereas the minus sign indicates that samples from these formulations cannot be separated into distinct clusters from the other samples of a different formulation using the associated pre-treatment. The contribution of the variability from the samples can be seen in Figures 3(a), 3(b) and 3(c). The baseline offset is evident in the exploded plot of the spectra in Figure 4 and the effect of smoothing and derivative pre-treatment is also evident as well (Note: The number of data points applied for smoothing the spectra during derivatisation were investigated and 21 data points was found to be optimal for discerning spectral features of the four formulations in spectral space). Particularly noticeable at the wavelength 1440 nm, one can observe a sharp band in the untreated spectra, which is attributed to Innovator A samples (Advil) [Note: This is due to the sucrose in the formulation as indicated in Table 2. The origin of this unique band is discussed by Davies and Miller. 12 The impact of band resolution from dispersive and FT is most noticeable in Figure 4, which is the expanded view of the first derivative and smoothed spectra. The result of applying the Savitzky Golay first derivative, 21 point smoothing filter is that it changes the directions of the of the Generic A and Generic B spectra relative to both the Innovator A and Innovator B spectra in the region between 1400 nm and 1500 nm. Principal component calculations The PCA models [either by Mahalanobis distance (MD) for the Foss and Büchi instruments or Euclidian distance (ED) for the Bruker instrument) resulted in correct and accurate predictions of all three validation sets. Mahalanobis distance is the statistical distance taking into account the variance of each variable and the correlation coefficients. In the case of a single

6 6 An NIR Comparison of Method Development Approaches Using A Drug Product Table 4. Data pretreatments from nm. Spectra treatment Cluster nm Innovator A Innovator B Generic A Generic B Untreated spectra + + Baseline corrected First derivative Second derivative + + Baseline corrected + + First derivative Baseline corrected Second derivative * Pre-treatments applied to the raw spectra that resulted in a separate and distinct cluster for any member of the four sample sets following principle component analysis are denoted by a plus sign, and pre-treatments that did not result in separate and distinct clusters are denoted by a minus sign. Plus signs, in bold, indicates that all four sample sets were successfully separated variable, it is the square of the distance (between two objects, or between an object and the centroid) divided by the variance a Euclidian distance is simply the distance between two variables and is calculated as the arithmetic difference, i.e. variable 1 variable 2. For a bi-variate model, the squared distances between two vectors in multi-dimensional space are the sum of squared differences in their coordinates. As noted previously, the algorithms used by each software were different by the calculations employed for score distances. The significance of this difference is that, even though the distances were calculated in the principal component space (PC space) for the Foss and Büchi calibration sets, and in wavelength space for the Bruker calibration set, on similarly pre-treated spectra, using the same wavelength range, and on different number of principal components, the models gave identical predictions for the validation sets (see Table 7). The reason that the PCA a From a glossary provided with the permission of Bryan Prazen of the Synovec Group, Department of Chemistry, University of Washington. Found at 18&type=EducationFeature&chId=9&page=1 calculation is invariant in either wavelength space or PC space is explained by De Maesschalck et al. 6 The supervised training by PCA, a discriminant analysis technique, relies on the measurement of distances between objects in order to achieve classification. As noted by De Maesschalck et al., this can occur in wavelength space or in PC space, and on normalised or unnormalised spectra. Another factor that De Maesschalck et al. point out, that may not be obvious to a casual user of NIR techniques, is that the MD and the ED can also be calculated using a smaller number of latent variables (PCs) obtained after PCA analysis instead of the original variables. In this case, the MD, however, does not need to correct for the covariance between the variables, since PCs are by definition orthogonal (uncorrelated). However, the way each of the residual PCs is weighted in the computation of the distance must be taken into account. The paper goes on to clarify the relationship between the ED and the MD, particularly how each is calculated in the original variable space and the PC space. It should be noted here that the Bruker software, OPUS 5.5, utilises an approach referred to as factorisation, which is explained as a spectral distance calculation. Table 7 gives the model values for each Table 5. Data pre-treatments from nm. Spectra treatment Cluster ( nm) Innovator A Innovator B Generic A Generic B Untreated spectra + Baseline corrected + + First derivative + + Second derivative + or yes? Baseline corrected First derivative Baseline corrected Second derivative + ** See footnote from Table 4

7 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) 7 Table 6. Model values for each calibration. Unscrambler 9.7 Vision 3.4 OPUS 5.5 NIRCal 5.2 Calibration set * Test set Validation set Raw spectra range nm nm cm cm 1 Number of raw data points Baseline transform ** Savitzky Golay First derivative (SG 1st deriv.) ( nm), using a 3rd order polynomial, 21 point smoothing PCA Baseline offset f(x) = x min(x) ( nm) Math treatments baseline correction (1000 nm) Manipulate offset normalisation after transforming spectra from wavenumber to wavelength ( nm) N/A N/A Change from wavenumber to wavelength Add constant, Constant = 1000 Shift neg to 0, (total 1501/1501) Change from wavenumber to wavelength N/A N/A N/A Absorbance Log 10 (1/ ) Modify /transform/ derivatives/s. Golay/ variables (1000 nm 2500 nm)/ 1 st derivative/ 21 points smoothing/ 3 rd order polynomial Task Mahalanobis Distance in Principal Component Space/ nm Math treatments Savitzky Golay First derivative/ Region minimum/1000 nm/ region maximum/ 2500 nm/21 point, cubic spline polynomial Mahalanobis distance in principal component space/spectral filtering/ wavelength min.1400 nm/ wavelength max nm PCs: 5 Threshold: Probability Level: Evaluate/Setup Identity test method/ Load method/ Parameters/ Preprocessing/First derivative 21 points/ Regions/ cm cm 1, cubic spline polynomial Factorisation: six factors (in wavelength space) Threshold: 0.25/ nm Linear filter (84075, 10032, 43284, 78176, 96947, , 95338, 79564, 56881, 29592, 0, 29592, 56881, 79564, 95338, , 96947, 78176, 43284, 10032, 84075, ) Mahalanobis Distance in Principal Component Space/ nm *Only seven tablets were packaged for sample 6UFGB (lot 6 USP/Foss/Generic B Rite Aid), all other lots from UFGB: 1,2,3,4,5,7, and 8 were packaged to contain eight tablets, six for the Calibration set and two for the test set. **Since the baseline algorithms were all significantly different and do not give sufficiently similar spectral pre-treatments, it was decided to drop this step and proceed directly to the derivative/smoothing step on all calibration data. The attempt to use normalisation to match the baseline correction from the Vision 3.4 and Unscrambler 9.7 software was met with another unexpected obstacle which forced the investigators to not use a baseline correction as the first step of pre-treating the calibration spectra. The normalisation pre-treatment algorithm resides in a different module than the IDENT module where the calibration model resides. In order to carryout the pre-treatment on the calibration and subsequent unknown samples, a macro that would pre-process the spectra and automatically load those spectra into OPUS IDENT as well as being able to analyse an unknown sample using offset normalisation spectral pre-processing and automatically output a result, would be required. Since this is only a requirement of the OPUS 5.5 software, it was decided to drop this step and proceed directly to the derivative/smoothing step on all calibration data. calibration from each instrument and the Camo software. Figure 4 shows the expanded view of the derivative spectra, and Figure 5 are the scores plots from each of the software (Should this phrase be deleted?).

8 8 An NIR Comparison of Method Development Approaches Using A Drug Product Innovator B 2a (a) Generic B Generic A Innovator A Innovator B (b) 2b Generic B Generic A Innovator A Innovator B 2c (c) Generic B Generic A Innovator A Figure 2. (a) Büchi PCA Score Plot (Unscrambler 9.7) using, (b) Foss PCA Score Plot (Unscrambler 9.7) and (c) Bruker PCA Score Plot (Unscrambler 9.7).

9 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) 9 (a) 3 log (1/R) Innovator A (b) Log (1/R) Log (1/R) (c) Wavelength Figure 3. (a) Büchi raw spectra plot (NIRCal 5.2), (b) Foss raw spectra plot (Vision 3.4) and (c) Bruker raw spectra plot (OPUS 5.5).

10 10 An NIR Comparison of Method Development Approaches Using A Drug Product Büchi (b) (c) (a) (d) Foss (b) (a) (c) 18 (d).0020 Bruker Line Plot (b).0005 (c) (a) nm nm nm nm nm nm nm nm nm (d) Variables Figure 4. Expanded view of Büchi, Foss and Bruker derivative spectra of calibration and test set: (a) Innovator A, (b) Generic A, (c) Generic B and (d) Innovator B.

11 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) 11 Table 7. Number of tablets identified. Instrument Lot Innovator A tablets Innovator B tablets Generic A Generic B Foss Bruker Büchi Twenty tablets were tested for each combination of instrument and lot, except for Lot 10 of Innovator B where only eight were tested on the Foss and Bruker instruments Validation results and discussion Table 7 lists a summary of the results of the validations sets from the, Bruker, Büchi and Foss instruments, respectively. The results demonstrate that all four objectives of the study were met. Four distinct computerised algorithms of PCA and factorisation were used to construct three separate spectral libraries from a common calibration and test set, each residing on different instruments in different laboratories and one residing on stand-alone software, the referee model. The use of the referee model helped establish the baseline values for the model parameters such as spectral range, spectra pre-treatment, calibration algorithms, etc. that could be used across the various software platforms despite the variations in instrument bandwidth, spectral data points, algorithms for smoothing, derivative and other calculations. Specifi cally, while it was found that a PCA model based on calculating Mahalanobis distance in PC space, second derivative Gap 20, second order polynomial was sufficient to model the calibration and test set on two of the three instruments, it became evident early on in the experiment that by using exactly the same model consisting of the same algorithm calculation and same pre-treatment routine was not possible. However, by the use of the referee software which, when loaded with the calibration set from all three instruments, one could easily determine how to optimise the spectral range, and smoothing and derivative pre-treatment parameters in order to achieve similar calibration and test set parameters. The limiting calibration setting was found to be the polynomial function, which effects how many points are used to calculate the smoothing and derivative function. As was shown in Figure 4, there is a very narrow spectral range (100 nm or cm 1 on which separation can be made on the four data sets in spectra space. Having met the first objective with some deviations, the second, third and fourth objectives were easily met, as the three models were successful at predicting the validation sets for each instrument, resulting in four distinct clusters in multidimensional space, each cluster representing the innovator and generic brand Ibuprofen formulations. Additionally, samples from Generic A lots 6GE0118 (CVS Brand) and P42058 (Rite Aid) were correctly identified as not belonging to Generic A lots comprising the calibration sets for all three data sets. It was observed that these samples have a distinct banding pattern in the region from 1400 nm 1500 nm from all other sample spectra. The most likely cause of this is assuredly due to the presence of a component that absorbs in the NIR region not found in those samples comprising the calibration set. As a result of the new band in the critical region between nm, these samples also produce a fi fth separate and distinct cluster in PC space. It must be mentioned here that this experiment, being performed in different laboratories within different organisational cultures, was totally driven by a protocol that was jointly crafted and agreed upon prior to execution. The authors feel that this is a key point since this experiment was designed to meet specific objectives, despite the fact that the instruments, software and personnel were at different locations. This, of course, was not the major factor. The major factor was trying to coordinate all of the steps within the protocol from within different organisations. While NIR experiments are generally described as non-destructive, fast and cost effective, when done on a large scale they require planning, discussion and coordination. This is rarely mentioned. It is hoped that this example may serve as a model for future applications that involve large sample sets and multi-organisations using multiple instrument software combinations. The current global pharmaceutical counterfeiting problem is one area that should benefit from examples like the one demonstrated in this paper. Conclusions The specific objective of the study was to obtain log 1/R spectra of four formulations of Ibuprofen (200 mg) from two branded

12 12 An NIR Comparison of Method Development Approaches Using A Drug Product and two store-branded Ibuprofen (200 mg) immediate-release tablets and use them to design, develop, validate and deploy a calibration model that can subsequently be used to correctly classify by discriminant analysis using PCA, log 1/R spectra from unknown samples (validation set) on NIR instruments of varying types and software configurations. This experiment was designed to study the impact that NIR instrument hardware and software configurations have on NIR method development. Several variables were detected and assessed. Spectrometer types, sample holders, spectral acquisition settings, data pre-treatments and PCA algorithms were studied. NIR method development was attempted by three different analysts on three different instruments located in two different laboratories. It was determined that even though identical samples were used for modelling and prediction, and the same calibration approach was tried on the accompanying software, spectra differences were observed due to the number of data points, and that these impact the ability to perform the same or similar data pretreatments in different software. Furthermore, the algorithms employed in each software platform may limit the ability to deploy a method developed on any single software platform to be deployed across different software platforms. However, despite the differences observed, it was possible to find a common method using each software that enabled accurate predictions of the validation samples when each model was used independent of instrument and software configuration. Knowing the sources of variability that impact the log 1/R NIR spectrum will minimise the overall prediction variability and increase the likelihood of correctly classifying by discriminant analysis, the log 1/R spectra from unknown samples subsequently measured and compared to the spectral library and classifi ed by the calibration model, when model parameters are used on different instrument and software combinations. It was found that using Savitzky Golay, first derivative, 21 point smoothing, third order polynomial, pre-treated spectra and either PCA or factorisation model (either by MD in principal component space for the Foss and Büchi or ED in wavelength space for the Bruker Factorisation method) resulted in different models but possessing the same accuracy capabilities (100%) for predicting samples comprising similar validation sets. Each model consisted of 192 calibration samples (except the Foss calibration set which had one missing tablet for a total of 191 calibration samples) and 64 test set samples developed for each NIR instrument, Each model correctly and accurately (100%) predicted 160 validation samples using the Büchi model, and 148 validation samples using the Bruker and Foss models (Graeme Batten thinks that this sentence should be deleted. Do you agree?). One validation sample set, a store-branded Ibuprofen (200 mg) immediate-release tablet, was correctly identified as not belonging to the samples represented in the calibration set by all three models. Based on these results, and despite differences in instrument configuration [dispersive or Fourier Transform (FT)], number of spectral data points, PCA or factorisation algorithms and validation modelling approach, exact and accurate spectroscopic results can be achieved using NIR spectroscopy for discriminate analysis. More importantly, this study shows that the same NIR method spectral range and pre-treatment parameters can be used and that nearly the same multivariate models can be obtained, despite instrumental and software differences, to accurately predict the identity of pharmaceutical dosage forms. Acknowledgements The authors are grateful for the contributions to this project from each of the following scientists: Darrell Abernathy, Rebecca Allen, Todd Cecil, Walter Hauck, Andrea Iwanik, Steven Lane, Samir Wahab and Patricia White from USP, Rudy Flach, Charles Petersen and Heather Coffin from Irvine, William Martin from the FDA, Michael Surgeary from Büchi, Cynthia Kradjel from Integrated Technical Solutions and Verne Hebard from Bruker. References 1. P. de Peinder, M.J. Vredenbregt, T. Visser and D. de Kaste, Detection of lipitor counterfeits: a comparison of NIR and Raman spectroscopy in combination with chemometrics, J. Pharmaceut. Biomed. Anal. 47, 688 (2008). doi: /j.jpba J. Luypaerta, D.L. Massart and Y. Vander Heyden, Near-infrared spectroscopy applications in pharmaceutical analysis, Talanta 72, 865 (2007). doi: /j. talanta Y. Roggo, P. Chalusa, L. Maurera, C. Lema-Martineza, A. Edmonda and N. Jenta, A review of near infrared spectroscopy and chemometrics in pharmaceutical technologies, J. Pharmaceut. Biomed. Anal. 44, 683 (2007). doi: /j.jpba A.K. Deisingh, Pharmaceutical counterfeiting, Analyst 130, 271 (2005). doi: /b407759h 5. C.A. Anderson, J.K. Drennen and E.W. Ciurczak, Pharmaceutical applications of near infrared spectroscopy, in Handbook of Near-Infrared Analysis, 3rd Edn, (Practical Spectroscopy Series Volume 35) Ed by D.A. Burns and E. W. Ciurczak. CRC Press, Boca Raton, Florida, USA, p. 585 (2008). 6. R. De Maesschalck, D. Jouan-Rimbaud and D.L. Massart, Tutorial The Mahalanobis distance, Chemometr. Intell. Lab Syst. 50, 1 (2000). doi: / S (99) USP general information chapter <1119> Near-Infrared Spectroscopy ( usp31nf26secondsupplement01.html. 8. S.H.F. Scafi and C. Pasquini, Identification of counterfeit drugs using near-infrared spectroscopy, Analyst 126, 2218 (2001). doi: /b106744n

13 A. Kazeminy et al., J. Near Infrared Spectrosc. 17, xxx xxx (2009) J. Workman Jr and J. Brown, A new standard practice for multivariate, quantitative infrared analysis-part I, Spectroscopy 11(2), 48 (1996). 10. J. Workman Jr and J. Brown, A new standard practice for multivariate, quantitative infrared analysis-part II, Spectroscopy 11(9), 24 (1996). 11. A. Savitzky and M.J.E. Golay, Smoothing and differentiation of data by simplified least squares procedures, Anal. Chem. 36, 1627 (1964). doi: / ac60214a A.M.C. Davies and C. Miller, Tentative assignment of the 1440-nm absorption band in the near-infrared spectrum of crystalline sucrose, Appl. Spectrosc. 42, 703 (1988).