- Email: [email protected]
Infrared hyperspectral imaging for qualitative analysis of pharmaceutical solid forms
Analytica Chimica Acta 535 (2005) 79–87
Infrared hyperspectral imaging for qualitative analysis of pharmaceutical solid forms Y. Roggo∗ , A. Edmond, P. Chalus, M. Ulmschneider F. Hoffmann-La Roche A.G., Analytical Business Process Support, Building 65, Room 516, Grenzacherstrasse, CH-4070 Basel, Switzerland Received 5 October 2004; received in revised form 14 December 2004; accepted 14 December 2004 Available online 21 January 2005
Abstract A multi-spectral imaging spectrometer records simultaneously spectra and spatial information of samples. The infrared (IR) imaging system used was the Hyperion 3000 microscope (Bruker Optics) equipped with a focal plane array (FPA) detector. The detector allows creating a 64 × 64 pixels image. For each pixel, a complete spectrum is acquired, which means that an IR image is in fact a data cube. Two methods for qualitative analyses of the data cube were applied: peak height and unfold principal component analysis (PCA). These methods were performed on two different pharmaceutical problems: the first one was the analysis of a contamination on the surface of a pharmaceutical solid dosage form and the second one was a set of six images of intact tablets with different dissolution properties. On the first data set, IR imaging and chemometrics identified the contamination (a concentration of wet dye). The imaging method applied on the second set allowed the determination of the main cause of the dissolution problem, which was the surface distribution of magnesium stearate. This study shows that infrared imaging can be useful for qualitative analysis and troubleshooting of pharmaceutical solid forms. © 2004 Elsevier B.V. All rights reserved. Keywords: Infrared imaging; Focal plane array detector; Unfold PCA; Solid pharmaceutical; Dissolution properties; Troubleshooting
1. Introduction Infrared (IR) spectroscopy is in evolution: instrumentation and data treatment are constantly improving. When a sample is analyzed by IR spectroscopy, its homogeneity is an important issue. A spectrometer integrates the spatial information and in case of the analysis of a solid form, the use of a mean spectrum on a surface can be a drawback. For example, in the pharmaceutical industry it is important to map the distribution of the active ingredients and the excipients in a tablet [1]. Therefore, more and more studies deal with spectroscopic imaging [2,3]. Infrared multi-spectral imaging is a recent development that combines the chemical information from spectroscopy with the spatial information [4]. In principle, it is possible to collect multi-spectral images with simple point detectors, i.e. the classical mapping with IR microscopes. However, ∗
Corresponding author. Tel.: +41 61 68 81 336. E-mail address: [email protected] (Y. Roggo).
0003-2670/$ – see front matter © 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2004.12.037
the array detectors measuring simultaneously with multiple detector elements reduce the recording time, provide uniform background and improve the signal to noise ratio [5]. Fig. 1 presents the general principle of an IR imaging system and the structure of the data obtained. A FPA is an optical detector placed at the focal plane of a spectrometer and it can be manufactured to be sensitive to ultraviolet, visible, near infrared or infrared radiations. Recent developments in optics allow the production of cooled and uncooled FPAs with different numbers of pixels from 64 × 64 up to 1024 × 1024 pixels and a different spectral ranges of detection (from 1 to 12 m) [6,7]. The mercury cadmium telluride (MCT) detector has become the dominant FPA in the IR region because of the coverage of the entire IR range. By using IR imaging, a new type of data structure needs to be analyzed: an IR data cube. A monochromatic image can be presented as a two dimensional n × m array describing the distribution of the light intensity where n and m are the numbers of digitalization steps, i.e. pixels, along the x and y directions. An infrared hyperspectral image is defined by
80
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
Fig. 1. Description of the infrared imaging system: (A) instrumentation; (B) data structure.
at least 50 planes [8]. In our study, 376 wavelengths were recorded, that means the result is a 3D n × m × 376 array (Fig. 1). The aim of this paper is to present methods for the qualitative analysis of an IR image, i.e. concatenation of data cubes, peak height and unfold principal component analysis. We will discuss also how IR imaging can be a new technology to analyze pharmaceutical solid forms. Two examples will be described: the identification of a contamination on the surface of a tablet and the explanation of dissolution problems by analyzing the surface of a second type of tablets.
2.2. FT-IR measurements
2. Materials and methods
2.3. Chemometrics methods
2.1. Pharmaceuticals samples
2.3.1. Data pre-treatments The spectra of the two data sets (contamination and dissolution problems) were normalized with the standard normal variate (SNV) method, i.e. the spectra were mean centered and scaled to unit variance by spectrum. The spectral data are reduced and centred [9] by the use of the following calculation:
2.1.1. Analysis of a contamination One tablet with the presence of a blue spot on the surface was analyzed. Three pure products were analyzed as references: the active ingredient (a molecule of the benzodiazepine family), the dye (indigo carmine) and the placebo (mixture of all of the excipients without the active ingredient: avicel, indigo carmine, magnesium stearate). 2.1.2. Dissolution problems Six samples were analyzed: three of them had ‘good’ dissolution properties and three failed the dissolution test (the ‘bad’ samples). Several ingredients were used as references: avicel PH101, the active ingredient, magnesium stearate, and poloxamer 188. The dissolution testing was accomplished after the IR imaging measurement, using the Sotax Dissolution Tester AT 6 and AT 7 smart paddle stirrers (Sotax AG, Allschwil, Switzerland) in artificial gastric juice (pH 1.2, 37 ◦ C, stirred at 75 rpm). The Perkin-Elmer Lambda 40 UV spectrophotometer (Perkin-Elmer AG, H¨unenberg, Switzerland) was utilized.
An Equinox 55 spectrometer (Bruker, Ettlingen, Germany) coupled with a Hyperion 3000 microscope equipped with a 64 × 64 MCT FPA detector were used to acquire the IR spectra between 3900 and 900 cm−1 at 16 cm−1 resolution (i.e. 376 data points) under N2 purge. The binning function (i.e. pixels are grouped together) was applied to improve the signal and finally an image of 16 × 16 pixels was acquired. The number of scans was 20 and the surface analyzed by one FPA measurement was a 270 m × 270 m area.
SNVi = (xi − x¯ ) /
− x¯ )2 (w − 1) i (xi
for i ∈ [3900 cm−1 ; 900 cm−1 ]. where xi is the log(1/R) value at the wavelength i, w the number of wavelengths, x¯ is the mean of the log(1/R) values (on each segment) and SNVi is corrected log(1/R) value at the wavelength i. For the second set (dissolution problem), the SavitszkyGolay [10] method was applied to correct the baseline shift (filter length: 19 points and filter order: 3) after the normalization.
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
81
Fig. 2. Concatenation of IR images and principles of the unfold principal component analysis: (A) IR image concatenation; (B) principal component analysis; (C) creation of the scores image.
2.3.2. Concatenation of hyperspectral data cubes An A × B image means an image of A pixels in x-axis and B pixels in y-axis. The data cubes (A × B × 376 wavelengths) of the samples and of the reference materials were grouped together in order to help the image interpretation. The concatenation was done along the x and y axis to create a larger data cube (Fig. 2).
Concerning the first data set (contamination analysis), nine FPA images (3 × 3 mapping) were acquired on the tablet with the blue spot, i.e. a 48 × 48 pixels image can be displayed. Each of the three references (placebo, active ingredient, and indigo carmine) was analyzed (one FPA measurement, i.e. 16 × 16 pixels) and the IR data cubes were concatenated to allow the display of an image of 16 pixels
Fig. 3. Methods to display IR images—peak height and PCA.
82
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
in x and 48 in y. All the data cubes, samples and references, were grouped in order to get a 64 × 48 pixels image (i.e. 64 × 48 × 376 matrix). For the second data set, six samples were analyzed. Two FPA images were recorded on each tablet, i.e. for each tablet a 32 × 16 pixels image was obtained. The three data cubes of the samples with good dissolution properties were concatenated in the y-dimension to obtain a 32 × 48 image. The same type of 32 × 48 image was produced for the bad samples. Finally, the two sets of ‘bad’ and ‘good’ samples for the dissolution test were concatenated in the x-dimension to obtain a 64 × 48 image. Then the final data cube is a 64 × 48 × 376 matrix. 2.3.3. Image construction Two different methods were used to extract the chemical information and to display IR images (Fig. 3) [11–13]. • Peak height: the absorbance values at a selected wavelength for each pixel are used to display an image. • Principal component analysis (PCA): PCA forms the basics of multivariate data analysis. The most important PCA application is to reduce the number of variables and to rep-
resent a multivariate data table in a low dimensional space [14]. Thus, the new variables (loadings) are linear combinations of the original absorbances. The new coordinates were computed as follows: T = Xc·P with T score matrix, Xc mean centered spectral matrix and P loading matrix. The ‘non-linear iterative partial least squares’ (NIPALS) algorithm [15] was used for the determination of loadings and scores. Fig. 3 presents the application of the PCA method on an IR data cube. The data cubes need to be unfolded before the PCA. The three dimensions (A × B × L) matrix is replaced by a two dimensions (A × B) × L matrix. The PCA is performed and the matrix is refolded.
2.3.4. Software The software package for the data acquisition was Opus (Bruker, Ettlingen, Germany). All NIR data were exported and computed with Matlab R12 (The Mathworks, Natick, USA) and the PLS toolbox (Eigenvector, Manson, USA). The peak height and PCA methods were computed with the software Isys v3.1 (Spectral Dimensions Inc., Olney, USA).
Fig. 4. Infrared image at specific wavelengths for the first data set (identification of the blue spot): (A) indigo carmine, λ = 1095 cm−1 ; (B) benzodiazepine, λ = 1681 cm−1 ; (C) placebo, λ = 979 cm−1 . B: benzodiazepine; P: placebo; IC: indigo carmine; Ref: reference.
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
83
Fig. 5. Results of the PCA for the first data set (identification of the blue spot) and comparison of the loadings with the reference spectra. PC1: placebo; PC2: benzodiazepine; PC3: indigo carmine; PC4: CO2 ; PC5: blue spot; (A) score images; (B) loadings; (C) reference spectra; B: benzodiazepine; P: placebo; IC: indigo carmine; Ref: reference. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
3. Results and discussion 3.1. Analysis of a contamination The two methods, peak height (Fig. 4) and PCA (Fig. 5), were applied. The acquisition size on the sample was 810 m × 810 m, i.e. nine FPA measurements. Images were displayed at 1095, 1681 and 979 cm−1 to evaluate the distribution of indigo carmine, active ingredient, and placebo, respectively. Nevertheless the characterization of the blue spot was impossible. The information concerning the blue spot must be contained in several peaks or overlapping bands lead to difficulties for the interpretation of the single wavelength images. However, this method was sufficient to map the active ingredient on the surface (Fig. 4). PCA was then applied to identify the blue spot. In the first part (Fig. 5), the spectra of the sample were grouped with the spectra of pure benzodiazepine, placebo, and dry indigo carmine to perform a PCA. The number of principal components was selected in order to have interpretable loadings and images. Table 1 presents the percentages of explained variances for each principal component. As a result, placebo was visible on the whole surface and the blue spot was not pure dry indigo carmine. Moreover, principal component (PC) 5 identified the blue spot. The loading
of PC5 was interpreted and compared to the references spectra (Fig. 6A) to characterize the blue spot. The loading of the PC5 and indigo carmine (IC) had common peaks but there were also some spectral differences between PC5 and IC in the range from 3490 to 3410 cm−1 due to the OH group. The hypothesis was that the blue spot was a mixture of indigo carmine mixed with others excipients with high residual moisture content. Indigo carmine in the 64 × 48 image was replaced by a new reference sample (indigo carmine, water, and avicel). Then a second PCA (Table 1) was performed on the new image to confirm this hypothesis.
Table 1 Variance captured by the principal components for the analysis of a contaminated sample PC
PCA1
PCA2
1 2 3 4 5 6
75.01 8.82 5.19 4.04 2.76 0.85
71.6 11.93 7.66 3.97 1.23 0.74
PCA1 was the results of the principal component analysis obtained with the first image of contamination and PCA2 with the image and the new internal reference. In bold: the PC which identifies the particle.
84
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
Fig. 6. Identification of the blue spot. (A) Comparison of the loading of PC5 and the dye spectra (indigo carmine). Gray spectrum: loading 5, black spectrum: indigo carmine. (B) Second principal component analysis with the new reference (indigo carmine + water + avicel). B: benzodiazepine, P: placebo, NR: new reference sample. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
The new reference and the blue spot are both highlighted at the same time by PC2 scores image (Fig. 6B). Therefore, we were able to conclude that the contamination was wet indigo carmine particles and placebo.
3.2. Dissolution problems A specific wavelength was selected for each of the four ingredients of the tablets: 1103 cm−1 for the poloxamer,
Fig. 7. Infrared image at specific wavelengths for the second data set (dissolution problems): (A) Avicel, λ = 1056 cm−1 ; (B) active ingredient, λ = 1635 cm−1 ; (C) poloxamer, λ = 1103 cm−1 ; (D) magnesium stearate, λ = 2908 cm−1 .
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
85
Fig. 8. Scores images obtained by PCA on the second data set (dissolution problems): PC1, poloxamer; PC2, magnesium stearate; PC3, active ingredient; PC4, avicel.
2908 cm−1 for the magnesium stearate, 1635 cm−1 for the active compound and 1056 cm−1 for the avicel. The images displayed at these wavelengths are shown in Fig. 7. The differences between the two groups of samples can be detected with single wavelength images. The results of the principal component analysis are presented in Figs. 8 and 9. Table 2 shows the explained variance for each principal component. The first loading (Fig. 9) can be attributed to poloxamer, the second loading to magnesium stearate, the third one to the active ingredient, and the fourth to avicel. However, the other ingredients can also participate to the loading because loadings are not pure component spectra. Especially, the loading of the Table 2 Variance captured by the principal components for the dissolution problem PC 1 2 3 4 5 6 In bold: PCs used for the analysis.
Variance captured (%) 74.55 9.49 4.81 3.14 1.17 0.74
Fig. 9. Loadings of the PCA on the second data set (dissolution problems) and comparison with the reference spectra: (A) PC1, poloxamer; (B) PC2, magnesium stearate; (C) PC3, active ingredient; (D) PC4, avicel.
86
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
Fig. 10. Image and spectra of the sample, which was contaminated (A) before normalization (B) after SNV normalization.
PC3, which was attributed to the active ingredient, might be contaminated by other components. This study, with single specific wavelength images and with PCA images, gave similar results and showed the differences between the two sets, i.e. separation of good and bad samples for the dissolution. The main differences were due to the distributions of magnesium stearate and the active compound. No differences were observed for the spatial distribution of the poloxamer and the avicel. The magnesium stearate was hydrophobic therefore it protected the kernels of the tablet from moistening and therefore decreased the dissolution rate and when a sample had more active ingredient on the surface the dissolution properties were increased. 3.3. Chemometric methods 3.3.1. Pre-treatments The normalization method (mean centering and scaling to unit variance by spectrum) changes the amplitude of each pixel spectrum individually. This operation reduces differences due to sample presentation such as path-length, but increases the noise. It is appropriate for qualitative analysis. However, it may degrade the data quality for a quantitative analysis [16]. Nevertheless, the quality of the spectra obtained with this spectrometer on these samples is low: noise and baseline shifts can be detected. The pre-treatments are applied to
correct the baseline shift. Fig. 10 shows the spectra and the image of the contamination before and after the SNV normalization. After the normalization, the spectra of the contaminated sample and the reference spectra can be compared. Moreover, the optical system is sensitive to vibrations. This observation confirms the drawbacks of the FPA detectors described by Tran [5]. Consequently, the use of data pre-treatments is essential for the interpretation and for the qualitative comparison of the IR images. 3.3.2. Image concatenation The advantage of the concatenation is to work with all the images at a time. The entire image is pre-treated with the same methods. The image comparison is simplified as all the images are displayed with the same color scale. Moreover, the identification of the sample pixels is improved by the concatenation of the images of the samples and the references. When a pixel has the same color as one of the reference sample for a specific wavelength or for the PCA scores image, we can assume that the pixel and the reference have the same chemical composition. This approach is similar to the matrix augmentation used with the methods of multivariate curve resolution. 3.3.3. Peak height compared to PCA Peak height is a classical method to interpret infrared spectra. The main advantage of the peak height method is the
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
selection of specific wavelengths. The disadvantage is that the bands are often overlapping and finding out a specific band becomes problematic. Principal component analysis solves the problem of wavelengths selection. The main advantages of PCA are that this method reduces the number of variables (wavelengths), i.e. the number of images to analyze. The drawback is the interpretation of the loadings. Loadings differ from the spectra of pure substances and their interpretation can be difficult because several chemical species contribute to one loading.
4. Conclusions IR imaging was applied on pharmaceutical tablets to solve two different problems. Concerning the analysis of a contamination, IR imaging led to the following conclusions: no external contamination has been detected by IR imaging and the blue spot was identified as wet dye. Concerning the dissolution issue, the chemical composition of the surface explained the differences in the dissolution rate: magnesium stearate protected the kernel of the tablet from the water action and therefore decreases the dissolution rate. Spectral imaging is a complex and multidisciplinary field. The commercialization of new FPAs is making infrared imaging more and more attractive. This method has proven
87
its potential for qualitative analysis of pharmaceutical products and can be used when spatial information becomes relevant for an analytical application.
References [1] R. Smith, M. Nelson, P. Treado, Proc. SPIE 3920 (2000) 14. [2] A. El-Hagasy, H. Morris, F. D’Amico, J. Drennen, J. Pharm. Sci. 90 (2001) 1298. [3] K. Chan, S. Hammond, S. Kazarian, Anal. Chem. 75 (2003) 2140. [4] M. Morris, Microscopic and Spectroscopic Imaging of the Chemical State, Marcel Dekker, New York, 1993. [5] C. Tran, Appl. Spectrosc. Rev. 38 (2003) 133. [6] J. Barton, R. Cannata, S. Petronio, SPIE 4721 (2002) 37. [7] L. Faraone, HgCdTe Infrared Detector Technology, Norosa Publishing House, India, 1998, pp. 721–728. [8] A.D. Brianco, G. Serafino, G. Spoeck, An introduction to spectral imaging. http://www.ctr.at/hypspbvdoc/IntSpImg.pdf. [9] R. Barnes, M. Dhanoa, S. Lister, Appl. Spectrosc. 43 (1989) 772. [10] A. Savistzky, M. Golay, Anal. Chem. 36 (1964) 1627. [11] K. Artyushkova, J. Fulghum, J. Electron. Spectrosc. 121 (2001) 33. [12] B. Budevska, Vibr. Spectrosc. 24 (2000) 37. [13] M. Antunes, J. Sim˜ao, A. Duarte, M. Esteban, R. Tauler, Anal. Chim. Acta 459 (2002) 291. [14] L. Lebart, A. Morineau, M. Piro, Statistiques Exploratoires Multidimensionnelles, Dunod, Bordas, Paris, 1997, pp. 32–34. [15] P. Geladi, B. Kowalski, Anal. Chim. Acta 185 (1986) 1–17. [16] Spectral Dimension, User’s Manual, Isys 3.1 Chemical Imaging Software, Olney, MD, 2004.
Infrared hyperspectral imaging for qualitative analysis of pharmaceutical solid forms Y. Roggo∗ , A. Edmond, P. Chalus, M. Ulmschneider F. Hoffmann-La Roche A.G., Analytical Business Process Support, Building 65, Room 516, Grenzacherstrasse, CH-4070 Basel, Switzerland Received 5 October 2004; received in revised form 14 December 2004; accepted 14 December 2004 Available online 21 January 2005
Abstract A multi-spectral imaging spectrometer records simultaneously spectra and spatial information of samples. The infrared (IR) imaging system used was the Hyperion 3000 microscope (Bruker Optics) equipped with a focal plane array (FPA) detector. The detector allows creating a 64 × 64 pixels image. For each pixel, a complete spectrum is acquired, which means that an IR image is in fact a data cube. Two methods for qualitative analyses of the data cube were applied: peak height and unfold principal component analysis (PCA). These methods were performed on two different pharmaceutical problems: the first one was the analysis of a contamination on the surface of a pharmaceutical solid dosage form and the second one was a set of six images of intact tablets with different dissolution properties. On the first data set, IR imaging and chemometrics identified the contamination (a concentration of wet dye). The imaging method applied on the second set allowed the determination of the main cause of the dissolution problem, which was the surface distribution of magnesium stearate. This study shows that infrared imaging can be useful for qualitative analysis and troubleshooting of pharmaceutical solid forms. © 2004 Elsevier B.V. All rights reserved. Keywords: Infrared imaging; Focal plane array detector; Unfold PCA; Solid pharmaceutical; Dissolution properties; Troubleshooting
1. Introduction Infrared (IR) spectroscopy is in evolution: instrumentation and data treatment are constantly improving. When a sample is analyzed by IR spectroscopy, its homogeneity is an important issue. A spectrometer integrates the spatial information and in case of the analysis of a solid form, the use of a mean spectrum on a surface can be a drawback. For example, in the pharmaceutical industry it is important to map the distribution of the active ingredients and the excipients in a tablet [1]. Therefore, more and more studies deal with spectroscopic imaging [2,3]. Infrared multi-spectral imaging is a recent development that combines the chemical information from spectroscopy with the spatial information [4]. In principle, it is possible to collect multi-spectral images with simple point detectors, i.e. the classical mapping with IR microscopes. However, ∗
Corresponding author. Tel.: +41 61 68 81 336. E-mail address: [email protected] (Y. Roggo).
0003-2670/$ – see front matter © 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2004.12.037
the array detectors measuring simultaneously with multiple detector elements reduce the recording time, provide uniform background and improve the signal to noise ratio [5]. Fig. 1 presents the general principle of an IR imaging system and the structure of the data obtained. A FPA is an optical detector placed at the focal plane of a spectrometer and it can be manufactured to be sensitive to ultraviolet, visible, near infrared or infrared radiations. Recent developments in optics allow the production of cooled and uncooled FPAs with different numbers of pixels from 64 × 64 up to 1024 × 1024 pixels and a different spectral ranges of detection (from 1 to 12 m) [6,7]. The mercury cadmium telluride (MCT) detector has become the dominant FPA in the IR region because of the coverage of the entire IR range. By using IR imaging, a new type of data structure needs to be analyzed: an IR data cube. A monochromatic image can be presented as a two dimensional n × m array describing the distribution of the light intensity where n and m are the numbers of digitalization steps, i.e. pixels, along the x and y directions. An infrared hyperspectral image is defined by
80
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
Fig. 1. Description of the infrared imaging system: (A) instrumentation; (B) data structure.
at least 50 planes [8]. In our study, 376 wavelengths were recorded, that means the result is a 3D n × m × 376 array (Fig. 1). The aim of this paper is to present methods for the qualitative analysis of an IR image, i.e. concatenation of data cubes, peak height and unfold principal component analysis. We will discuss also how IR imaging can be a new technology to analyze pharmaceutical solid forms. Two examples will be described: the identification of a contamination on the surface of a tablet and the explanation of dissolution problems by analyzing the surface of a second type of tablets.
2.2. FT-IR measurements
2. Materials and methods
2.3. Chemometrics methods
2.1. Pharmaceuticals samples
2.3.1. Data pre-treatments The spectra of the two data sets (contamination and dissolution problems) were normalized with the standard normal variate (SNV) method, i.e. the spectra were mean centered and scaled to unit variance by spectrum. The spectral data are reduced and centred [9] by the use of the following calculation:
2.1.1. Analysis of a contamination One tablet with the presence of a blue spot on the surface was analyzed. Three pure products were analyzed as references: the active ingredient (a molecule of the benzodiazepine family), the dye (indigo carmine) and the placebo (mixture of all of the excipients without the active ingredient: avicel, indigo carmine, magnesium stearate). 2.1.2. Dissolution problems Six samples were analyzed: three of them had ‘good’ dissolution properties and three failed the dissolution test (the ‘bad’ samples). Several ingredients were used as references: avicel PH101, the active ingredient, magnesium stearate, and poloxamer 188. The dissolution testing was accomplished after the IR imaging measurement, using the Sotax Dissolution Tester AT 6 and AT 7 smart paddle stirrers (Sotax AG, Allschwil, Switzerland) in artificial gastric juice (pH 1.2, 37 ◦ C, stirred at 75 rpm). The Perkin-Elmer Lambda 40 UV spectrophotometer (Perkin-Elmer AG, H¨unenberg, Switzerland) was utilized.
An Equinox 55 spectrometer (Bruker, Ettlingen, Germany) coupled with a Hyperion 3000 microscope equipped with a 64 × 64 MCT FPA detector were used to acquire the IR spectra between 3900 and 900 cm−1 at 16 cm−1 resolution (i.e. 376 data points) under N2 purge. The binning function (i.e. pixels are grouped together) was applied to improve the signal and finally an image of 16 × 16 pixels was acquired. The number of scans was 20 and the surface analyzed by one FPA measurement was a 270 m × 270 m area.
SNVi = (xi − x¯ ) /
− x¯ )2 (w − 1) i (xi
for i ∈ [3900 cm−1 ; 900 cm−1 ]. where xi is the log(1/R) value at the wavelength i, w the number of wavelengths, x¯ is the mean of the log(1/R) values (on each segment) and SNVi is corrected log(1/R) value at the wavelength i. For the second set (dissolution problem), the SavitszkyGolay [10] method was applied to correct the baseline shift (filter length: 19 points and filter order: 3) after the normalization.
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
81
Fig. 2. Concatenation of IR images and principles of the unfold principal component analysis: (A) IR image concatenation; (B) principal component analysis; (C) creation of the scores image.
2.3.2. Concatenation of hyperspectral data cubes An A × B image means an image of A pixels in x-axis and B pixels in y-axis. The data cubes (A × B × 376 wavelengths) of the samples and of the reference materials were grouped together in order to help the image interpretation. The concatenation was done along the x and y axis to create a larger data cube (Fig. 2).
Concerning the first data set (contamination analysis), nine FPA images (3 × 3 mapping) were acquired on the tablet with the blue spot, i.e. a 48 × 48 pixels image can be displayed. Each of the three references (placebo, active ingredient, and indigo carmine) was analyzed (one FPA measurement, i.e. 16 × 16 pixels) and the IR data cubes were concatenated to allow the display of an image of 16 pixels
Fig. 3. Methods to display IR images—peak height and PCA.
82
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
in x and 48 in y. All the data cubes, samples and references, were grouped in order to get a 64 × 48 pixels image (i.e. 64 × 48 × 376 matrix). For the second data set, six samples were analyzed. Two FPA images were recorded on each tablet, i.e. for each tablet a 32 × 16 pixels image was obtained. The three data cubes of the samples with good dissolution properties were concatenated in the y-dimension to obtain a 32 × 48 image. The same type of 32 × 48 image was produced for the bad samples. Finally, the two sets of ‘bad’ and ‘good’ samples for the dissolution test were concatenated in the x-dimension to obtain a 64 × 48 image. Then the final data cube is a 64 × 48 × 376 matrix. 2.3.3. Image construction Two different methods were used to extract the chemical information and to display IR images (Fig. 3) [11–13]. • Peak height: the absorbance values at a selected wavelength for each pixel are used to display an image. • Principal component analysis (PCA): PCA forms the basics of multivariate data analysis. The most important PCA application is to reduce the number of variables and to rep-
resent a multivariate data table in a low dimensional space [14]. Thus, the new variables (loadings) are linear combinations of the original absorbances. The new coordinates were computed as follows: T = Xc·P with T score matrix, Xc mean centered spectral matrix and P loading matrix. The ‘non-linear iterative partial least squares’ (NIPALS) algorithm [15] was used for the determination of loadings and scores. Fig. 3 presents the application of the PCA method on an IR data cube. The data cubes need to be unfolded before the PCA. The three dimensions (A × B × L) matrix is replaced by a two dimensions (A × B) × L matrix. The PCA is performed and the matrix is refolded.
2.3.4. Software The software package for the data acquisition was Opus (Bruker, Ettlingen, Germany). All NIR data were exported and computed with Matlab R12 (The Mathworks, Natick, USA) and the PLS toolbox (Eigenvector, Manson, USA). The peak height and PCA methods were computed with the software Isys v3.1 (Spectral Dimensions Inc., Olney, USA).
Fig. 4. Infrared image at specific wavelengths for the first data set (identification of the blue spot): (A) indigo carmine, λ = 1095 cm−1 ; (B) benzodiazepine, λ = 1681 cm−1 ; (C) placebo, λ = 979 cm−1 . B: benzodiazepine; P: placebo; IC: indigo carmine; Ref: reference.
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
83
Fig. 5. Results of the PCA for the first data set (identification of the blue spot) and comparison of the loadings with the reference spectra. PC1: placebo; PC2: benzodiazepine; PC3: indigo carmine; PC4: CO2 ; PC5: blue spot; (A) score images; (B) loadings; (C) reference spectra; B: benzodiazepine; P: placebo; IC: indigo carmine; Ref: reference. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
3. Results and discussion 3.1. Analysis of a contamination The two methods, peak height (Fig. 4) and PCA (Fig. 5), were applied. The acquisition size on the sample was 810 m × 810 m, i.e. nine FPA measurements. Images were displayed at 1095, 1681 and 979 cm−1 to evaluate the distribution of indigo carmine, active ingredient, and placebo, respectively. Nevertheless the characterization of the blue spot was impossible. The information concerning the blue spot must be contained in several peaks or overlapping bands lead to difficulties for the interpretation of the single wavelength images. However, this method was sufficient to map the active ingredient on the surface (Fig. 4). PCA was then applied to identify the blue spot. In the first part (Fig. 5), the spectra of the sample were grouped with the spectra of pure benzodiazepine, placebo, and dry indigo carmine to perform a PCA. The number of principal components was selected in order to have interpretable loadings and images. Table 1 presents the percentages of explained variances for each principal component. As a result, placebo was visible on the whole surface and the blue spot was not pure dry indigo carmine. Moreover, principal component (PC) 5 identified the blue spot. The loading
of PC5 was interpreted and compared to the references spectra (Fig. 6A) to characterize the blue spot. The loading of the PC5 and indigo carmine (IC) had common peaks but there were also some spectral differences between PC5 and IC in the range from 3490 to 3410 cm−1 due to the OH group. The hypothesis was that the blue spot was a mixture of indigo carmine mixed with others excipients with high residual moisture content. Indigo carmine in the 64 × 48 image was replaced by a new reference sample (indigo carmine, water, and avicel). Then a second PCA (Table 1) was performed on the new image to confirm this hypothesis.
Table 1 Variance captured by the principal components for the analysis of a contaminated sample PC
PCA1
PCA2
1 2 3 4 5 6
75.01 8.82 5.19 4.04 2.76 0.85
71.6 11.93 7.66 3.97 1.23 0.74
PCA1 was the results of the principal component analysis obtained with the first image of contamination and PCA2 with the image and the new internal reference. In bold: the PC which identifies the particle.
84
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
Fig. 6. Identification of the blue spot. (A) Comparison of the loading of PC5 and the dye spectra (indigo carmine). Gray spectrum: loading 5, black spectrum: indigo carmine. (B) Second principal component analysis with the new reference (indigo carmine + water + avicel). B: benzodiazepine, P: placebo, NR: new reference sample. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
The new reference and the blue spot are both highlighted at the same time by PC2 scores image (Fig. 6B). Therefore, we were able to conclude that the contamination was wet indigo carmine particles and placebo.
3.2. Dissolution problems A specific wavelength was selected for each of the four ingredients of the tablets: 1103 cm−1 for the poloxamer,
Fig. 7. Infrared image at specific wavelengths for the second data set (dissolution problems): (A) Avicel, λ = 1056 cm−1 ; (B) active ingredient, λ = 1635 cm−1 ; (C) poloxamer, λ = 1103 cm−1 ; (D) magnesium stearate, λ = 2908 cm−1 .
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
85
Fig. 8. Scores images obtained by PCA on the second data set (dissolution problems): PC1, poloxamer; PC2, magnesium stearate; PC3, active ingredient; PC4, avicel.
2908 cm−1 for the magnesium stearate, 1635 cm−1 for the active compound and 1056 cm−1 for the avicel. The images displayed at these wavelengths are shown in Fig. 7. The differences between the two groups of samples can be detected with single wavelength images. The results of the principal component analysis are presented in Figs. 8 and 9. Table 2 shows the explained variance for each principal component. The first loading (Fig. 9) can be attributed to poloxamer, the second loading to magnesium stearate, the third one to the active ingredient, and the fourth to avicel. However, the other ingredients can also participate to the loading because loadings are not pure component spectra. Especially, the loading of the Table 2 Variance captured by the principal components for the dissolution problem PC 1 2 3 4 5 6 In bold: PCs used for the analysis.
Variance captured (%) 74.55 9.49 4.81 3.14 1.17 0.74
Fig. 9. Loadings of the PCA on the second data set (dissolution problems) and comparison with the reference spectra: (A) PC1, poloxamer; (B) PC2, magnesium stearate; (C) PC3, active ingredient; (D) PC4, avicel.
86
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
Fig. 10. Image and spectra of the sample, which was contaminated (A) before normalization (B) after SNV normalization.
PC3, which was attributed to the active ingredient, might be contaminated by other components. This study, with single specific wavelength images and with PCA images, gave similar results and showed the differences between the two sets, i.e. separation of good and bad samples for the dissolution. The main differences were due to the distributions of magnesium stearate and the active compound. No differences were observed for the spatial distribution of the poloxamer and the avicel. The magnesium stearate was hydrophobic therefore it protected the kernels of the tablet from moistening and therefore decreased the dissolution rate and when a sample had more active ingredient on the surface the dissolution properties were increased. 3.3. Chemometric methods 3.3.1. Pre-treatments The normalization method (mean centering and scaling to unit variance by spectrum) changes the amplitude of each pixel spectrum individually. This operation reduces differences due to sample presentation such as path-length, but increases the noise. It is appropriate for qualitative analysis. However, it may degrade the data quality for a quantitative analysis [16]. Nevertheless, the quality of the spectra obtained with this spectrometer on these samples is low: noise and baseline shifts can be detected. The pre-treatments are applied to
correct the baseline shift. Fig. 10 shows the spectra and the image of the contamination before and after the SNV normalization. After the normalization, the spectra of the contaminated sample and the reference spectra can be compared. Moreover, the optical system is sensitive to vibrations. This observation confirms the drawbacks of the FPA detectors described by Tran [5]. Consequently, the use of data pre-treatments is essential for the interpretation and for the qualitative comparison of the IR images. 3.3.2. Image concatenation The advantage of the concatenation is to work with all the images at a time. The entire image is pre-treated with the same methods. The image comparison is simplified as all the images are displayed with the same color scale. Moreover, the identification of the sample pixels is improved by the concatenation of the images of the samples and the references. When a pixel has the same color as one of the reference sample for a specific wavelength or for the PCA scores image, we can assume that the pixel and the reference have the same chemical composition. This approach is similar to the matrix augmentation used with the methods of multivariate curve resolution. 3.3.3. Peak height compared to PCA Peak height is a classical method to interpret infrared spectra. The main advantage of the peak height method is the
Y. Roggo et al. / Analytica Chimica Acta 535 (2005) 79–87
selection of specific wavelengths. The disadvantage is that the bands are often overlapping and finding out a specific band becomes problematic. Principal component analysis solves the problem of wavelengths selection. The main advantages of PCA are that this method reduces the number of variables (wavelengths), i.e. the number of images to analyze. The drawback is the interpretation of the loadings. Loadings differ from the spectra of pure substances and their interpretation can be difficult because several chemical species contribute to one loading.
4. Conclusions IR imaging was applied on pharmaceutical tablets to solve two different problems. Concerning the analysis of a contamination, IR imaging led to the following conclusions: no external contamination has been detected by IR imaging and the blue spot was identified as wet dye. Concerning the dissolution issue, the chemical composition of the surface explained the differences in the dissolution rate: magnesium stearate protected the kernel of the tablet from the water action and therefore decreases the dissolution rate. Spectral imaging is a complex and multidisciplinary field. The commercialization of new FPAs is making infrared imaging more and more attractive. This method has proven
87
its potential for qualitative analysis of pharmaceutical products and can be used when spatial information becomes relevant for an analytical application.
References [1] R. Smith, M. Nelson, P. Treado, Proc. SPIE 3920 (2000) 14. [2] A. El-Hagasy, H. Morris, F. D’Amico, J. Drennen, J. Pharm. Sci. 90 (2001) 1298. [3] K. Chan, S. Hammond, S. Kazarian, Anal. Chem. 75 (2003) 2140. [4] M. Morris, Microscopic and Spectroscopic Imaging of the Chemical State, Marcel Dekker, New York, 1993. [5] C. Tran, Appl. Spectrosc. Rev. 38 (2003) 133. [6] J. Barton, R. Cannata, S. Petronio, SPIE 4721 (2002) 37. [7] L. Faraone, HgCdTe Infrared Detector Technology, Norosa Publishing House, India, 1998, pp. 721–728. [8] A.D. Brianco, G. Serafino, G. Spoeck, An introduction to spectral imaging. http://www.ctr.at/hypspbvdoc/IntSpImg.pdf. [9] R. Barnes, M. Dhanoa, S. Lister, Appl. Spectrosc. 43 (1989) 772. [10] A. Savistzky, M. Golay, Anal. Chem. 36 (1964) 1627. [11] K. Artyushkova, J. Fulghum, J. Electron. Spectrosc. 121 (2001) 33. [12] B. Budevska, Vibr. Spectrosc. 24 (2000) 37. [13] M. Antunes, J. Sim˜ao, A. Duarte, M. Esteban, R. Tauler, Anal. Chim. Acta 459 (2002) 291. [14] L. Lebart, A. Morineau, M. Piro, Statistiques Exploratoires Multidimensionnelles, Dunod, Bordas, Paris, 1997, pp. 32–34. [15] P. Geladi, B. Kowalski, Anal. Chim. Acta 185 (1986) 1–17. [16] Spectral Dimension, User’s Manual, Isys 3.1 Chemical Imaging Software, Olney, MD, 2004.