Using MicroNIR Spectroscopy and Supervised Pattern Recognition in Identification of Pharmaceutical Raw Materials

Table of Contents

Introduction
Experimental Details
     Materials
     Spectra Collection
     Spectral Pretreatment and Chemometric Analysis
Results and Discussion
     Model Transferability
     Large-Scale Classification
Conclusion

Introduction

Raw material identification (RMID) or verification of the packaging label is a quality-control practice commonly employed in the pharmaceutical industry. The increasing global footprint of the supply chain and public health concerns resulting from mislabeled or contaminated materials have driven many regulatory bodies to require inspection of every barrel in every shipment of materials used in pharmaceutical drugs. Traditionally, pharmaceutical RMID depended on laboratory-based analytical techniques such as wet chemistry, chromatography and titrations among others. Most of these techniques are time consuming, destructive in nature and labor intensive, and thus it is challenging to handle an enormous number of analyzes.1

In recent years, vibrational spectroscopy, including mid-infrared (mid-IR), near-infrared (NIR) and Raman spectroscopy, has been widely accepted in the pharmaceutical industry for RMID due to its non-destructive nature, fast data acquisition and minimal sample preparation. Especially, with substantial progress in portable mid-IR, NIR and Raman spectrometers, on-site and in situ analysis of an increasing number of samples has become practical for material identification, which opens up more application opportunities.2

Among the three vibrational spectroscopic techniques, IR and NIR measure absorbance, while Raman measures scattering. Raman is sensitive to the change in the polarizability of a vibrating molecule, and NIR and IR are sensitive to the change in the dipole moment of a vibrating molecule. When compared to NIR, mid-infrared is less popular in RMID due to the strong absorption coefficient in the mid-IR spectral range, which limits the path length into the samples and sometimes needs dilution of the samples using infrared transparent materials.3 In general, Raman and NIR are complementary in nature. Both these techniques have a wide range of applications in pharmaceutical analysis,4,5 however they have their own advantages and disadvantages.6 Raman spectroscopy has exceptional molecular selectivity, can be effortlessly used in a non-contact fashion through common container materials, and is free of water interference from aqueous solutions. However, the high energetic laser power may decompose sensitive samples, and interference from fluorescent molecules can be a limitation. Conversely, NIR spectroscopy does not suffer from the fluorescence problem and can also measure through glass or plastic containers. The complexity of the spectra is the limiting factor of NIR, thus low molecular selectivity, resulting from combination bands and vibrational overtones, which require the use of multivariate data analysis. Over the past decade, the algorithms and computing power have improved dramatically permitting NIR to become user friendly and more powerful. NIR has been selected as the analytical tool in this article for pharmaceutical RMID.

Generally, major pharmacopoeias have adopted near-infrared techniques. The suitability of NIR instrumentation for application in pharmaceutical testing has been addressed by the United States Pharmacopoeia (Chapter 1119)7 and the European Pharmacopoeia (Chapter 2.2.40)8. Luypaert et al. reviewed a variety of NIR applications for pharmaceutical material identification,4 such as identifying commonly used excipients and active pharmaceutical ingredients (API),9,10 distinguishing between closely related substances,11,12 and classifying varied polymorphic forms of the same product.13,14 Very recently, Grout included NIR material qualification outputs with statistical process control (SPC) charts (through historical trending) to link material attributes to both process behavior and product quality, which brings about rapid material qualification on receipt with improved understanding of process performance.15 Additionally, miniaturized NIR spectrometers became commercially available in the last couple of years. Their performance for pharmaceutical applications has been successfully demonstrated,2,16 which indeed facilitates more flexible and practical use of NIR spectrometers for on-site material identification.

Near-infrared-based pharmaceutical RMID requires the help of chemometric tools for classification due to the complex nature of NIR spectra, which is basically a pattern recognition problem. The pattern recognition techniques are divided into two categories, the supervised and the unsupervised methods, with the former being most commonly used in pharmaceutical applications.17 Initially, in a classification study, unsupervised principal component analysis (PCA) is often used in order to show discriminating tendencies. This is followed by the application of more effective supervised analyses for improved discrimination.18 Correlation based methods, distance based methods, linear discriminant analysis (LDA), soft independent modeling of class analogy (SIMCA), and partial least squares discriminant analysis (PLS-DA) are classical methods for the supervised classification.17 For instance, Blanco and Romero built a library including NIR spectra for 125 different raw materials using the correlation coefficient as the discriminating criterion.9 Dreassi et al. utilized LDA to distinguish pharmaceutical compounds with a range of physical properties.19 Krämer and Ebel demonstrated the discrimination of microcrystalline and powdered celluloses as well as cellulose and cellulose ethers by SIMCA.12 Chemical quality of 7-aminocephalosporanic acid was assessed by Andre from a large number of production lots using PLS-DA.20 Some additional methods were also discussed in the review article by Roggo et al.17

A number of challenges are yet to be addressed in spite of the successful applications of NIR for pharmaceutical RMID in the past. Model transferability, i.e., transferring the chemometric model from one or more master instruments to multiple target instruments, is one such challenge. Often, the tasks of RMID can be fulfilled on a regular basis by using multiple instruments. Built-in differences and changes induced by wear and varying environments in the instrument response function can render a model established on one instrument invalid on another. However, it is often not economical and/or practical to develop models for individual instruments. Thus, transferable robust models that need the least number of master instruments are considered to be highly desirable for RMID. Another challenge of NIR-based RMID is dealing with a large library of NIR spectra when the total number of classes reaches hundreds, which is not uncommon for RMID, as there are a variety of APIs and excipients with varied physical properties from different manufacturers and different lots involved. In this case, some conventional pattern classifiers suffer from the resolution issue that the chemometric models’ discrimination power is diluted by the huge number of classes and thus unable to distinguish smaller differences among them.

In recent years, the support vector machine (SVM), initially popular in the neural networks and machine learning community, has been introduced to chemometrics and has proven to be powerful in NIR spectra classification.21,22 However, this classifier has been less studied in NIRbased pharmaceutical RMID. This article successfully addresses the challenges of model transferability and largescale classification in NIR-based RMID by the use of SVM.

Moreover, miniature MicroNIR spectrometers were used for all of the measurements. The MicroNIR spectrometer is designed to use a novel thin-film linear variable filter (LVF) as the dispersive element on top of an InGaAs array detector. The filter coating is physically tapered with position, resulting in a constant change in the center wavelength of the filter with position. This ultra-compact spectrometer offers low power consumption, portability, stability, ruggedness and high-speed measurement, and has been successfully used in several applications.16,23–26 The use of MicroNIR spectrometers with exceptional performance consistency and the powerful SVM algorithm makes reliable and rapid on-site RMID highly achievable.

Experimental Details

Materials

Two sets of materials were used in this work. The first set consisted of 19 of the most frequently used APIs and excipients, including ascorbic acid, acetaminophen, aspirin, benzocaine, cellulose, caffeine, corn starch, fructose, (hydroxypropyl)methyl cellulose (HPMC), hydroxypropyl cellulose (HPC), ibuprofen, lactose, magnesium stearate (Mg-stearate), poly(ethylene oxide) (PEO), polyvinylpyrrolidone (PVP), polysorbate 80, sodium starch glycolate (SSG), titanium dioxide (TiO2), and talc. For each compound, two to three different products were purchased, any these products were different in physical properties and/or grades, or were from different manufacturers. The samples were divided into lots A, B, and C. This set of samples was used to study model transferability. The second set consisted of 253 commonly used APIs and excipients that were provided by one of the collaborators in the pharmaceutical industry. Multiple samples were collected, for each compound, in order to include natural variability. This set of samples was used to examine large-scale classification.

Spectra Collection

Miniature MicroNIR Pro 1700 spectrometers developed and commercialized by Viavi Solutions Inc. (formerly JDSU, Santa Rosa, CA) were employed for all data acquisition. All spectra were gathered using MicroNIR Pro spectrometer software version 2.1 (Viavi Solutions Inc.) in the range of 908–1676 nm. A reference measurement was performed on the MicroNIR almost 15 min after the lamps were turned on and every hour thereafter while performing scans. A 99% diffuse reflectance panel was used for the 100% reference value and the 0% reference value was obtained by leaving the tungsten lamps on with an empty vial holder. This scenario was used to account for any scattered light from the sample vial holder.

The spectra of the pharmaceutical materials were collected in ambient conditions. As shown in Figure 1, all samples were presented to the spectrometer housed in 14 mm diameter borosilicate glass vials with measurements carried out through the bottom of the vials. The operator used a vial holder that slips over the end of the spectrometer to effortlessly introduce and remove samples while maintaining an optimal 3 mm distance between the material and the sapphire window of the spectrometer. After every scan, the vial was rotated approximately 10–15º. Each scan had an integration time of 10 ms with spectrum averaged over 50 collections.

Hanhheld MicroNIR

Figure 1. MicroNIR spectrometer equipped with a vial holder and tethered to a rugged 7’’ Windows 8.1 tablet for pharmaceutical raw material identification.

For the set of 19 compounds, three vials were prepared for every sample in lots A, B, and C, respectively. Data acquisition was performed on three different days and at different times of the day using varied vials of the same sample to take into account varying ambient temperature. A minimum of five spectra were collected for each vial. Six spectrometers were used to collect data from the set of 19 pharmaceutical compounds in order to study the model transferability. For the set of 253 compounds, a minimum of 20 spectra were collected from multiple samples for each compound. Figure 1 also illustrates the portability of the device for on-site and in situ pharmaceutical RMID with spectra acquisition by the ultra-compact MicroNIR spectrometer tethered to a rugged seven-inch Windows 8.1 tablet.

Spectral Pretreatment and Chemometric Analysis

All the steps of spectral processing and chemometric analysis were executed using Matlab (The MathWorks, Inc.). All of the spectra collected were pretreated using Savitzky–Golay first derivative followed by standard normal variate (SNV). Soft independent modeling of class analogy (SIMCA), LDA, PLS-DA, SVM, and quadratic discriminant analysis (QDA) were applied to all of the preprocessed data. Their performance was compared for large-scale classification and model transferability. It should be noted that Matlab was chosen in this study to automatize the procedure of handling multiple chemometric models. In fact, the steps involved in spectral pretreatment, model validation and model building for the majority of the algorithms tested in this study can be performed using The Unscrambler (Camo Software Inc.), which is the default data processing software for MicroNIR products.

Results and Discussion

It is essential to select suitable chemometric models to enable reliable NIR-based pharmaceutical RMID. SVM finds increasing interest in chemometrics recently as a powerful classifier even though it is not as well-known in pharmaceutical RMID. The focus of this article was to test the performance of SVM in pharmaceutical RMID. The key concept of SVM is the use of maximized margin hyperplanes to define decision boundaries separating data points into varied classes, which provides SVM extremely good generalization capabilities. In this work, two different kernel functions, linear (SVM-linear) and radial basis function (RBF) (SVM-rbf), were used to test SVM models. A third SVM model using linear kernel and a hierarchical scheme (hier-SVM-linear) was also tested. In light of recent development of hierarchical SVM classifiers,27,28 an ubiquitous hierarchical scheme was produced and used in this work, which involves multilevel classification with upper-level classification refining lower-level classification that comprises of classes with chemical similarity to attain more accurate prediction. For comparison, four classical models were selected for testing, including SIMCA, LDA, PLSDA, and QDA. Soft independent modeling of class analogy (SIMCA) considers each class separately by performing PCA for each class and place additional emphasis on similarity within a class than on discrimination between classes. PLS-DA also emphasizes the discrimination between classes by rotating the principal components (PCs) such that a maximum separation among classes is obtained. QDA and Linear discriminant analysis (LDA) focus on finding linear or quadratic discriminant functions in order to maximize separation among the different classes. From a practical perspective, several factors are considered to compare these models in this work, including prediction success rate, scope of the spectral library, model building time, and model transferability.

Model Transferability

At first, all models were analyzed using the library of 19 pharmaceutical compounds. For this type of library with a comparatively small number of materials, a successful classification of all the compounds is required based on literature data and past experience. However, what needs to be proven is the instrument-to-instrument transfer of methods. The method transferability is vital for point-of-use NIR applications, particularly when employed in large scale and when a method is built in one location and is then required to be used at several different locations. Therefore, this study was to compare model transferability of the seven different models while using MicroNIR spectrometers. In order to carry out this, the researchers used six different MicroNIR spectrometers to collect six libraries of spectral data, covering variations from operators, instruments, chemicals, and environmental factors. They determined the number of spectrometers needed for developing a robust model that will be successful when employed to spectrometers that were not part of the training set.

For every chemometric model, the process started with developing a training model with the use of the consolidated library of training data gathered by different numbers of spectrometers (nTUx, x = 1–6). Then, the model was employed to make predictions for the test set gathered by the remaining spectrometers. This process was repeated for each case (x = 1–6) until consolidated libraries collected through all possible combinations of spectrometers were employed for training. For each repetition, the prediction success rate was defined as the number of correct predictions divided by the total number of predictions obtained and the average value reported. In the case of nTU6, when all the six spectrometers were employed for training, the training set contained spectra chosen by the Kennard–Stone sample selection scheme,29 while the remaining spectra served as the test set.

The Table 1 shows the average value of prediction success rate of each case for each model, and the number of spectrometers required to achieve 100% classification prediction was defined for each model. It can be clearly seen that the prediction success rate was growing with the number of spectrometers for training until it achieved 100%. The hier-SVM-linear and SVM-linear models only needed one spectrometer for training to obtain perfect prediction, and the SVM-rbf model required two. On the contrary, five or six spectrometers were needed to obtain perfect prediction for the other models. These results show that SVM algorithms outperformed all the other algorithms with regard to model transferability.

Table 1. Prediction accuracy as a function of number of spectrometers to build training models.

Classifier nTU1 nTU2 nTU3 nTU4 nTU5 nTU6 # Units for 100%
SIMCA 97.40 98.85 99.50 99.81 99.96 100 6
PLS-DA 99.66 99.68 99.90 99.96 99.99 100 6
LDA 86.41 97.84 98.83 99.98 100 100 5
QDA 86.47 97.88 98.85 99.98 100 100 5
SVM-rbf 99.66 100 100 100 100 100 2
SVM-linear 100 100 100 100 100 100 1
Hier-SVM-linear 100 100 100 100 100 100 1

As an example, the training set obtained by one spectrometer was used to predict the test set obtained by another spectrometer using the SVM model. As shown in Figure 2, the clear partitions of classes as well as successful predictions are visualized by the PCA-SVM plot, where the collected data was plotted against PC1 and PC2 and SVM calculations generated the class boundaries. The filled triangles in each color-coded class signify the support vectors that constrain the margin width between different classes and the open triangles signify the remaining training data. The color-coded stars signify the predicted data with the color indicating the real identity of the test sample. As shown in Figure 2, the prediction success rate for the data was 100%. Since the data structure of the spectra has 121 dimensions and the PCA-SVM plot shows only two, the few TiO2 data not completely enclosed by the boundary require a higher dimensional space in order to visualize the classification.

PCA-SVM plot for 19 pharmaceutical compounds based on the SVM model.

Figure 2. PCA-SVM plot for 19 pharmaceutical compounds based on the SVM model.

It must be noted that for the library of 19 pharmaceutical compounds, with SIMCA and PLS-DA, prediction accuracy of >98.8% was obtained with only two instruments for training. This outstanding instrument-to-instrument method transferability can be credited to the reliable performance of MicroNIR spectrometers physically, spectrally, and optically, which was enabled by the photometric stability and wavelength calibration reproducibility of the MicroNIR spectrometers.23

The researchers conducted more stringent model validation to further prove the robustness of the SVM model and its excellent transferability. The materials in the library of 19 pharmaceutical compounds were from three separate lots (lots A, B, and C) with three products different in physical properties and/or grades, or from different manufacturers for each compound. So, the researchers were able to use data from one lot (lot A) as the training set, while data from the remaining two lots (lots B and C) was used the external testing set. Three validation tests were prepared to systematically test the effect of material and instrument variation and strength of the calibration model. For the baseline case, the same-unit–same-lot test (SUSL), the datasets combined from all three lots were used with half of the data for training, while the other half for testing. The six spectrometers were analyzed respectively. For the second case, the same-unit–cross-lot test (SUXL), the lot A data from one unit was used as the training set, while the lot B and lot C data from a different unit was used as the testing set. The six spectrometers were analyzed respectively. For the third case, the cross-unit–cross-lot test (XUXL), the lot A data collected from one unit was used as the training set, while the lot B and lot C data from a different unit was used as the testing set. In total, pairwise cross-unit validation was carried out between the six spectrometers with 30 assessments. The XUXL case replicated the real world situation, where the calibration model has to be transferred to different instruments, applied to materials from different lots and/or sources, and used at different test locations. For each case, six models were compared. The hier-SVM-linear model was not included for the test here, because theoretically, the results from hier-SVM-linear model are projected to be the same as from the SVMlinear model when the number of classes is comparatively small.

The validation results are shown in Figure 3 and Table 2. Individual prediction success rates are shown in Figure 3 for each case: (a) SUSL; (b) SUXL; and (c) XUXL. Average prediction success rates and the corresponding regular deviation across various spectrometers or different combinations of spectrometers are shown in Table 2. As shown in Figure 3a and Table 2, all the models performed well and most of the prediction success rates reached nearly 100%. The lowest prediction success rate for sixth spectrometer for both QDA and LDA models was 92.44%. For the SUXL case, the prediction success rate remained above 90% for all the models using all the spectrometers. However, from Figure 3b and Table 2 it can be clearly seen that the two SVM models outperformed other models with most of the prediction success rates more than 97% except for the second spectrometer, where the prediction success rate was 94.23% using the SVM-rbf model. SIMCA provided the lowest prediction success rate compared to other models. As shown in Figure 3c and Table 2, the prediction success rate for the XUXL case significantly reduced when using QDA, LDA, and SIMCA. The lowest success rate for the QDA and LDA models was below 62%. The lowest success rate for the SIMCA model was below 60%. The SVM models outperformed other models in terms of both the average success rate and consistency in the high success rate (>94%) for all the data points as shown in Figure 3c. Between the two SVM models, SVM-linear model performed better with prediction success rate of 100% for 29 out of the 30 pairs, except when using the first spectrometer for testing and the second spectrometer for training, where the success rate was 96.43%, which was still higher compared to the success rates achieved by all the other models using the same two spectrometers. PLS-DA performed better than QDA, LDA, and SIMCA with prediction success rate ranging from 88.26% to 97.08%. These validation results clearly show that the SVM-linear model is better in terms of both model transferability and generalization capability.

Figure 3. Model validation using different lots of material and different MicroNIR units. (a) Same-unit-same-lot (SUSL) validation; (b) same-unit-cross-lot (SUXL) validation; (c) cross-unit-cross-lot (XUXL) validation. Unit# denotes the spectrometer number. T#P# denotes the spectrometer number for training (T#) and the spectrometer number for testing or prediction (P#).

Table 2. Model validation using different lots of materials and different MicroNIR spectrometers.

Classifier SUSL SUXL XUXL
AVGa STDb AVG STD AVG STD
SIMCA 100 0 90.93 0.22 83.92 8.92
PLS-DA 99.98 0.06 93.94 1.03 92.69 1.94
LDA 98.49 2.99 93.52 1.32 85.64 7.93
QDA 98.49 2.99 93.55 1.25 85.73 7.91
SVM-rbf 99.72 0.47 98.38 2.33 98.01 2.26
SVM-linear 100 0 99.54 1.14 99.88 0.65

a AVG: Average
b STD: Standard deviation

Large-Scale Classification

The library of 253 pharmaceutical compounds was used to examine the capability of the chemometric models in order to solve the large-scale classification problem. Various types of differences were contained in this library. The key difference was chemical structure, but for a number of compounds, there were also differences in potency, particle size, and formulation (coated versus uncoated) of the same materials. The objective was to test if all these 253 compounds with both physical and chemical differences can be discriminated simultaneously.

Figure 4a shows how the raw spectra of different compounds displayed broad bands and large variability in baseline, shape, and intensity over the whole spectral range. Spectral pretreatment is required minimize or eliminate variability unrelated to the main spectral features for classification. Usually, the unwanted variation is caused by uncontrollable physical effects such as changes in refractive index, non-homogeneous distribution of the particles, sample morphology, sample packing/density variability, etc. However, it must be noted that some spectral features linked with physical properties are important here because the researchers also wanted to differentiate compounds with various physical properties besides chemically different compounds. Therefore, Savitzky–Golay first derivative was selected followed by SNV for moderate spectral pretreatment. More resolvable peaks were observed and the baseline drift was reduced in the pretreated spectra as shown in Figure 4b.

Figure 4. NIR spectra of the training set from 253 pharmaceutical compounds. (a) Raw spectra; (b) pretreated spectra.

All the seven models were then applied to the pretreated spectra of the 253 pharmaceutical compounds. Half of the spectra were used as the test set and the other half was used as the training set. The results are shown in Table 3 and the prediction accuracy is compared. The PLS-DA model performed most poorly among all these models. A required condition for PLS-DA to work reliably is that each class is tight and uses a small and separate volume in X-space comprising of the multivariate data, which was not satisfied by the dataset. Experience reveals that PLS-DA is more suitable for a small number of classes,30 and thus not appropriate for large-scale classification. Furthermore, it took over 20 hours to develop the PLS-DA model in this study. QDA and LDA demonstrated excellent performance for this dataset, which can possibly be attributed to excellent settings of this dataset satisfying the desired conditions for QDA and LDA, such as multivariate normal distribution of the class population as well as equal co-variances in the classes.31 It only took seconds to develop these models. However, it must be noted that QDA and LDA did not perform well when compared with other models in terms of model transferability (Table 1, Table 2, and Figure 3). In addition, when an additional compound with very less number of spectra was included (imbalanced dataset), the performance of QDA and LDA significantly worsened, while there were only small changes in prediction success rate for the other models (data not shown). For this dataset, the soft independent modeling of class analogy (SIMCA) also performed well. This good performance could be explained by the unique feature of SIMCA that PCA is used to each group separately and the number of PCs is chosen individually and not together for all groups, which enables an optimal dimension reduction in each group to reliably classify new objects.32 However, it took minutes to develop the SIMCA model, much slower compared to the QDA/LDA models, and the model transferability of SIMCA was among the worst (Table 1, Table 2, and Figure 3).

Table 3. Comparison of different models for classification of the 253 pharmaceutical compounds.

Classifier No. of spectra Prediction success rate (%) No. of missed predictions
SIMCA 2566 97.54 63
PLS-DA 2566 85.23 379
LDA 2566 99.61 10
QDA 2566 99.73 7
SVM-rbf 2566 85.78 365
SVM-linear 2566 96.57 88
Hier-SVM-linear 2566 100 0

One of the key features of SVM models is that they can work in a kernel-induced feature space enabling both linear and nonlinear modeling.21,22 In this study, the linear kernel (SVM-linear) performed well with a slightly lower prediction success rate compared to SIMCA, whereas the nonlinear kernel (SVM-rbf) did not perform well. It must be noted that parameter optimization is often carried out to develop a SVM model, one parameter (C) for the linear kernel and two parameters (C and γ) for the RBF kernel, which generally involves a thorough search algorithm and takes a long computation time. From the practical point of view, to build a robust method with minimum model building time and efforts, the researchers deliberately skipped the parameter optimization process in order to test the SVM models using the default settings (C = 1.0 and γ = 1/number of variables). The prediction success rate of the SVMlinear model was 96.57%, which was excellent considering that no optimization was carried out, and it only took seconds to develop the model. The excellent performance of the linear kernel could be linked with the linear relationship between the concentration and the absorbance of the material according to Beer’s law. In fact, if optimization was carried out, the performance of SVM-rbf was also good whereby the prediction success rate increased from 85.78% to 98.52%.

In order to further enhance the performance of SVM, a hierarchical scheme was used with the linear kernel (hier-SVMlinear) without parameter optimization. This approach increased the prediction accuracy to 100%. In pattern classification, the variation between the probability of assigning the sample to the winner class (the maximum probability) and the probability of assigning the sample to the second place class (the second maximum probability) is the solution for prediction accuracy. In this study, the multi-class classification was achieved by using pairwise comparisons. The probability analyzed based on a multi-class probability estimate by combining the pairwise class probabilities with the use of the algorithm proposed by Wu et al.33 In Figure 5, both the maximum probability and the second maximum probability of categorizing all the samples in the test set are shown. It is very clear that for all the spectra, the variations between the two probabilities were important, more than 0.33 for all the data except one of 0.08, signifying the robust discrimination power of hier-SVM-linear. Though it only took seconds to develop the hierSVM-linear model, it was slightly longer than the SVM-linear model.

Figure 5. Estimation probabilities of class membership.

In order to further confirm the generalization capability of the SVM models, external validation was performed. Spectra from 14 compounds out of 19 compounds, which are also included in the library of 253 compounds, were employed as the testing set. These 14 compounds are polysorbate 80, HPC, PVP, SSG, acetaminophen, talc, ascorbic acid, corn starch, HPMC, lactose, benzocaine, Mg-stearate, TiO2 and caffeine. In order to be compared with the testing set, for these 14 chemically different materials in the library of 253 compounds, spectra from various compounds of the same chemical structure were combined to remove physical differences. After the consolidation, there were 231 classes in the training set. This validation test represents possibly the most challenging condition, and often a real world situation, where a much larger number of classes, including chemically similar compounds to the compounds in the testing set, were employed in the training set than in the testing set. It must be noted that the spectra in the testing sets were collected in the lab using six spectrometers, while the training set was collected from various samples with a use of a different spectrometer (the seventh spectrometer) in a different environment by an external collaborator in the pharmaceutical industry. Hence, this external validation test represented real-world application. It can not only show the predictive power of the model, but also show the model transferability.

The individual prediction success rate with the use of different spectrometers for the testing sets and the average value based on all the six spectrometers for each model are shown in Table 4. In both model transferability and predictive power, support vector machine (SVM) models have clearly shown excellent performance. The average prediction success rates achieved by the three SVM models (>90% in most cases) were significantly higher compared to the other models. Moreover, the excellent performance was more consistent across different spectrometers. The hierSVM-linear model has performed well, with prediction success rates ranging between 92.85% and 97.92% across the six spectrometers, which agrees with the previous results. The researchers closely examined the misclassification results for the SVM models and observed that a majority of the spectra, which were classified incorrectly, were predicted as the classes of chemically similar materials. PLS-DA performed better among the other models. This model exhibited pretty good model transferability, as shown in Table 4 and Figure 3. However, it took longer time (>20 h) to develop the model and the prediction accuracy was not enough for large-scale classification. In this validation test, QDA and LDA performed very poorly. Successful on-site and in situ pharmaceutical RMID using NIR requires a robust chemometric model with reliable model transfer from instrument to instrument, high prediction success rate, the capability to control a large number of materials with both physical and chemical differences, as well as simple model settings and short model building time for easy building and rebuilding of model when required. Support vector machine (SVM) modeling, particularly the hier-SVM-linear algorithm, meets all these requirements, and it can be potentially used as a powerful classification technique in pharmaceutical industry.

Table 4. Model validation using the large-scale classification model to predict samples from different sources with different MicroNIR units

Classifier Ta-Unit7
Pb-Unit1
T-Unit7
P-Unit2
T-Unit7
P-Unit3
T-Unit7
P-Unit4
T-Unit7
P-Unit5
T-Unit7
P-Unit6
AVGc STDd
SIMCA 84.84 76.10 84.75 84.92 85.40 83.90 83.32 3.57
PLS-DA 85.71 85.89 85.00 85.71 85.71 85.69 85.62 0.31
LDA 33.02 31.97 57.75 45.87 67.78 70.57 51.16 16.86
QDA 33.10 32.13 58.33 47.78 68.41 70.73 51.75 16.91
SVM-rbf 91.67 92.08 93.42 90.32 90.79 86.67 90.83 2.30
SVM-linear 92.30 92.95 95.33 92.54 92.38 87.80 92.22 2.44
Hier-SVM-linear 94.92 93.03 97.92 95.71 95.56 92.85 95.00 1.89

a T: Training
b P: Prediction
c AVG: Average
d STD: Standard deviation

Conclusion

In this study, researchers showed the use of MicroNIR spectrometers for NIR-based pharmaceutical RMID and resolved two challenges in this area, large-scale classification and model transferability. The successful application can be attributed to the reliable instrument-to-instrument performance of MicroNIR spectrometers and robust SVM modeling with fast model building, accurate classification, excellent method transferability, and simple implementation. With rapid, reliable and nondestructive analysis capability, it is highly promising to use the portable and ultracompact MicroNIR spectrometers for on-site and in situ pharmaceutical RMID to examine each barrel in every shipment of materials used in the manufacturing of pharmaceutical drugs for the fulfillment of safety standards and quality in pharmaceutical industry.

References

1. M.R. Siddiqui, Z.A. AlOthman, N. Rahman. ‘‘Analytical Techniques in Pharmaceutical Analysis: A Review’’. Arab. J. Chem. (2013).

2. D. Sorak, L. Herberholz, S. Iwascek, S. Altinpinar, F. Pfeifer, H.W. Siesler. ‘‘New Developments and Applications of Handheld Raman, Mid-Infrared, and Near-Infrared Spectrometers’’. Appl. Spectrosc. Rev. 2011. 47(2): 83–115.

3. M.A. Druy. ‘‘Molecular Spectroscopy Workbench Applications for Mid-IR Spectroscopy in the Pharmaceutical Process Environment’’. Spectroscopy. 2004. 19(2): 60–63.

4. J. Luypaert, D.L. Massart, Y. Vander Heyden. ‘‘Near-Infrared Spectroscopy Applications in Pharmaceutical Analysis’’. Talanta. 2007. 72(3): 865–883.

5. T. Vankeirsbilck, A. Vercauteren, W. Baeyens, G. Van der Weken, F. Verpoort, G. Vergote, J.P. Remon. ‘‘Applications of Raman Spectroscopy in Pharmaceutical Analysis’’. Trends Anal. Chem. 2002. 21(12): 869–877.

6. B. Swarbrick. ‘‘Review: Advances in Instrumental Technology, Industry Guidance and Data Management Systems Enabling the Widespread Use of Near Infrared Spectroscopy in the Pharmaceutical/biopharmaceutical Sector’’. J. Near Infrared Spectrosc. 2014. 22(3): 157–168.

7. ‘‘Near-Infrared Spectrophotometry’’. United States Pharmacopoeia USP 38 NF33. 2015.

8. ‘‘Near-Infrared Spectroscopy’’. European Pharmacopoeia. 8th ed. 2013.

9. M. Blanco, M.A. Romero. ‘‘Near-Infrared Libraries in the Pharmaceutical Industry: A Solution for Identity Confirmation’’. Analyst. 2001. 126(12): 2212–2217.

10. W.B. Mroczyk, K.M. Michalski. ‘‘Application of Modern Computer Methods for Recognition of Chemical Compounds in NIRS’’. Comput. Chem. 1998. 22(1): 119–122.

11. K. Kreft, B. Kozamernik, U. Urleb. ‘‘Qualitative Determination of Polyvinylpyrrolidone Type by near-Infrared Spectrometry’’. Int. J. Pharm. 1999. 177(1): 1–6.

12. K. Kra¨mer, S. Ebel. ‘‘Application of NIR Reflectance Spectroscopy for the Identification of Pharmaceutical Excipients’’. Anal. Chim. Acta. 2000. 420(2): 155–161.

13. M. Blanco, J. Coello, H. Iturriaga, S. Maspoch, C. Pe´rez-Maseda. ‘‘Determination of Polymorphic Purity by near Infrared Spectrometry’’. Anal. Chim. Acta. 2000. 407(1–2): 247–254.

14. M. Blanco, A. Villar. ‘‘Development and Validation of a Method for the Polymorphic Analysis of Pharmaceutical Preparations Using near Infrared Spectroscopy’’. J. Pharm. Sci. 2003. 92(4): 823–830.

15. B. Grout. ‘‘Application of Near Infrared Conformance Trending for Material Quality, Container Consistency and Minimisation of Process Risk’’. J. Near Infrared Spectrosc. 2014. 22(3): 169–178.

16. M. Alcala`, M. Blanco, D. Moyano, N. Broad, N. O’Brien, D. Friedrich, F. Pfeifer, H.W. Siesler. ‘‘Qualitative and Quantitative Pharmaceutical Analysis with a Novel Handheld Miniature near-Infrared Spectrometer’’. J. Near Infrared Spectrosc. 2013. 21(6): 445–457.

17. Y. Roggo, P. Chalus, L. Maurer, C. Lema-Martinez, A. Edmond, N. Jent. ‘‘A Review of Near Infrared Spectroscopy and Chemometrics in Pharmaceutical Technologies’’. J. Pharm. Biomed. Anal. 2007. 44(3): 683–700.

18. I. Storme-Paris, H. Rebiere, M. Matoga, C. Civade, P.A. Bonnet, M.H. Tissier, P. Chaminade. ‘‘Challenging Near Infrared Spectroscopy Discriminating Ability for Counterfeit Pharmaceuticals Detection’’. Anal. Chim. Acta. 2010. 658(2): 163–174.

19. E. Dreassi, G. Ceramelli, P. Corti, S. Lonardi, P.L. Perruccio. ‘‘Near-Infrared Reflectance Spectrometry in the Determination of the Physical State of Primary Materials in Pharmaceutical Production’’. Analyst. 1995. 120(4): 1005–1008.

20. M. Andre. ‘‘Multivariate Analysis and Classification of the Chemical Quality of 7-Aminocephalosporanic Acid Using Near-Infrared Reflectance Spectroscopy’’. Anal. Chem. 2003. 75(14): 3460–3467.

21. J. Luts, F. Ojeda, R. Van de Plas, B. De Moor, S. Van Huffel, J.A.K. Suykens. ‘‘A Tutorial on Support Vector Machine-Based Methods for Classification Problems in Chemometrics’’. Anal. Chim. Acta. 2010. 665(2): 129–145.

22. O. Devos, C. Ruckebusch, A. Durand, L. Duponchel, J.-P. Huvenne. ‘‘Support Vector Machines (SVM) in Near-Infrared (NIR) Spectroscopy: Focus on Parameters Optimization and Model Interpretation’’. Chemom. Intell. Lab. Syst. 2009. 96(1): 27–33.

23. D.M. Friedrich, C.A. Hulse, M. von Gunten, E.P. Williamson, C.G. Pederson, N.A. O’Brien. ‘‘Miniature Near-Infrared Spectrometer for Point-of-Use Chemical Analysis’’. Proc. SPIE. 2014. 8992: 899203.

24. C.G. Pederson, D.M. Friedrich, C. Hsiung, M. von Gunten, N.A. O’Brien, H.-J. Ramaker, E. van Sprang, M. Dreischor. ‘‘Pocket-Size Near-Infrared Spectrometer for Narcotic Materials Identification’’. Proc. SPIE. 2014. 9101: 91010O.

25. N.A. O’Brien, C.A. Hulse, D.M. Friedrich, F.J. Van Milligen, M.K. von Gunten, F. Pfeifer, H.W. Siesler. ‘‘Miniature Near-Infrared (NIR) Spectrometer Engine for Handheld Applications’’. Proc. SPIE. 2012. 8374: 837404.

26. J.J.R. Rohwedder, C. Pasquini, P.R. Fortes, I.M. Raimundo, A. Wilk, B. Mizaikoff. ‘‘iHWG-mNIR: A Miniaturised Near-Infrared Gas Sensor Based on Substrate-Integrated Hollow Waveguides Coupled to a Micro-NIR-Spectrophotometer’’. Analyst. 2014. 139(14): 3572–3576.

27. S. Liu, H. Yi, L. Chia, D. Rajan. ‘‘Adaptive Hierarchical Multi-Class SVM Classifier for Texture-Based Image Classification’’. In: Multimedia and Expo (ICME), IEEE International Conference on. Amsterdam, the Netherlands: 6–8 July 2005. doi: 10.1109/ICME.2005.1521640.

28. C.N. Silla, A.A. Freitas. ‘‘A Survey of Hierarchical Classification across Different Application Domains’’. Data Min. Knowl. Discov. 2011. 22(1–2): 31–72.

29. R.W. Kennard, L.A. Stone. ‘‘Computer Aided Design of Experiments’’. Technometrics. 1969. 11(1): 137–148.

30. L. Eriksson, T. Byrne, J. Johansson, J. Trygg, C. Vikstro¨m. Multi- and Megavariate Data Analysis: Basic Principles and Applications. Umetrics Academy, 2013.

31. W. Wu, Y. Mallet, B.Walczak,W. Penninckx, D.L. Massart, S. Heuerding, F. Erni. ‘‘Comparison of Regularized Discriminant Analysis Linear Discriminant Analysis and Quadratic Discriminant Analysis Applied to NIR Data’’. Anal. Chim. Acta. 1996. 329(3): 257–265.

32. K. Varmuza, P. Filzmoser. Introduction to Multivariate Statistical Analysis in Chemometrics. Boca Raton, FL: CRC Press, 2009.

33. T.-F. Wu, C.-J. Lin, R.C. Weng. ‘‘Probability Estimates for Multi-Class Classification by Pairwise Coupling’’. J. Mach. Learn. Res. 2004. 5: 975–1005.

This information has been sourced, reviewed and adapted from materials provided by Viavi Solutions.

For more information on this source, please visit Viavi Solutions.

Tell Us What You Think

Do you have a review, update or anything you would like to add to this article?

Leave your feedback
Submit