Prediction of the particle size and flow characteristics of powder blends for tableting by near-infrared spectroscopy and chemometrics

The purpose of this research was to apply near-infrared (NIR) spectroscopy in combination with chemometrics to predict particle size and flow characteristics of a meloxicam powder blends for tableting. In order to develop calibration models for particle size (mean particle size, poly-dispersion index), and flow properties (angle of repose and time of flow) prediction, the NIR reflection spectra of different meloxicam powder blends prepared according to an experimental design were analyzed using different preprocessing methods by partial last-square (PLS) regression followed by leaveone-out cross-validation. Very good prediction ability was found for mean particle size, poly-dispersion index, angle of repose, and time of flow in models in whose development no preprocessing spectrum was applied. Also, a good prediction was found preprocessing spectrum such smoothing moving average for particle size characteristics, and unit vector normalization for powder flow properties. Therefore, NIR-chemometric methods developed in this work can be useful for the prediction of the granulometric properties and parameters related to the flowability of the meloxicam powder blends and may be used as process analytical technology (PAT) tools for process control during meloxicam tablets manufacturing.


INTRODUCTION
In the manufacturing process of tablets powder characteristics as homogeneity and free flowing are essential to ensure uniformity of mass and drug content per dose unite. Both homogeneity and flowability are influenced by particle size and particle size distribution. In this context methods for easy assessing the particle size characteristics and the flowability properties of powder blends for tableting are of great importance for the pharmaceutical industry [1,2].
NIR spectroscopy in combination with chemometrics allows rapid analysis of different characteristics of the intermediate or the finite product during the manufacturing of the tablets and it is an important toolbox in Process Analytical Technology (PAT). Direct analysis of intact solid dosage forms or pharmaceutical powder blends (as intermediate product) is considered to be an important goal for NIR analysis in the pharmaceutical industry, with the increasing needs for in-line, on-line, or at-line testing [3,4,5].
Particle size characteristics (such are: mean particle size and poly-dispersion index) and flow characteristics (such are: time of flow and angle of repose) are well-known and important parameters of powders used in the pharmaceutical industry for manufacturing solid dosage forms [6,7,8].
The NIR spectra of a solidus contain both chemical and physical information and have the potential to be used to measure one or multiple characteristics [9,10,11]. In the early applications of the near-infrared spectroscopy in pharmaceutical, they were focused on using different preprocessing methods to reduce or remove the physical interference, such as particle size or particle shape information, in order to improve the models' performance for chemical prediction [3,12,13,14]. Instead of reducing or removing the physical information from the NIR spectra, they can be extracted and can be used for the direct measurement of the particle size and flow characteristics of powder blends for tableting. A method is using chemometrics techniques to direct modeling the particle size characteristics or the flow properties with NIR spectra. This research work aimed to develop a calibration model for predicting the particle size and flow properties of powder blends for tableting in order to be used for in-line or at-line monitoring the technological process of meloxicam tablet manufacturing.

Materials
Meloxicam was provided by Uquifa, Spain; isomalt was provided by BENEO-Palatinit, Germany; microcrystalline cellulose and sodium starch glycolate was provided by JRS Pharma, Germany; and magnesium stearate was provided by UNDESA, Spain.

Preparation of powder blends for NIR calibration
The qualitative-quantitative formula of powder blends for meloxicam tablets preparation is presented in Table 1. Powder mixing: meloxicam, microcrystalline cellulose, isomalt, and sodium starch glycolate were mixed for 5 minutes in a planetary mixer (Erweka, Germany). Additionally, magnesium stearate was added, and the final powder blends were mixed for more than one minute.
Subsequently, the powder blends were sieved into three different particle size classes (Table 2). Then, using these different particle size classes samples with known composition were prepared in order to build a calibration model, according to the ratio from the DoE matrix (Table 4).  For the powder characteristics calibration, synthetic powder blends with known properties were prepared according to a design of experiment (DoE) with three factors and two levels. The input variables of the DoE were the granulometric classes and the levels were the mixing ratios. The output variables of the DoE were granulometric characteristics of the powder (particle size and poly-dispersion index) and flow properties (angle of response and time of flow) ( Table  3). The matrix of the DoE is shown in Table 4.

NIR spectroscopic analysis
NIR spectra of powder blends were registered using a NIR spectrometer (Antaris II FT-NIR Analyzer from TermoElectron, SUA). The spectra were recorded using an indium gallium arsenide (InGaAs) detector in reflectance sampling configuration, equipped with a system for sample rotation during the measurements in order to obtain representative spectra. Each reflectance spectrum was recorded by integrating 32 scans taken from 11,000 to 4,000 cm-1 at 8 cm-1 resolution.

Data processing
In order to construct the calibration methods, different spectra preprocessing methods were used to enhance the information of interest for the study and to decrease the influence of the side information contained in the spectra. The spectra preprocessing methods were smoothing, first derivative (FD). second derivative (SD), standard normal variate (SNV), unit vector normalization (UNV) [11]. The ability and the efficiency of tested calibration models were evaluated using the Root Mean Square Error of Cross Validation (RMSECV) using the following formula:

Software
Unscrambler software package from Camo X, Norway, was used to perform Partial Least Squares (PLS) regression. The calibration set (Table 4), which consists of the component mixtures in a suitable combination, was prepared according to a design of experiments built using the Modde software. Camo software allows validation of the models by full cross-validation, a procedure that consists of an iterative calibration by removing, in turn, each spectrum from the training set and then predicting the excluded sample with that calibration [2,9,14].

Particle size characteristics of powder blends
The powder blends were sieved using AS Basic sieve shaker (Retsch, Germany) equipped with a set of 4 sieves (100, 200, 300, 400 m). The mean particle size and poly-dispersion index of powder blends were determined according to a well-known method [15].

Flow properties of the powder blends
The angle of repose and the time of flow were determined according to the European Pharmacopeia methods [16].

Particle size and flow characteristics of the synthetic powder blends
In order to develop a calibration model for the particle size and the flow characteristics prediction, ten synthetic powder blends according to the experimental design matrix (table 4) were prepared. The results of particle size and flow characteristics of the prepared synthetic powder blends are presented in Table 5.
where, Y true = true properties Y pred = predicted properties n = number of training samples Modde software was used to find correlations between fractions used to obtain synthetic powder blends and particle size and flow characteristics of obtained powder blends. The results are shown in figure 1.
According to obtained results ( Figure 1) a very good correlation was found between the fraction used in the preparation of the synthetic mixtures and particle size and flow characteristics, respectively. A high amount of fraction 0-100 leads to obtaining synthetic powder blends with a small mean diameter, large poly-dispersion index, and low flow properties. On the other hand, a large amount of fraction 200-300 conducts to obtaining a synthetic powder blend with high mean diameter, small poly-dispersion index, and good flow properties (small angle of repose and small time of flow).

Spectra investigation
The development of a calibration model is an iterative activity that consists of checking different spectral preprocessing methods, as well as their combination with different spectral regions. Both the whole spectral range and specific spectral regions containing strong bands and different spectral preprocessing were evaluated in order to find good calibration models. The NIR spectra of the calibration powder blends are shown in figure 2a. In figure 2b is shown the spectra of the three fractions used in the calibration model. As shown in figure 2.b. significant differences are present especially in the range 7,000-4,000 cm-1 of the spectrums. This region was used for model calibration. calibration model according to the experimental design (b) a b

Multivariate calibration for particle size characteristics
Different multivariate calibration models were applied to find a direct correlation between the NIR spectra of ten calibration powder blends and their particle size characteristics. Cross-validation has been used to select the optimum calibration model [17,18]. The tested models were evaluated regarding the predictive abilities by plotting the actual known properties against the predicted properties and calculation the correlation coefficient (R). Also, as a diagnostic test for examining the errors in the predicted characteristics of the tested models, the RMSECV was used. The RMSECV indicates both the precision and the accuracy of the predictions. The correlation coefficients (R2) and the RMSECV obtained on the synthetic powder blends from the calibration matrix of theirs NIR reflection spectra using the PLS calibration models for prediction of the particle size are shown in Table 6. The model was selected based on the following criteria: the smallest number of factors (principal components), it has the smallest RMSECV value, the highest R2 value, and the RMSECV value for that model is not significantly greater than RMSECV for the model with one or more additional factors. The RMSECV values plotted as a function of PLS factors obtained using different spectral preprocessing methods for prediction the particle size are shown in Figure 3.
According to the RMSECV decrease and its correlation coefficient, two PLS factors were found to be optimum for using PLS methods without any pretreatment and smoothing -moving average pretreatment.

Results of control samples for particle size characteristics
In order to evaluate the prediction ability of the NIR chemometrics methods, six control samples were analyzed using the chemometric methods found above and the reference sieves method. Tables 7 and  8 have listed the results for the predictive ability and those obtained by sieves analysis on the six control samples regarding particle size characteristics (mean particle size and in for poly-dispersion index).
The data arrays of the control samples obtained using the reference sieves method and by applying the chemometric methods were compared. No statistically significant difference (P-type 1 error > 0.5) was found between the means obtained for the three arrays of predictive and reference data, considering a confidence level of 95%. Therefore, a good similarity can be considered between the obtained results using the sieves reference method and the proposed NIR chemometric methods and prove the performance of the NIR chemometric methods.

Multivariate calibration for powder flow
The correlation coefficients (R2) and the RMSECV obtained on the synthetic powder blends from the calibration matrix of theirs NIR reflection spectra using the PLS calibration models for prediction of the angle of repose are shown in Table 9. The RMSECV values plotted as a function of PLS factors obtained using different spectral preprocessing methods for predicting the angle of repose are shown in Figure 4.

FIGURE 4. Plotting the RMSECV function of the PLS factors for the angle of repose
According to the RMSECV decrease and its correlation coefficient, for the prediction of the angle of repose was found to be optimum 6 PLS factors when using PLS methods without any pretreatment and 5 PLS factors when using PLS methods with Unit Vector Normalization pretreatment.
The correlation coefficients (R2) and the RMSECV obtained on the synthetic powder blends from the calibration matrix of theirs NIR reflection spectra using the PLS calibration models for prediction of the time of flow are shown in Table 10. The RMSECV values plotted as a function of PLS factors obtained using different spectral preprocessing methods for predicting the time of flow are shown in Figure 5.
According to the RMSECV decrease and its correlation coefficient, for the prediction of the time of flow were found to be optimum 2 PLS factors when using PLS methods without any pretreatment and 2 PLS factors when using PLS methods with Unit Vector Normalization pretreatment.

Results of control samples for powder flow characteristics
In order to evaluate the prediction ability regarding the powder flow characteristics of the NIR chemometrics methods, six control samples were analyzed using the chemometric methods found above and European Pharmacopeia reference methods. Table 11 has listed the results for the   The data arrays of the control samples obtained using the European Pharmacopeia reference methods and by applying the chemometric methods were compared. No statistically significant difference (P-type 1 error > 0.5) was found between the means obtained for the three arrays of predictive and reference data, considering a confidence level of 95%. Therefore, a good similarity can be considered between the obtained results using the European Pharmacopeia reference methods and the proposed NIR chemometric methods and prove the performance of the NIR chemometric methods regarding determination the flow characteristics of powder blend for tableting (the angle of response and the time of flow).
Similar results were obtained by Porfire et all on indapamide powder blends for tableting [7]. Therefore, the developed NIR methods can predict the pharmaceutical properties of powder blends very useful during the tableting process, as are particle size and flowability, and can be exploited as PAT tools.

CONCLUSIONS
NIR spectroscopy in combination with chemometrics was evaluated to direct assay the physical characteristics of meloxicam powder blends for tableting of interest for pharmaceutical processing as particle size and flowability. Different PLS calibration models were developed and evaluated regarding the prediction of the mean particle size, the poly-dispersion index, the angle of repose, and the time of flow.
The results show a very good prediction ability of those parameters and allow to be determinate directly from NIR reflection spectra of powder blends for tableting, without any sample preparation. Such quick NIR -chemometric methods can be used for on-line, in-line, or at-line monitoring of the meloxicam tablets manufacturing process and are helpful in achieving the goals of the process analytical technology (PAT).