Analysis of prediction errors in partial least squares calibration models using near infrared spectroscopy by three experimental systems

Abstract

This dissertation is focused on understanding the absorption and scattering effects of solid materials in the near infrared (NIR) spectral region and their impact on the prediction errors observed in NIR calibration models developed using partial least squares (PLS) regressions. Four different studies were performed using three experimental settings with four levels of heterogeneity of the materials. The first study consisted in the use of polypropylene films varying the number of layers stacked together which provided a system with reduced heterogeneity. NIR spectra were acquired using two experimental setups with the integrating sphere module of a Fourier transform NIR (FT-NIR) spectrometer. The depth of penetration of the radiation into the polymer layers was estimated using the O-H stretching mode related to first and second overtones of talc, which ranged from 2.95 to 3.12 mm. PLS models were developed using 30 film layers and bias values were not significantly different from zero at the 95% confidence level. Seven spectral regions were evaluated using different spectral preprocessing, the results showed that optical sampling is unbiased and there is an absence of systematic error by the NIR method. A calibration model using 50 film layers was also evaluated and it presented high statistical errors and bias due the depth of penetration of NIR radiation (optical sampling). This study highlights the lack of systematic error in the NIR method as long as the calibration is representative of the variation to be modelled by PLS regression. A second study was performed using two polymer films (polypropylene and polyethylene) with similar thickness to vary the heterogeneity of the samples and to evaluate the prediction errors observed in PLS models due to light scattering. Two FT-NIR were used to acquire the spectra of the samples. The spectra from the first instrument was used to develop the calibration models. NIR spectra from both instruments obtained on three days chosen at random order were used as prediction set to evaluate the linearity and reproducibility of the calibration model. Calibration models were developed based on polyethylene percent content varying the placement and composition below the infinite depth of the radiation. The results based on ANOVA of the predictions shown that PLS models using second derivative as preprocessing in the spectral region of 6500 – 5000 cm-1 provided low residual values with no statistical differences on both instruments. This study provides a straightforward and economic analytical method to test the linearity and reproducibility of two FT-NIR instruments using low heterogeneous polymer films. The third study was developed for real time determination of drug concentration, powder density, and porosity of powder blends at low active pharmaceutical ingredient (API) concentration (3.00 %w/w) within a feed frame. The feed frame provides the most representative stage for measurement of API before the final process. However, changes in the materials’ physical properties (e.g. powder density, particle size, flowability, and cohesivity) have a significant effect on NIR spectra. Therefore, this represents a challenge in the development of the calibration model. NIR calibration models using second derivative as spectral preprocessing explained the changes in API concentration, bulk density, and porosity of the powder blends with low error and bias values. The fourth study shows an applied case in a commercial manufacturing plant in Puerto Rico. Tablets with a combination medicine with two APIs at low concentration were analyzed by PLS regression models for real time release testing (RTRt) in a continuous manufacturing (CM) process. This study provides a better understanding of changes in the manufacturing process and their impact in the predictions of NIR calibration, furthermore, the evaluation serves for the improvement of control strategies in the manufacturing of a drug product.