Development of Near-IR Spectroscopic and Imaging Methods for Process Control of Pharmaceutical Powder Blending

The Process Analytical Technology (PAT) initiative, undertaken by the Food and Drug Administration (FDA), paves the way for improvement of drug manufacturing through real-time measurements that allow better process understanding. In this work, Near-Infrared spectroscopy (NIRS) was used to monitor powder blending through optical ports mounted on the blender. Preliminary studies demonstrated that active ingredient concentration and processing conditions affected the mixing end point. The advantages and limitations of different chemometric algorithms in blend uniformity analysis were evaluated, with an ultimate goal of developing a global model that could be used for prediction of future samples. The limitation of Curve-Fitting to Mean Standard Deviation-Time Data technique as a sole measure of homogeneity was apparent because it measures spectral variability without being sensitive to potency changes. In contrast, pattern recognition approaches, such as Soft Independent Modeling of Class Analogies (SIMCA), were more reliable owing to their sensitivity to %potency and variability changes. The NIR-analyzed sample mass was estimated to be between 5.5 and 6.4 mg, highlighting the importance of monitoring blends from multiple locations to ensure representative sampling. An experimental design approach was used to characterize the effect of humidity, component concentration, and blender speed on mixing end point. All three variables were shown to significantly impact the blending process. Furthermore, humidity and concentration had a significant effect on particle size and density of powder mixtures. Qualitative algorithms such as SIMCA and Bootstrap Error-adjusted Single-sample Techniques (BEST) were evaluated. Optimization of NIR models was achieved by spectral processing, and training set sample selection. The models developed were successful in predicting blend homogeneity of independent blend samples. A quantitative NIR model for blending end point prediction was also developed. Process signature was built into NIR models by using spectra of actual blend experiments. Evaluation of principal component regression (PCR), partial least squares (PLS) and multi-term linear regression (MLR) showed that a single wavelength-linear regression yielded optimum results. The blending profiles predicted by the NIR quantitative model correlated well to those determined by the UV method. Characterization of intra-shell versus inter-shell powder mixing kinetics and its implication in sensor positioning was also performed