Improvement of Pharmaceutical near Infrared Calibrations by Empirical Target Distribution Optimisation

Abstract

Using process samples to develop near infrared prediction models can generate artefacts in the calibration set, potentially reducing model performance. Due to the inherent variability of manufacturing processes, the assumption that all samples from a given batch have the same concentration for the compound of interest is incorrect. A tablet press will produce compacts with different drug concentrations, distributed around a nominal value, due to mechanical effects, vibrations and powder segregation that occur in the hopper. Differences in tablet crushing strength may also be observed for those same reasons. Consequently, using such samples and their associated nominal concentrations for the development of a calibration model for the prediction of content uniformity of tablets might produce models with rather high error while able to cope with a large range of process variability. The objective of this work was to investigate empirical approaches for optimising the target (reference) concentrations of prediction samples to best match the underlying spectral variance in an attempt to mitigate interferences that may be responsible for select portions of calibration error. Three approaches were developed. The first used samples presenting a low residual from an original model to re-predict high residual samples. The second approach was an iterative search of the best target value for each sample. The third method used target values generated from a normal distribution. These approaches were compared with the classical slope and bias correction methods on their ability to predict two independent validation sets. While several methods showed significant over-fitting and a high validation error, the iterative search routine enhanced calibration performance compared to post-regression correction methods and was proven to be a viable alternative to current industry practices.