Feasibility study of NIR for surveillance of a pharmaceutical process, including a study of different preprocessing techniques

Multivariate calibration models for predicting the initial reactant and an impurity in a synthesis of contrast media have been developed using near‐infrared (NIR) spectroscopy. The reaction is presently monitored off‐line by HPLC. The calibration models are based on spectra measured on‐line. The process solution is chemically complex and the intention of this study was to enhance knowledge about the predictive ability of the models. Different data‐preprocessing and variable selection methods reported in the literature have been tested in order to enhance chemical information and reduce irrelevant variability due to physical influences on measured spectra such as scattering and temperature. In addition to standard preprocessing methods such as normalization, differentiation and multiplicative scatter correction, orthogonal signal correction and optimized scaling were tested. The modelling was performed by partial least squares (PLS), except for optimized scaling which uses principal component regression (PCR). Optimized scaling on differentiated data gave the lowest prediction error. The NIR region from 1100 to 1900 nm gave better models than variable selection and fewer variables. Models with correlation coefficients for best fit equal to 0.998 and 0.961 were obtained for the initial reactant and the impurity, respectively.