Process monitoring in pharmaceutical applications using near-infrared spectroscopy and simple modelling techniques.

摘要

The overuse of sophisticated spectral data processing has been blamed for slow acceptance of near-infrared spectroscopy (NIRS) into mainstream pharmaceutical analysis. In this research, NIRS was investigated for application to pharmaceutical process monitoring using simple data modelling techniques. Firstly, NIRS was employed to monitor the functional group conversion of molecules bound to stationary supports. Using second derivative spectra it was possible to follow an esterification reaction in real-time despite concerns of NIR sensitivity to low quantities of material. No sample preparation was required for this novel application and measurements were made non-invasively in the presence of solvents and reagents. Crude assessments of the reaction kinetics were also made by following changes in absorbance as a function of time. As an example of quality control in late product development, a variety of data processing algorithms were compared for the determination of the exact water content in lactose monohydrate samples. The change from bound to bound and free water over the required water content range caused complications when modelling the NIR spectral data. Many good models were developed which could be used in place of the more laborious reference method (Karl Fischer titration). A novel, less commonly used semi- quantitative algorithm, Polar Qualification System, was also investigated and was successfully used to qualify lactose samples. Finally, in the production environment, on-line monitoring of a drug content was performed on a marketed dry powder inhalation product. Investigations were made to determine the effective sample size measured by a NIR fibre optic probe, so that comparisons could be made with conventional methods of measuring blend homogeneity. Calibration models were developed off-line using multiple linear regression with up to three wavelengths. On-line data were then fitted to these models, however, the quality of the predictions were poorer than desired, but nevertheless provided useful information.