Calibration sampling paradox in near infrared spectroscopy: a case study of multi-component powder blend

The objective of this study was to illustrate the sampling paradox resulting from the different strategies of spectral acquisition while preparing and implementing the calibration models for prediction of blend components in multi-component cohesive blends. A D-optimal mixture design was used to create 24 blending runs of the formulation consisting of chlorpheniramine maleate, lactose, microcrystalline cellulose and magnesium stearate. Three strategies: (a) laboratory mixing and static spectral acquisition, (b) IBC mixing and static spectral acquisition and (c) IBC mixing and dynamic spectral acquisition were investigated for obtaining the most relevant and representative calibration samples. An optical head comprising a sapphire window mounted on the lid of the IBC was used for static and dynamic NIR spectral acquisition of the powder blends. For laboratory mixed samples, powders were blended for fixed period of 30 min and later on scanned for NIR spectra. For IBC mixed blends, the spectral acquisition was carried out in-line for 2 min and stopped for static spectral acquisition. The same cycle was repeated for the next 28 min. Partial least square (PLS) calibration models for each component were built and ranked according to their calibration statistics. Optimal calibration models were selected from each strategy for each component and used for in-line prediction of blend components of three independent test runs. Although excellent statistics were obtained for the PLS models from the three strategies, significant discrepancies were observed during prediction of the independent blends in real time. Models built using IBC mixed blends and dynamic spectral acquisition resulted in the most accurate predictions for all the blend components, whereas models prepared using static spectral acquisition (laboratory mixed and IBC) showed erroneous prediction results. The prediction performance differences between the models obtained using the different strategies could be explained in the context of relevancy and representative sample collection at the initial stage of calibration model building.