New developments in multivariate calibration

This work reviews both current methods and new developments in treating spectral interferences in multivariate calibration. Essentially all pretreatment techniques fall into six categories. Spectral transformations used for the linearization of spectral data. Variety of techniques such as offset and slope corrections, derivatives, baseline corrections, row-wise normalizations, wavelets etc. These techniques are essentially targeted for spectral intensity variations etc. Techniques handling specific interferences such as optical path length differences (scattering etc), known spectral interferences and external (experimental or process) parameters. Techniques removing spectral variation that is not correlated with response data. Ensemble methods estimating spectral interferences and using data augmentation in order to get robust multivariate models. Selectivity enhancements using either net analyte signal (NAS) or science-based calibration (SBC). In overall, current methods and practices provide a very different means to treat spectral data. We would like to put emphasizes on guaranteed selectivity using SBC which potentially delivers clear opportunities over traditional chemometrics. The key question is how to ensure selectivity and treat spectral interferences efficiently at the same time. For example, most of the techniques do not guarantee selectivity as they are influenced by spurious and unspecific correlations when it comes to PLS or equivalent modeling following the pretreatment step. The NAS method guarantees selectivity but does not properly account for interferences. It operates like an ordinary CLS method. Selectivity is sacrificed at the expense of sensitivity. It is crucial for industry to become aware that selectivity is not guaranteed. The current practice in industry is that predictive ability is over-appreciated as pretreatments are often selected based on what improves model and its predictive ability. There is nothing bad in this approach if models are validated properly. Unfortunately, we can hardly ever validate models against all unknown spectral interferences in the process. In this work we focus on SBC as it provides a means to handle spectral noise and to guarantee selectivity. We try to show that there is an alternative and cost effective approach to develop robust, sensitive and selective models and methods. Such measures are especially important in pharmaceutical industry where PAT developments are in focus. The method described here is directly applicable to process monitoring, method development, quality control and calibration transfer. The method is also directly applicable to all spectroscopic instruments whether UV, visible, NIR, IR or Raman. Simple demonstration on a multi-component system is provided as an example on the results achievable with new SBC calibration method including cross-sensitivities and selectivity measures