Infrared spectroscopy: instrumental factors affecting the long-term validity of chemometric models.

Data from instrumental techniques such as mid-infrared spectroscopy are increasingly being analyzed for sample identification and classification by chemometric methods based on principal component analysis (PCA). However, even modern spectrometers are subject to instability. This may affect PCA, because PCA selects the variables with the largest variance. This paper investigates the relative effects of sources of instrumental instability using a model developed for fruit puree classification. Single-beam spectra, potentially useful for on-line analysis, saw their overall intensity decrease as the infrared source output and/or the detector sensitivity declined. Consequently, single-beam spectra were mainly differentiated by their overall intensity and had to be used with caution in the long term because this strongly affected the analyses. Absorbance spectra were not sensitive to source or detector decay but showed, in the long term, subtle band shape changes and frequency shifts. While these changes were not found to influence analyses involving very different samples, they diminished the success of analyses of data sets with small intrinsic variance. Where there was large spectral differences between sample classes, instrument-related factors were insignificant. However, where spectral differences were more subtle (with a single class), instrumental effects became more important. Suggestions are given to reduce the instrumental and experimental interferences on chemometric analyses, both when recording spectra and for managing spectral databases.