A bootstrap-based strategy for spectral interval selection in PLS regression

Bootstrap-based methods have been applied for spectral variable selection in near (NIR) and mid-infrared (MIR) spectroscopy applications. In this paper, an extension of those methods for the selection of spectral intervals instead of single spectral variables is proposed. This approach, interval partial least square (PLS)-Bootstrap (iPLS-Bootstrap), was compared against the PLS-Bootstrap method and the use of the whole spectral region for model development. These methods were tested on a NIR spectral dataset obtained from at-line monitoring of an industrial fermentation process, by correlating the spectra with the concentration of the active pharmaceutical ingredient (API). The performance of the models was evaluated based on the predictive ability for both cross-validation and external validation. For the dataset used, iPLS-Bootstrap enabled to improve the model predictive ability, with a greater impact on external validation. The decrease observed in RMSEP relative to the full-spectrum and PLS-Bootstrap model was, respectively, 14 and 6%. Copyright (C) 2008 John Wiley & Sons, Ltd.