Chemometric Analysis of Diffuse Reflectance-Absorbance Fourier Transform Infrared Spectra Using Rule Induction Methods: Application to the Classification of Eubacterium Species

Abstract

Strains representative of four Eubacterium species were analyzed by using diffuse reflectance-absorbance Fourier transform infrared spectroscopy. To identify important wavenumber regions for the classification of these 22 bacterial isolates, we investigated three rule induction methods and various spectral preprocessing regimes. In this study both univariate and multivariate classification and regression trees (CART) methods and the fuzzy multivariate rule-building expert system (FuRES) method were exploited. It was found that the FuRES method was superior in terms of prediction, whereas the rules proposed by the univariate CART method were easier to interpret in terms of which wavenumbers in the IR spectra were important for bacterial class separation. Scaled and detrended FT-IR spectra and first-order numerical differentiation preprocessing steps were necessary to obtain optimal classification models. Finally, a reduction in the classification error for the CART-based methods was observed by analyzing the compressed B-spline coefficients rather than the original spectra representation. The spectral interpretation of these rules is in agreement with analyses using the uncompressed representation.