Uncertainty assessment in FT-IR spectroscopy based bacteria classification models

The number of Fourier Transform infrared spectroscopy (FT-IR) applications for discrimination between microorganisms at different levels has increased substantially over recent years. Appropriate spectra handling and processing algorithms is a requirement for successful method implementation. Chemometrics methods such as Principal Component Analysis (PCA), discriminant Partial Least-Squares Regression (PLSDA), and Soft Independent Modelling of Class Analogy (SIMCA) have found particular relevance in this context. Despite its applicability, robustness of multivariate calibration is still a key issue for effective application of FT-IR spectroscopy for microorganisms classification. Resampling methods provide a very interesting alternative to estimate dispersion measurements in the context of PCA, PLSDA and SIMCA modelling. This work focuses on the comparison of PCA, PLSDA and SIMCA approaches for bacteria discrimination based on FT-IR spectra. Bias and uncertainty for all implementations in terms of bacteria classification prediction are assessed using two non-parametric resampling methods — jackknife and bootstrap. A total of 73 samples of Acinetobacter baumanni, Enterococcus faecium and Staphylococcus aureus/epidermidis were used for calibration, and 32 samples from an independent samples group were used for model testing. Resampling strategies were applied to each method to provide a dispersion measure of the classifications for the testing set. Employed bootstrapping and jackknifing methods demonstrated to be valid alternatives to estimate bias and variance for non-supervised and supervised microorganisms discrimination models.