Towards improvement in classification of Escherichia coli, Listeria innocua and their strains in isolated systems based on chemometric analysis of visible and near-infrared spectroscopic data

Abstract

This study investigated the classification of Escherichia coli and Listeria innocua at species and strain levels using transflectance near infrared (NIR) spectroscopy together with various chemometric methods. NIR spectra were collected from a series of dilutions of bacterial suspensions in phosphate buffered saline. Different spectral pre-processing methods were applied to the raw spectra during model calibration. Partial least squares discriminant analysis (PLS-DA) was used to develop calibration models while the least squares support vector machine (LS-SVM) technique was employed to improve difficult classifications. Besides calibration models based on all wavelengths, competitive adaptive reweighted sampling (CARS) was implemented for the first time to select some important wavelengths for establishing simplified models in order to classify bacterial strains. Results indicated that, when LS-SVM and CARS were used, the overall correct classification rates (OCCRs) and model simplicity were generally greatly improved over results obtained by PLS-DA. For classification of E. coli and L. innocua at species level, 100% of samples were correctly classified using only three wavelengths (1884, 1886 and 1890nm). For E. coli strain identification, use of CARS and LS-SVM produced an OCCR of as high as 85.2% for prediction while PLS-DA using all wavelengths could only attain an OCCR of 48.2% for the same task. Classification of L. innocua strains was also substantially improved using the same strategy and the highest OCCR achieved was 66.7%. This study demonstrated that CARS and LS-SVM were useful tools for enhancing classification of bacteria.