Near infrared spectroscopy for classification of bacterial pathogen strains based on spectral transforms and machine learning

The potential of near-infrared spectroscopy in classifying individual bacterial strains from different species was investigated in this study. Bacterial samples in liquid nutrient culture were collected periodically (0, 6 and 12 h) during incubation and their spectra were acquired in the near-infrared (NIR) range of 1000-2500 nm. Spectral transforms, including absorbance (A), transmittance (T) and Kubelka-Munk (KM) units were explored in order to enhance classification performance. Partial least squares discriminant analysis (PLS-DA), radial basis function neural network (RBF) and support vector machine (SVM) were used in classification model development. The results illustrated that nonlinear methods such as SVM and RBF neural network outperformed PLS-DA, where the overall correct classification rates (OCCRs) were both above 96%. Successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS) and random forest (RF) were employed to reduce spectral redundancy and to identify important wavelengths for simplifying models. The RF model yielded the best predictions as indicated by the shortest modeling time and the excellent OCCRs (100%) for both calibration and prediction. The overall results demonstrated the suitability of NIR spectroscopy with RF for the simultaneous classification of water-borne pathogenic strains from different species.