Discrimination of hazardous bacteria with combination laser-induced breakdown spectroscopy and statistical methods

Visualization of laser-induced breakdown spectroscopy data of mouse organs based on the feature extraction method

At present, there is no comprehensive and systematic research on laser-induced breakdown spectroscopy (LIBS) data visualization. In particular, the LIBS spectra of biological samples have large noise and weak signals, which seriously affect the feature visualization. Here, three commonly used sample visualization methods were compared, and a new method was applied for tissue sample visualization. We used the LIBS mapping technique to obtain LIBS spectra of different organ slice samples from mice. LIBS spectral distribution was visualized after extracting the region of interest. The three spectral visualization methods are compared, and the performance of visualization algorithms is quantitatively analyzed. The potential of heat-diffusion for the affinity-based transition embedding (PHATE) method highlights the details of the LIBS spectral distribution while maintaining the overall structure of the data. The correlation coefficient between dimensionality reduction data and raw data is 0.97, and the average distance between samples of different categories is 0.64, which are superior to those of traditional principal component analysis (PCA), multidimensional scaling (MDS), and t-distributed stochastic neighbor embedding (t-SNE). The results show that the PHATE method can serve as a very promising LIBS spectral visualization tool.

Study on direct identification of bacteria by laser-induced breakdown spectroscopy

Bacteria are everywhere in the natural environment. Although most of them are harmless, there are still some hazardous bacteria that will harm human health, so it is particularly important to identify bacteria quickly. Compared with traditional time-consuming and complicated identification methods, laser-induced breakdown spectroscopy (LIBS) is one of the potential technologies for rapid identification of bacteria. In this paper, six weakly active bacteria, including Escherichia coli, Enterococcus faecalis, Bacillus megaterium, Bacillus thuringiensis, Pseudomonas aeruginosa and Bacillus subtilis, are taken as analysis samples. The thawed bacteria are placed in deionized water, and then uniformly smeared on five kinds of substrates to verify the feasibility of using LIBS to identify these bacteria. Spectrum filtering, normalization and principal component analysis (PCA) are used to preprocess the spectra, and a multi-class identification method based on the one-against-all linear kernel function of support vector machine (SVM) is proposed to establish the prediction model. The identification performance is evaluated by using precision and recall. The experimental results show that high-purity graphite is the best substrate with the least interference to the LIBS spectrum of bacteria. The prediction precision of these six bacteria is 77.27%, 92.86%, 84.21%, 94.12%, 81.82% and 76.92%, respectively, recall is 85%, 100%, 94.12%, 80%, 81.82% and 75% respectively, and the identification rate is 84.17%. It can be seen that the direct identification of bacteria can be preliminarily realized by smearing bacteria on the graphite substrate and analyzing its LIBS spectra, which provides a feasible way for simple, rapid and on-site bacterial identification.

Construction of classification models for pathogenic bacteria based on LIBS combined with different machine learning algorithms

Bacteria, especially foodborne pathogens, seriously threaten human life and health. Rapid discrimination techniques for foodborne pathogens are still urgently needed. At present, laser-induced breakdown spectroscopy (LIBS), combined with machine learning algorithms, is seen as fast recognition technology for pathogenic bacteria. However, there is still a lack of research on evaluating the differences between different bacterial classification models. In this work, five species of foodborne pathogens were analyzed via LIBS; then, the preprocessing effect of five filtering methods was compared to improve accuracy. The preprocessed spectral data were further analyzed with a support vector machine (SVM), a backpropagation neural network (BP), and k-nearest neighbor (KNN). Upon comparing the capacity of the three algorithms to classify pathogenic bacteria, the most suitable one was selected. The signal-to-noise ratio and mean square error of the spectral data after applying a Savitzky-Golay filter reached 17.4540 and 0.0020, respectively. The SVM algorithm, BP algorithm, and KNN algorithm attained the highest classification accuracy for pathogenic bacteria, reaching 98%, 97%, and 96%, respectively. The results indicate that, with the support of a machine learning algorithm, LIBS technology demonstrates superior performance, and the combination of the two is expected to be a powerful tool for pathogen classification.

Predictive data clustering of laser-induced breakdown spectroscopy for brain tumor analysis

Limited by the lack of training spectral data in different kinds of tissues, the diagnostic accuracy of laser-induced breakdown spectroscopy (LIBS) is hard to reach the desired level with normal supervised learning identification methods. In this paper, we proposed to apply the predictive data clustering methods with supervised learning methods together to identify tissue information accurately. The meanshift clustering method is introduced to compare with three other clustering methods which have been used in LIBS field. We proposed the cluster precision (CP) score as a new criterion to work with Calinski-Harabasz (CH) score together for the evaluation of the clustering effect. The influences of principal component analysis (PCA) on all four kinds of clustering methods are also analyzed. PCA-meanshift shows the best clustering effect based on the comprehensive evaluation combined CH and CP scores. Based on the spatial location and feature similarity information provided by the predictive clustering, the PCA-Meanshift can improve diagnosis accuracy from less than 95% to 100% for all classifiers including support vector machine (SVM), k nearest neighbor (k-NN), soft independent modeling of class analogy (Simca) and random forests (RF) models.