Fig. 2: Classification performances for strain identification of 20 Pseudomonas aeruginosa clinical isolates using single-excitation and multi-excitation raman spectroscopy.

a Nine machine learning classifiers (GradBoost, LogReg, SVM, RF, PLS-DA, PCA-LDA, PCA-GradBoost, PCA- LogReg, PCA-SVM) were applied to each of the three Raman spectral datasets (532 nm, 785 nm, and multi-excitation) for strain identification of 20 Pseudomonas aeruginosa clinical isolates. Across all classifiers, the multi-excitation approach was found to outperform both single-excitation approaches with respect to the macro-averaged F1 score (indicated by an ‘X’). The highest performing of these, SVM, achieved a macro-averaged F1 score of 0.80, 0.81, and 0.87 for the 532 nm, 785 nm, and multi-excitation Raman excitations, respectively, and was selected among the nine classifiers for further investigation. The confusion matrices for the (b) 532 nm, (c) 785 nm, and (d) multi-excitation Raman datasets using the SVM algorithm were used to compare strain label assignments across the three approaches. Strains uniquely mislabelled by the single-excitation approaches were found to be corrected using the combined multi-excitation dataset. e The per class F1 score for each dataset was also compared, to evaluate performance and stability for both strains and excitation wavelengths. Overall, the multi-excitation approach was found to outperform or match the highest performing single-excitation approach with respect to per strain accuracy and F1 score for 11 out of the 20 investigated strains. In all other cases, the multi-excitation approach achieved the second-best performance, out of the three datasets, save for PA30 which classified poorly across all approaches.