Fig. 3: Classification accuracies for antibiotic-sensitivity profiling of 20 Pseudomonas aeruginosa clinical isolates using single-excitation and multi-excitation Raman spectroscopy.

Nine machine learning classifiers were applied to predict the sensitivities of 20 Pseudomonas aeruginosa strains to four antibiotics: (a) ceftazidime, (b) ciprofloxacin, (c) imipenem, and (d) tobramycin, using both single-excitation (532 nm, and 785 nm), and multi-excitation Raman approaches. The highest performing classifier, with respect to the adjusted F1 score, was identified for each antibiotic-sensitivity classification task and selected for further investigation. Of the applied classifiers, SVM was found to be the highest performing model for all four antibiotics, with an adjusted F1 score of (a) 0.85, 0.81, and 0.88, (b) 0.94, 0.86, and 0.94, (c) 0.88, 0.86, and 0.92, and (d) 0.89, 0.87, and 0.93, for the 532 nm, 785 nm, and multi-excitation datasets, respectively. In all cases, the multi-excitation approach was found to outperform or match both single-excitation approaches. To further compare the difference in performances across the three excitation approaches, the per strain accuracies of each SVM model was evaluated. The heatmap displays the accuracy of the SVM model in correctly predicting the sensitivity of each strain using the three spectral datasets. Strains are grouped by their true sensitivity to the respective antibiotic to visualise class imbalance for each antibiotic-sensitivity characterisation task. A red–yellow–green colourmap is used to display overall classification where red represents an overall resistant classification, and green an overall sensitive classification.