Photonic Data Science

Increasing antimicrobial resistance (AMR) has developed into an enormous health burden. Here, a systematic investigation was conducted to evaluate the discriminative performance of Raman spectroscopy between different resistance classes (Susceptible, ESBL, CRE, VRE, VSE) in common clinical isolates (Escherichia coli, Klebsiella pneumoniae, Klebsiella oxytoca, Citrobacter freundii, Acinetobacter baumanii, Enterococcus faecium). Two different Raman spectroscopic methods (UVRR in bulk and 785 nm excitation directly on the Petri dish) and four different machine learning algorithms (PCA-LDA, PLS-DA, PCA-SVM, PCA-RF) were tested aiming the application of a decision-tree using a 3-step approach composing of species classification, differentiation of susceptible from resistant strains within the species and differentiation of ESBL and CRE as AMR subclasses within the class of antibiotic-resistant strains. In species classification, the two Raman methods yield similar results in all applied models. When attempting the differentiation of susceptible vs resistant strains in the intraspecies level, 785 nm overall outperformed UVRR and PCA-SVM and PLS-DA provided higher discriminative power compared to PCA-LDA and PCA-RF. For the discrimination of ESBL vs CRE isolates UVRR was not suitable as a method and 785 nm excitation provided correct identification of all 9 strains when using PCA-SVM and PLS-DA, confirming stability over replicate-to-replicate variations. Raman spectra from 785 nm excitation directly on the Petri dish combined with PCASVM and PLS-DA are suitable for diagnostic application of Raman spectroscopy in hospital settings. These results are the first step of a long journey in the development of Raman spectroscopy for microbiological documentation and extraction of AMR-related information in infectious diseases.