NEW YORK — Mass spectrometry can be used to gauge whether clinical bacterial samples may be resistant to certain antimicrobial treatments and possibly guide treatment choices, a new study has found.
Current culture-based methods to determine whether a bacterial sample is antibiotic resistant can take up to 72 hours, leading to a time frame during which patients may be treated with too narrow or too broad an antibiotic.
To shorten the time it takes to determine antimicrobial resistance, researchers from the University of Basel and elsewhere in Switzerland developed a database of bacterial and fungal matrix-assisted laser desorption/ionization-time of flight, or MALDI-TOF, mass spectra profiles and their associated antibiotic resistance profiles. As they reported in Nature Medicine on Monday, they used their database to develop classifiers to predict the antimicrobial resistance of different pathogens with high accuracy. A retrospective clinical case study further suggested that the approach would have changed clinical treatment in about 14 percent of cases.
"These findings exemplify the potential of classifiers to optimize antibiotic treatment and assist antibiotic stewardship efforts using real clinical cases," ETH Zurich's Karsten Borgwardt, Basel's Adrian Egli, and colleagues wrote in their paper.
The researchers developed their database, which they dubbed the Database of Resistance Information on Antimicrobials and MALDI-TOF Mass Spectra, or DRIAMS, by collecting MALDI-TOF mass spectra and resistance information on more than 30,000 clinical isolates from four different Swiss clinical labs. The largest collection within DRIAMS, called DRIAMS-A came from University Hospital Basel and included 145,341 mass spectra. Most of the spectra could be generated from clinical samples within 24 hours.
The researchers divided DRIAMS-A — which they used for most of their analyses — into training and testing datasets and applied three different machine learning approaches to predict antimicrobial resistance. They in particular focused their analyses on three key clinical pathogens, Staphylococcus aureus, Escherichia coli, and Klebsiella pneumoniae, and antibiotics used to treat infections they cause.
For all three, the researchers reported a high overall performance. The classifier could predict S. aureus resistance to oxacillin with 80 percent accuracy as well as E. coli and K. pneumoniae resistance to ceftriaxone currently with 74 percent accuracy for both. For 31 of the 42 antibiotics studied, the classifier generated could correctly classify resistance with 80 percent accuracy.
The researchers noted, however, some drawbacks to their approach. The performance of classifiers trained at one site was not transferable to mass spectra generated at other sites, likely due to differences in phylogenetic strains present in each region, but also due to technical variability. Similarly, the predictive power of the classifier was better for samples collected close in time to those used in the training set, likely due to similar changes over time. This suggested to the researchers that any classifier applied clinically should be retrained regularly with newly generated local data.
Still, the classifier could influence treatments used clinically. In a retrospective clinical case study, the researchers reviewed 416 cases with positive S. aureus, E. coli, and K. pneumoniae cultures. For 63 of these, an infectious disease specialist was consulted to help guide antibiotic treatment. The researchers applied this approach to that subset of cases to see whether their classifier would have suggested an alternative course of treatment. For most, 54 cases, the algorithm would not have suggested a change, but for nine, it would have. That change, for seven of the cases, would have been a de-escalation of the antibiotic therapy.
"Our retrospective clinical case study shows that our classifier might have a beneficial impact on patient treatment and promote antibiotic stewardship," the researchers wrote in their paper.