An epidemiological framework for improving the accuracy of whole-genome sequence-based antimicrobial resistance surveillance in <i>Salmonella</i>

Whole-genome sequence-based surveillance of bacteria for determinants of antimicrobial resistance (AMR) promises many advantages over traditional, wet-lab approaches. However, adjustments to parameters used to identify genetic determinants from sequencing data can affect results and interpretation of the important determinants in circulation. Using a dataset of whole-genome sequences from 1633 isolates of <i>Salmonella</i> Heidelberg and <i>S.</i> Kentucky collected from surveillance of Canadian poultry production, we queried the genomic data using an in silico AMR detection tool, StarAMR, applying a range of parameter values required for the detection pipeline to test for differences in detection accuracy. We compared the results from each iteration to phenotypic antimicrobial susceptibility results, and generated estimates of sensitivity and specificity using regression models that controlled for the effects of multiple sampling events and variables, and interactions between covariates. Results from our analyses revealed small, yet significant effects of the input parameters on the sensitivity and specificity of the AMR detection tool, and these effects differed based on the serovar and drug class in question. Findings from this study may have implications for the incorporation of whole-genome sequence-based approaches to the surveillance of AMR determinants in bacteria sampled from food products and animals related to food production.