FIGURE 2.

Confusion matrices showcasing agreement between susceptible-intermediate-resistant (SIR) classification of 128 Salmonella isolates obtained using phenotypic resistance testing (denoted as the matrix “Reference”) and five in silico methods (denoted as the matrix “Prediction”) for 15 antimicrobials. Accuracy values above each matrix denote the percentage of correctly classified instances out of all instances. Kappa values above each matrix denote Cohen’s kappa coefficient for the matrix, reported as a percent. For the phenotypic method, SIR classification was determined using NARMS breakpoints for Salmonella (accessed March 23, 2020). For the in silico antimicrobial resistance (AMR) determinant detection approaches, combinations of five pipelines and one to five AMR determinant databases were tested; isolate genomes that harbored one or more AMR determinants previously known to confer resistance to a particular antimicrobial were categorized as resistant to that antimicrobial (“R”), while those which did not were categorized as susceptible (“S”; see Supplementary Table 2 for all detected AMR determinants and their associated resistance classifications). For all AMR determinant detection methods, isolates that showed intermediate phenotypic resistance to an antimicrobial were categorized as susceptible (“S”) rather than resistant, as this classification produced slightly better accuracy scores for all pipeline/database combinations. For AMR determinant detection methods that relied on nucleotide BLAST (i.e., ABRicate and BTyper), the confusion matrix obtained using the nucleotide identity and coverage threshold combination that produced the highest accuracy are shown (Supplementary Figure 3); for all other methods, confusion matrices obtained using default detection parameters are shown.