Considerations when using discriminant function analysis of antimicrobial resistance profiles to identify sources of fecal contamination of surface water in Michigan.

The goals of this study were to (i) identify issues that affect the ability of discriminant function analysis (DA) of antimicrobial resistance profiles to differentiate sources of fecal contamination, (ii) test the accuracy of DA from a known-source library of fecal Escherichia coli isolates with isolates from environmental samples, and (iii) apply this DA to classify E. coli from surface water. A repeated cross-sectional study was used to collect fecal and environmental samples from Michigan livestock, wild geese, and surface water for bacterial isolation, identification, and antimicrobial susceptibility testing using disk diffusion for 12 agents chosen for their importance in treating E. coli infections or for their use as animal feed additives. Nonparametric DA was used to classify E. coli by source species individually and by groups according to antimicrobial exposure. A modified backwards model-building approach was applied to create the best decision rules for isolate differentiation with the smallest number of antimicrobial agents. Decision rules were generated from fecal isolates and applied to environmental isolates to determine the effectiveness of DA for identifying sources of contamination. Principal component analysis was applied to describe differences in resistance patterns between species groups. The average rate of correct classification by DA was improved by reducing the numbers of species classifications and antimicrobial agents. DA was able to correctly classify environmental isolates when fewer than four classifications were used. Water sample isolates were classified by livestock type. An evaluation of the performance of DA must take into consideration relative contributions of random chance and the true discriminatory power of the decision rules.