Bayesian estimation of sensitivity and specificity of culture- and PCR-based methods for the detection of six major non-O157 Escherichia coli serogroups in cattle feces

Abstract

Non-O157 Shiga toxin-producing Escherichia coli (non-O157 STEC, O26, O45, O103, O111, O121, and O145) are foodborne pathogens of public health importance. Culture and PCR-based methods have been developed for the detection of these serogroups in cattle feces. The objectives of this study were to evaluate diagnostic sensitivity and specificity of PCR- and culture-based methods for the detection of the six non-O157 serogroups, and to estimate their true prevalence in cattle feces, using a Bayesian latent class modeling approach that accounts for conditional dependence among the three methods. A total of 576 fecal samples collected from the floor of pens of finishing feedlot cattle during summer 2013 were used. Fecal samples, suspended in E. coli broth, were enriched and subjected to three detection methods: culture (involving immunomagnetic separation with serogroup specific beads and plating on a selective medium), conventional (cPCR), and multiplex quantitative PCR (mqPCR) assays. Samples were considered serogroup positive if the sample or the recovered isolate tested positive by PCR for an O gene of interest; neither Shiga toxin (stx) nor intimin (eae) genes were assessed. Prior information on the performance of the three methods was elicited from three subject experts. Culture was generally the least sensitive and most specific of the 3 tests across serogroups, mqPCR was generally the most sensitive test and cPCR more specific than mqPCR. Sensitivity analysis indicated that posterior inferences on test performance and prevalence were susceptible to prior specification in cases where few or no detections present in the data for selected combinations of diagnostic methods (i.e. extreme category problem). Our results characterize performance of detection methods and true prevalence of non-O157 serogroups, thus informing necessary adjustments for test bias in risk modeling.