Bayesian estimation of true prevalence, sensitivity and specificity of three diagnostic tests for detection of Escherichia coli O157 in cattle feces

Abstract

Cattle are a reservoir for Escherichia coli O157 and they shed the pathogen in their feces. Fecal contaminants on the hides can be transferred onto carcasses during processing at slaughter plants, thereby serving as a source of foodborne infection in humans. The detection of E. coli O157 in cattle feces is based on culture, immunological, and molecular methods We evaluated the diagnostic sensitivity and specificity of one culture- and two PCR-based tests for the detection of E. coli O157 in cattle feces, and its true prevalence using a Bayesian implementation of latent class models. A total of 576 fecal samples were collected from the floor of pens of finishing feedlot cattle in the central United States during summer 2013. Samples were enriched and subjected to detection of E. coli O157 by culture (immunomagnetic separation, plating on a selective medium, latex agglutination, and indole testing), conventional PCR (cPCR), and multiplex quantitative PCR (mqPCR). The statistical models assumed conditional dependence of the PCR tests and high specificity for culture (mode = 99%; 5th percentile = 97%). Prior estimates of test parameters were elicited from three experts. Estimated posterior sensitivity (posterior median and 95% highest posterior density intervals) of culture, cPCR, and mqPCR was 49.1% (44.8–53.4%), 59.7% (55.3–63.9%), and 97.3% (95.1–99.0%), respectively. Estimated posterior specificity of culture, cPCR, and mqPCR were 98.7% (96.8–99.8%), 94.1% (87.4–99.1%), and 94.8% (84.1–99.9%), respectively. True prevalence was estimated at 91.3% (88.1–94.2%). There was evidence of a weak conditional dependence between cPCR and mqPCR amongst test positive samples, but no evidence of conditional dependence amongst test negative samples. Sensitivity analyses showed that overall our posterior inference was rather robust to the choice of priors, except for inference on specificity of mqPCR, which was estimated with considerable uncertainty. Our study evaluates performance of three diagnostic tests for detection of E. coli O157 in feces of feedlot cattle which is important for quantifying true fecal prevalence and adjusting for test error in risk modeling.

Introduction

Shiga toxin-producing Escherichia coli O157 is a foodborne pathogen of significant impact on beef industry and public health (Callaway et al., 2013). It can cause mild to bloody diarrhea in humans which may progress to hemolytic uremic syndrome (Centers for Disease Control and Prevention CDC, 1993, Hussein, 2007) that can be fatal in children, the elderly and immuno-compromised individuals. In the United States, E. coli O157 is responsible for an estimated 63,153 illnesses, 2138 hospitalizations and 20 deaths annually (Scallan et al., 2011). Cattle are a major reservoir of this pathogen (Gyles, 2007). E. coli O157 colonizes the hindgut of cattle and is shed in their feces, which can contaminate cattle hides and the environment (Jacob et al., 2010). Subsequently, hides may serve as main source of contamination of carcasses and beef products during processing (Arthur et al., 2008).

Laboratory methods used for the detection of E. coli O157 in cattle feces, on hides, and in beef products include traditional culture methods, immunological, and molecular based diagnostic methods (Barkocy-Gallagher et al., 2002, Deisingh et al., 2004). The culture-based detection method exploits the specific biochemical characteristics of the pathogen for identification. It involves sample enrichment in a selective broth, followed by immunomagnetic bead separation, plating on selective agar medium, and confirmation of isolates by biochemical tests, agglutination, and/or PCR detection of serotype-specific and virulence genes (Elder et al., 2000, Bai et al., 2010, Jacob et al., 2014). Immunological-based detection methods use monoclonal or polyclonal antibodies to capture the pathogen specific antigen. Examples include ELISA-based assays (Thompson et al., 2007, Shen et al., 2014), microplate enzyme immunoassays (Hoefer et al., 2011), and fluorescence polarization assays (Nielsen et al., 2007). Molecular-based detection methods, which amplify serotype-specific DNA targets, include conventional PCR (Arthur et al., 2005, Islam et al., 2006, Bai et al., 2012, Jacob et al., 2014), real-time/quantitative PCR (Stefan et al., 2007, Jacob et al., 2014, Luedtke et al., 2014, Verstraete et al., 2014, Noll et al., 2015), and loop mediated isothermal amplification (Dong et al., 2014). However, none of these methods constitute gold standards for the detection of E. coli O157 in cattle feces, hides, or beef products. Furthermore, quantitative assessment of the diagnostic performance (sensitivity and specificity) of these methods is lacking.

Estimating the sensitivity and specificity of the detection methods is crucial for the evaluation of the true E. coli O157 prevalence and the true public health risk. In the United States, estimated apparent prevalence of E. coli O157 depends on cattle type and season, and based on a recent systematic review and meta-analysis study, it ranges from 0.0% to 55.0% in cattle feces (Ekong et al., 2015). On hides, estimates of apparent prevalence ranged from 0.0% to 100.0%; on beef carcasses during slaughter, apparent prevalence estimates ranged from 0.0% to 93.3% in pre-eviscerated carcasses, and from 0.0% to 2.0% in chilled carcasses (Ekong et al., 2015). It is possible that due to imperfect sensitivity and specificity of the diagnostic tests currently available, apparent prevalence estimates reported may misrepresent the true prevalence of E. coli O157 due to misclassification of some of the tested samples. Therefore, there is need to assess the diagnostic performance of current detection methods using fecal samples from naturally shedding commercial cattle.

In the absence of a gold/reference standard, latent class models implemented through Bayesian estimation procedures have been successfully used to estimate true disease prevalence and diagnostic performance for two or more diagnostic tests applied to samples from test subjects of unknown underlying disease status (Berkvens et al., 2006, Fablet et al., 2010, Haley et al., 2011, Paradis et al., 2012, Paul et al., 2013). Published standards for reporting diagnostic accuracy using Bayesian Latent Class Models were followed (Kostoulas et al., 2017, Appendix A in supplementary file). The objective of this study was to evaluate the diagnostic performance (diagnostic sensitivity and specificity) of culture-based, conventional PCR, and multiplex quantitative PCR tests used for the detection of E. coli O157 in cattle feces, and to estimate the true prevalence of E. coli O157 in feces of naturally shedding central US feedlot cattle, using latent class models in a Bayesian approach.