Bayesian estimation of diagnostic accuracy of fecal culture and PCR-based tests for the detection of Salmonella enterica in California cull dairy cattle

Epidemiological studies of low prevalence disease problems are often hindered by the high cost of diagnostic testing. The objective of this study was to evaluate PCR screening of both individual and pooled fecal samples from culled dairy cows for the invA gene of Salmonella followed by culture to determine if the sensitivity and specificity were comparable to the results from traditional culture methods applied to individual samples. Cows from six different dairies were sampled in all four seasons. A total of 240 individual cow fecal samples, 24 fecal pools and 24 pools of 24-hour tetrathionate enrichment broth were tested. Diagnostic sensitivity of PCR screening followed by culture of PCR positive or indeterminate samples (i.e PCR-CUL method) was lower than that of culture (CUL) when applied to individual fecal samples (94.8%, 99.5%), however the specificity was comparable (99.6% and 97.7% respectively). For pools of five fecal samples and pools of five, 24 h tetrathionate broth samples, the specificity of both tests were comparable (∼98%); however, their sensitivity was only comparable in pooled fecal samples (∼93%) but greater for culture compared to PCR-CUL in pooled broth samples (∼99% versus ∼93%). Compared to culture results from testing of individual fecal samples, testing pooled fecal samples by culture had a relative sensitivity of 74% and relative specificity of 96%, testing pooled fecal samples by PCR-CUL resulted in relative sensitivity of 90% and relative specificity of 96%. Testing of pooled 24-hour enrichment broth by PCR-CUL increased the relative sensitivity and specificity to 100%. PCR testing followed by culture of positive or indeterminate samples is a time saving alternative to traditional methods. In addition, pooling of samples may be a useful method for decreasing cost if study aims can accommodate a moderate loss of relative sensitivity.


INTRODUCTION
Nontyphoidal Salmonella enterica infections are estimated to be the leading cause of foodborne illnesses in the United States resulting in over a million cases, about 20,000 hospitalizations, and more than 400 deaths annually (Scallan et al., 2011). Foods of animal origin are important sources of Salmonella infections in humans (Buncic & Sofos, 2012;Pires et al., 2009;Scallan et al., 2011). In a 2002 multi-state study in the US, Varma et al. (2006) identified consumption of undercooked ground beef as a risk factor for infection with multi-drug resistant Salmonella Newport infection. About 18% of ground beef produced in the United States are sourced from cull dairy cows (NAHMS, 1996). Beef carcass contamination with Salmonella sp. may occur during slaughter. The prevalence of Salmonella sp. in cull dairy cattle in the US had been reported to range from 0.0% to 93.0% depending on the season and day of the week that the samples were collected (Troutt et al., 2001).
Traditionally, the prevalence of Salmonella sp. has been determined using individual sample culture methods. However, these methods take several days to complete and require the culturing of a large number of samples if the prevalence of Salmonella is low. Singer et al. (2006) showed that the use of PCR on pools of five fecal samples can improve the speed and efficiency of detecting Salmonella spp. in dairy cattle feces. However, the study was not conducted on cull dairy cattle and had several limitations, including non-random selection of both cattle in the study and the individual samples included in each pool; not culturing pooled samples; and not serotyping Salmonella sp. isolates (Singer et al., 2006). Abu Aboud et al. (2016) in a study conducted on seven California dairies, estimated the crude and seasonal prevalence of Salmonella fecal shedding in cull dairy cattle based on pooled and individual fecal culture. In this current study, we estimate the prevalence of Salmonella fecal shedding in California cull dairy cattle based on the traditional culture (CUL) and a combined PCR-culture (PCR-CUL) tests using individual cow fecal samples (IF), pools of five fecal samples (FP), and pooled enrichment broths from five fecal samples after 24 h of incubation (BP). In addition, we estimated and compared the diagnostic accuracy of CUL and PCR-CUL tests for the detection of Salmonella enterica in IF, FP, and BP samples.

Fecal sampling
Individual fecal samples were collected from each of 240 cull dairy cattle from 6 dairies in the San Joaquin Valley of California between 2015 and 2016, specifically during summer (August-September), fall (October-November), winter (January-March), and Spring (April-May). The characteristics of the dairies are provided in Table 1 and the surveyweighted proportion for culling reasons are included in Table 2. Each dairy was sampled quarterly over a 12 month period to control for seasonal effects. At each sampling event,

Notes.
a Study year and seasons included summer (July 1-September 30, 2015), fall (October 1-December 31, 2015), winter (January 1-March 31, 2016) and spring (April 1-June 30, 2016). b Due to sample ID labelling error, 1 sample was not reconciled with dairy record (Herd 5, winter sample). c Totals and percents do not add up to 239 or 100%, respectively, due to multiple cull reasons. d The category labeled other reasons included the following conditions: unknown illness, gastrointestinal disorder, poor udder confirmation, undiagnosed fever, pneumonia or eye disease.
individual fecal samples were collected rectally, using clean palpation sleeves, from 10 cull cows within 24 h prior to sale from the dairy. For herds culling more than 10 cows on a given sampling date, a predetermined random number list was used to select the 10 cows for sampling. Samples were transferred to 2 oz plastic snap cap milk tubes immediately after collection and transported on ice to the Diary Epidemiology Laboratory, Veterinary Medicine Teaching and Research Center, Tulare, CA.

Pooling and enrichment
For each IF sample, approximately 8 g of feces were transferred to a sterile 15 ml polypropylene tube for individual culture. A predetermined random number list was used to select individual samples that were incorporated into each 5 sample pool and 2 g of feces from each of the 5 IF samples were transferred to a 50 ml sterile polypropylene tube (total of 10 g of feces each) and thoroughly mixed to create a FP of five samples. IF and FP samples were submitted to the CAHFS Tulare laboratory on the day of collection. Laboratory personnel were blinded to the dairy of origin and the individual samples contained in each pool. One gram of feces from either IF or FP samples were placed into a tetrathionate (TT) selective enrichment broth containing 0.01% brilliant green and 0.02% iodine at a 1:10 sample-to-broth ratio, incubated for 18-24 h at 37 ± 2 • C. Following incubation 0.75 ml of broth from each of the IF samples were combined to create broth pools (BP) from 5 individual fecal samples. The selection of individual broth samples for pooling followed the random selection performed for the fecal pools. A total of 14 cultures and 14 PCR tests were performed for each group of 10 fecal samples collected from a given dairy on a given day: Ten IF samples, two FP samples and two BP samples. The FP and the BP were comprised of the same IF samples.

Traditional culture (CUL)
An aliquot of incubated TT broth from each IF, BP and FP sample was inoculated onto two culture media (minimally-selective xylose lysine deoxycholate [XLD] plate and a highly-selective Xylose Lactose Tergitol 4 [XLT-4] plate) and incubated for 18-24 h at 37 ± 2 • C. Up to three H 2 S positive, Salmonella suspect colonies from each set of plates were subcultured separately, onto a 5% sheep blood agar-MacConkey agar bi-plate which was incubated for 18-24 h at 37 ± 2 • C. One colony from each biplate was used for biochemical testing which included triple sugar iron (TSI), urea, motility indole ornithine (MIO), citrate, O-nitrophenyl-beta-D-galactopyranoside (ONPG), and lysine iron agar (LIA) slants. Individual colonies with compatible biochemical test results (Quinn, 2011) were identified to the serogroup (Difco antiserum, BD Diagnostics, Sparks, MD) and serovar (Salmonella antisera, Statens Serum Institut, Denmark) level as described using the White-Kauffmann-Le Minor scheme (Grimont & Weill, 2007).

Salmonella PCR (PCR-CUL)
PCR testing of IF, FP and BP for the detection of Salmonella spp. was performed in a 96 well format by the CAHFS Davis bacteriology laboratory. Testing was based on the invA gene target as described previously (Suo et al., 2010)

Estimation of diagnostic sensitivity and specificity of tests
The diagnostic sensitivity and specificity of CUL and PCR-CUL for the detection of Salmonella sp. based on IF, FP, and BP were estimated using Bayesian no gold standard models as described by Hui & Walter (1980). We assumed conditional dependence between CUL and PCR-CUL because the PCR-CUL method included a culture confirmation following initial screening by PCR. Conditional dependence was modeled and estimated as conditional correlation according to (Dendukuri & Joseph, 2001;Gardner et al., 2000;Georgiadis et al., 2003). To adjust for the study's survey sampling the data was expanded based on the sampling weights (total number of cows available for sale on a given sampling day/number of cows randomly identified and sampled). External prior information for the model parameters were specified using beta distributions. A non-informative prior (dbeta(1,1)) was specified for all model parameters except for the specificity of CUL which was considered high (mode = 99%; 5th percentile = 97%) (Ekong et al., 2017). In addition, for the individual cow model, the prior for PCR-CUL sensitivity was set with a mode of 85.0% and 5th percentile of 50%, and PCR-CUL specificity was set with a mode of 95% and 5th percentile of 50% based on expert opinion. All models were fitted using three chain models, each consisting of 100,000 iterations with burn-in period of 10,000 iterations. Model convergence was assessed by visual inspection of the Gelman-Rubin diagnostic plots, deviance information criterion (DIC) and effective number of parameters estimated (pD).

Estimation of relative sensitivity and specificity of pooling
The The Bayesian analyses were conducted using OpenBUGS R version 3.2.3 rev 1012, calculation of relative sensitivity and specificity were conducted using Stata 14.2 (College Station, TX).

RESULTS
Of the 240 individual fecal samples 60 were positive for Salmonella sp. on culture (CUL). Of the 60-culture positive individual fecal samples, 54 were positive for Salmonella sp. by PCR -and follow-up culture (PCR-CUL), 4 were PCR positive, follow-up culture negative, and 2 were negative on PCR. Eighteen (18) FP were positive by both culture and by PCR-culture while 30 pools were negative on both. Nineteen (19) BP were positive on both culture and PCR-culture, 28 were negative on both and 1 was positive by culture but negative by PCR-culture (Table 3). The survey weighted results are presented in brackets in Table 3. For the pooled samples, 24/48 pools had no positive samples on IF culture or PCR-CUL and FP and BP were negative on both culture and PCR-CUL. Fifteen of 48 (15/48) pools had 1 or more IF positive on culture with the same number of positive on IF PCR-CUL and these pools had both FP and BP positive on both culture and PCR-CUL. Two of 48 pools (2/48) had 1 or 3 IF samples positive on both culture and PCR-CUL while the FP sample was negative on both culture and PCR-CUL, but the BP was positive on both culture and PCR-CUL (pools 11 and 48). Three of 48 (3/48) pools had 1 IF sample that was positive on culture but negative on PCR-CUL with all FP and BP testing being negative (pools 14, 21 and 39). Two of 48 (2/48) pools had 5 IF samples positive on culture with only 4 IF positive on PCR-CUL but all FP and BP testing was positive (pools 24, 38). One of 48 (1/48) pools had no samples positive on IF testing, FP was positive by both culture and PCR, but BP was negative on both. Lastly, one of 48 (1/48) pools had 1 IF sample that was positive on culture but negative on PCR, all FP testing was negative, but BP was positive by culture and negative by PCR. MALDI testing on 18 IF that were PCR positive or inconclusive but which did not have Salmonella sp. isolated on culture identified a variety of bacteria including Citrobacter amalonaticus (6 samples), Citrobacter koseri (1), Citrobacter freundii (2) Citrobacter spp. (5), Proteus mirabilis (4), E. coli (2), Enterobacter aerogenes, Enterobacter spp., Klebsiella pneumoniae, Klebsiella spp., and Leclercia adecarboxylata (1 each). MALDI testing on PCR positive or inconclusive but Salmonella culture negative BP (1 pool) and FP (4 pools) identified similar organisms. Salmonella sp. isolates were serogrouped and there were 4, 49, 2, 1 and 5 serogroup B, C1, C2, D1 and grouping sera negative isolates, respectively. The serogroup D1 isolate was presumed to be Salmonella Dublin and was not serotyped. Serotyping of isolates in other serogroups identified S. Typhimurium (4 isolates), S. Mbandaka (1), S. Montevideo (47), S. Rissen (1), S. Bardo (2), S. Cerro (4) and 1 untypeable isolate.
The Bayesian survey-weighted estimates of Salmonella sp. true prevalence in cull dairy cattle feces varied among the study herds ranging from 0.53% (95% PI: 0.01-2.78) to 67.3% (95% PI: 57.6-76.1) during the study period (Table 4). Furthermore, the estimated true prevalence varied by season ranging from 20.7% to 35.2% (Table 4). At the level of IF samples, diagnostic sensitivity of culture exceeded that of PCR-CUL (99.5% versus 94.8%) however the specificity was comparable (97.7% and 99.6% respectively). For FP and BP samples the specificity of both tests were comparable (∼98%), however, their sensitivity was only comparable in FP samples (∼93%) but greater for culture compared to PCR-CUL in BP samples (∼99% versus ∼93%) ( Table 5).
Pooling of fecal samples versus broth did not result in differences in sensitivity or specificity for either CUL or PCR-CUL tests with the exception of CUL's sensitivity for BP being higher than that of FP (98.6% versus 93.7%) (   (136) 39 (68) 36 (95) 18 (80) 37 (55) 11 (32) 180 (466) Fecal pool + + 0 (0) 0 (0) 3 (10) 6 (29) 2 (3) 7 (17) 18 (59) (14) 5 (12) 2 (9) 6 (9) 1 (3) 30 (75) Broth pool (28)  8 (14) 5 (11) 0 (0) 6 (9) 1 (3) 28 (65) Notes. a Actual data representing the total number of cows randomly identified and sampled at each dairy. b Data was expanded based on the sampling weights (total number of cows available for sale on a given sampling day/number of cows randomly identified and sampled) to adjust for the studys survey sampling. c +, Positive; −, Negative. d Results of enriched broth of fecal samples tested in series using PCR for Salmonella invA gene followed by culture when testing positive or inconclusive by PCR.   in Table 6. Relative sensitivity and relative specificity estimates for CUL compared to PCR-CUL were comparable with the exception of relative sensitivity for broth pools (87% vs. 83%). Relative specificity of fecal pools compared to broth pools were similar for both Notes. a Dependence model specifying uniform priors for all parameters except for the specificity of culture. b Additional priors specified for sensitivity and specificity of PCR-Culture. CUL and PCR-CUL. However, relative sensitivity for CUL of broth pools was greater than that of fecal pools.

DISCUSSION
PCR testing is a relatively modern technique that offers a significant improvement in turnaround times when applied to routine laboratory diagnostic testing. Similarly, the use of sample pooling when testing large populations with low disease or pathogen prevalence serves to improve study economics by decreasing the overall amount of testing that needs to be done to identify affected individuals (Aly et al., 2012;Murai et al., 2014). In the study reported here, we evaluated a diagnostic approach that utilized both PCR testing and sample pooling as applied to an epidemiological study of Salmonella in cull dairy cattle in California. A total of 18 IF samples were positive or inconclusive for Salmonella by PCR but an isolate could not be recovered. The inability to isolate Salmonella from these samples may have been due to low numbers of Salmonella, inactivated strains, or non-viable Salmonella; or recovery may have been compromised by other bacteria in the sample (e.g., Citrobacter sp, swarming Proteus). It is also possible that the bacteria isolated in these samples cross-reacted on the PCR to yield false positive results. The addition of the culture step following the screening PCR may have helped increase the specificity of the PCR-CUL testing.
Posterior estimates of diagnostic sensitivity and specificity for CUL and PCR-CUL methods for the detection of fecal Salmonella in adult dairy cows showed high and comparable results for IF samples, FP, and BP samples. The diagnostic specificity of CUL and PCR-CUL tests observed in the current study are comparable in magnitude to those recorded for culture and PCR methods for the detection of other foodborne fecal pathogens, Escherichia coli O157 and non-O157 serogroups (Ekong et al., 2017;Ekong et al., 2018). In the Ekong et al. (2017) and Ekong et al. (2018) studies, diagnostic sensitivity of the quantitative PCR method was greater than that of culture. In this current study, diagnostic sensitivity of CUL was slightly greater than that of PCR-CUL, however, their 95% confident intervals overlapped. Conversely, the diagnostic specificity of CUL and PCR-CUL in this study were comparable to those seen in the Ekong et al. (2017) and Ekong et al. (2018) studies. While pooling decreased sensitivity for both tests, testing FP and BP by PCR-CUL produce comparable estimates of sensitivity, pending further testing of the IF sample constituting the positive pools. Subsequently, identifying individual samples that are positive for Salmonella will be subject to the sensitivity and specificity of the individual sample test chosen. A second option which may offer cost saving is pooling fecal samples compared to broth samples. However, investigators should expect a decrease in sensitivity due to pooling of fecal compared to broth samples as estimated from current study. A higher sensitivity of BP may be explained by propagation of Salmonella in enriched broth. This decrease in sensitivity is of lesser magnitude when testing using PCR-CUL compared to CUL alone (90% vs. 74%). The estimated mean relative sensitivity of CUL for FP and BP in this study are slightly higher and had less variability as compared to the estimates reported in the study of Abu Aboud et al. (2016) which may be explained by differences over time despite being the same herds. However, these studies produced comparable estimates (mean and 95% CI) for relative specificity of CUL method.
Overall, there was high variability in the estimated prevalence of Salmonella in feces of cull dairy cows across the study herds. Lower prevalence (ranging from 0.53% to 11.1%) was recorded on four dairies, while high prevalences (57.0% and 67.3%) were recorded on 2 dairies. The lower prevalence observed on some dairies in this study is similar to estimates recorded in previous studies of dairy cattle across U.S. and in California. Wells et al. (2001) reported an overall prevalence of fecal shedding of Salmonella of 5.4% in milking cows and 18.1% in cows to be culled across US dairies and 14.9% for cull dairy cows at markets across the US. In a 2007 study, Lombard et al. (2012) reported an overall fecal Salmonella prevalence in individual cows of 14% in dairies across the U.S., with a prevalence of 3.9% in the West region of the US. In addition, Blau et al. (2005) in a 2002 study of dairy herds across 21 states of the US, reported a fecal Salmonella prevalence of 7.3% in milk cows. However, none of these previous studies were adjusted for test imperfection.