INTRODUCTION
Campylobacter spp. is the most common bacterial cause of human gastrointestinal disease in most developed countries [Reference Olson, Nachamkin, Szymanski and Blaser1]. C. jejuni is the most common species in human campylobacteriosis followed by C. coli [Reference Gillespie2, Reference Tam3]. Both species are frequently found in the alimentary tracts of a wide range of animals [Reference Newell and Fearnley4] with C. jejuni being most associated with the contamination of poultry flocks and poultry products, while C. coli is found predominantly in pigs [Reference Miller5, Reference Thakur6].
Poultry and poultry products remain one of the most important sources of human campylobacteriosis. A baseline survey carried out at European Union (EU) level, found the prevalence of Campylobacter-colonized broiler batches was 71·2% overall but this varied greatly between Member States (MS) from 2% to 100% [7]. C. jejuni and C. coli were found in 60·8% and 41·5%, respectively, of positive batches but the species distribution was highly variable across the EU. C. jejuni was the most common in 19 MS while in seven MS the frequency of C. coli ranged from 76·1% and 97·7% of the species identified. The reason for these large differences in the species prevalence in broiler flocks is unknown. However, microbiological methods based on culture and biochemical identification do not provide a true measure of the prevalence of the different species present in a sample as both the selective media used and the incubation atmosphere may have an impact on the species recovered [Reference Moran, Kelly and Madden8–Reference Williams11], and at best provide an approximation of the prevalence of these species in broiler flocks.
Polymerase chain reaction (PCR)-based methods have been applied for the detection of Campylobacter spp. directly from caecal contents, and some methods are able to identify both C. jejuni and C. coli in samples with greater sensitivity than conventional culture methods [Reference Randall12]. The accurate estimation of the true prevalence of these two species in broiler flocks may allow a better understanding of their epidemiology in such populations and assessment of their relative importance for human infection.
During 2007–2009 a UK-wide, 3-year survey of broiler flocks was conducted to estimate the prevalence of Campylobacter-infected batches of birds at slaughter [Reference Lawes13] by direct culture of caecal samples. A prerequisite for the determination of the actual prevalence of each Campylobacter species and hence their epidemiology is knowledge of the sensitivity and specificity of the culture test for these organisms. These performance criteria should be evaluated by comparison to a perfect (gold standard) reference test; however, for Campylobacter such a gold standard is not available to accurately determine infection status. As the true status of each batch was unknown, this created difficulties in assessing the true sensitivity and specificity of the caecal culture test. In recent years, the use of Bayesian modelling to estimate the sensitivity and specificity of a diagnostic test has been applied for this purpose [Reference Gardner14–Reference Arnold16]. For a Bayesian model to infer the sensitivity and specificity of a culture method a second conditionally independent test needs to be applied in parallel to ensure there are, at least, as many degrees of freedom in the data as there are parameters requiring estimation [Reference Gardner14]. To this end, a PCR test was applied to the caecal samples in parallel with culture. The objective of this study was to apply a Bayesian framework to evaluate standard direct culture and PCR for the detection of C. jejuni and C. coli in broiler caecal samples, and use this framework to estimate the true prevalence of these species in broiler flocks in the UK.
MATERIALS AND METHODS
The Campylobacter status of a sample was considered ‘mixed’ when both C. jejuni and C. coli were detected in a sample. Samples, where C. jejuni alone was present, were denoted as ‘C. jejuni’ (i.e. a non-mixed sample), and similarly for ‘C. coli’.
Sample and data collection
Caecal samples were collected from broiler slaughter batches as part of a 3-year randomized national prevalence survey [Reference Lawes13]. Briefly, per slaughter batch, a single caecum was collected for sampling from 10 different broilers at the time of evisceration. Samples were selected according to the month of sampling and flock depopulation status (i.e. whether birds had previously been taken from the flock, sometimes known as ‘thinning’). Caecal content was obtained from the caecum, stored at 4°C and tested within 1 week after collection by bacteriological culture and PCR.
Culture of pooled caecal sample
The method used for the detection and speciation of Campylobacter spp. in caeca was in accordance with the technical specifications set out in Annex I of Commission Decision 2007/516/EC and as reported previously [Reference Lawes13]. For each slaughter batch, caecal contents were pooled from each of the ten caeca individually by homogenizing 0·02 g from each caecum in 2 ml phosphate buffered saline (PBS) (0·1 m, pH 7·2). A 10 μl volume of this was plated on mCCDA agar (CM739 base with SR155 supplement; Oxoid, UK) and incubated at 41·5 ± 1°C microaerobically (84% N2, 10% CO2, 6% O2). Plates were examined at 24 h and 48 h for grey flat, irregular and spreading colonies typical of Campylobacter spp. [Reference Fitzgerald, Whichard, Nachamkin, Nachamkin, Szymanski and Blaser17]. Up to five suspect colonies were subcultured micro-aerobically as described onto 7% sheep blood agar with 0·1% cyclohexamide (CM0055, Oxoid) before confirmation and species identification based on phenotypic methods described in ISO 10 272–1:2006(E) [18]. The detection limit for the culture method for caecal samples using mCCDA has been estimated as 102 c.f.u./g [Reference Rodgers19] and for the PCR as 105 c.f.u./g [Reference Randall12].
Each sample, was considered culture positive if at least one colony recovered was confirmed as thermophilic Campylobacter spp. Speciation tests were performed on one single colony per positive batch as described previously [Reference Lawes13].
DNA extraction and PCR test
Extractions were performed within 7 days of caecal collection using an ExtractMaster™ Fecal DNA Extraction kit (Epicentre Biotechnologies, USA). A 250 μl sample of caecal suspension in 750 μl PBS was centrifuged for 5 min at 13 000 g and the pellet was re-suspended in 5 μl Tris buffer (1 m, pH 8) and extraction continued following the manufacturer's protocol. DNA preparations were stored at −20°C until PCR testing for C. jejuni and C. coli as described previously [Reference Randall12]. A cycle threshold (Ct) of between 10 and 32 was viewed as a positive result for either the mapA probe (C. jejuni) or ceuE probe (C. coli). A negative result was recorded for mapA and ceuE probes if the Ct value was >32 and the Ct value of the internal amplification control (IAC) probe was <40; however, when there was no Ct value for the IAC probe the result was invalid.
Statistical methods
All statistical analysis was conducted in WinBUGS 1.4, using a modified version of the approach used in [Reference Gardner14], in which a Bayesian method is proposed for estimating the sensitivity and specificity of two tests applied to two populations in the absence of a gold standard. In this study, there were four populations:
-
(1) First batch removed from flock, non-summer (October–March).
-
(2) First batch removed from flock, summer (June–September).
-
(3) Previously partly depopulated, non-summer (October–March).
-
(4) Previously partly depopulated, summer (June–September).
Each pooled caecal sample was tested by PCR and culture. As the culture result was based on a single colony, it was only possible to detect one species of Campylobacter, although both C. jejuni and C. coli may have been present in the sample. In light of this, there were 12 possible outcomes for each sample tested.
For each population k = 1, … , 4, the observations were condensed into a descriptive 12-dimensional vector y k . These vectors, for each of the four populations, were assumed to have independent multinomial sampling distributions,
$${\bf y}_{\bf k} \sim {\rm multinomial}(n_k, {\bf p}_{\bf k} ),$$
where n k represents the number of observations in population k, and p k denotes the vector of probabilities associated with each of the 12 possible outcomes.
Elements of the vector p k are defined by weighting the sensitivity and specificity with the proportion of samples deriving from each of the four possible population statuses: C. jejuni only (π j ), C. coli only (π c ), both species (π m ), or Campylobacter-free, with the sum of the prevalence estimates for each species constrained to be no more than 1. A detailed description of the elements of the vector p k is given in the Supplementary material (Table S1).
It was assumed that the sensitivity of PCR to detect C. jejuni and C. coli would be unaffected by whether the sample contained both species. Similarly, the sensitivity of culture was estimated for both C. jejuni and C. coli for samples with separate estimates when only one species of Campylobacter was present and for mixed samples. Furthermore, for culture the sensitivity of Campylobacter spp. is likely to be affected by the relative proportions of mixed species samples in the population, therefore account was taken of the relative proportions of mixed/C. jejuni/C. coli samples in each population.
Priors
The model was initially run with non-informative priors throughout, except for specificity of culture. The sensitivity of the model results to the choice of priors was examined by (i) running the model with informative priors for the sensitivity of culture (as given in Table 1) and (ii) with non-informative priors for all parameters. Where informative priors were used, beta-distributed priors were based on previous studies [Reference Vidal9, Reference Randall12, Reference Woldemarium20] (Table 1). BetaBuster software (University of California, Davis, USA) was used to determine the parameters for each variable.
Priors used for the Bayesian model to estimate the sensitivity and specificity of PCR and caecal culture for detection of Campylobacter in broiler chickens
Data for the prior elicitation for the sensitivity of culture to detect C. jejuni and C. coli were derived from a previous study, where caeca, boot swabs and faecal samples were collected in parallel from 36 flocks [Reference Vidal9]. In this study, Bayesian methods were used to estimate the sensitivity of each of the sampling methods to detect C. jejuni and C. coli, where the samples were from a mixture of flocks with both species or C. jejuni alone (no flocks had C. coli only). The required prior for the sensitivity of culture to detect C. coli in a mixed sample, λ cm , was available from [Reference Vidal9]. The priors for the sensitivity of culture to detect C. jejuni in a non-mixed sample (λ jj ), and a mixed sample (λ jm ), were obtained by splitting the positive flocks into C. jejuni only and flocks with both C. jejuni and C. coli, and estimating the sensitivity of culture of C. jejuni in each case. For the sensitivity of culture of mixed samples, a Dirichlet distribution was used to represent the priors, to ensure that λ jm , λcm were each between 0 and 1 and that the sum of the probabilities λ cm , λjm , and the probability of a false-negative mixed sample (1–λ cm –λ jm ) summed to 1 (see Supplementary material for further details and choice of priors).
We assumed that PCR would correctly classify the Campylobacter species. Due to the possibly imperfect nature of the hippurate test for speciation, some misclassification was allowed for culture in the model. A similar approach was adopted as for setting priors for sensitivity of culture of mixed samples, i.e a Dirichlet distribution was used for λ jj , λjc . It was assumed that λ cj = λ jc , i.e. an equal probability of misclassification for either species. We also assumed vague (uniform in the range 0–1) priors for the specificity and sensitivity of PCR, and for the batch prevalence for each group of birds (season and thinning status).
The cut-off value used for designating a sample positive by PCR was Ct⩽32 [Reference Randall12]. However, the Bayesian model was also used to explore the impact of increasing the cut-off to 36 on the sensitivity and specificity of PCR to detect each Campylobacter species.
RESULTS
Overall, samples from 436 slaughter batches, originating from 22 abattoirs in the UK were tested by PCR and culture methods. There were very few samples with only C. coli present; the majority also contained C. jejuni detected by PCR (the number of mixed samples greatly outnumbered the number of C. coli PCR positives). Furthermore, PCR appeared to be less sensitive than culture, with a much higher number of PCR-negative samples being positive for culture (19, 1, 21 and 12 for populations 1–4, respectively; Table 2) compared to samples that were negative by culture but positive by PCR (3, 6, 15 and 4 for populations 1–4, respectively; Table 2). The apparent prevalence of C. jejuni was markedly higher (range 47·3–66·5%) than that of C. coli (range 8·8–22%) in all four populations tested by culture (Table 3).
Number of Campylobacter-positive batches from PCR by species (Ct < 32 to be designated positive) and the respective result of culture methods, applied to the four broiler populations, based on caecal samples (taken from broilers at slaughter as part of a national prevalence survey)
Number of Campylobacter-positive batches from PCR (for Ct < 32 and Ct < 36) and culture methods applied to the four broiler populations, based on caecal samples (taken from broilers at slaughter as part of a national prevalence survey)
n.a., Not applicable.
There was also evidence of a higher sensitivity of culture to detect C. coli, in line with the prior for relative sensitivity of culture to detect both species in a sample, as for mixed samples a total of 31 (3 + 6+11 + 11) were positive for C. jejuni by culture, compared to 52 (3 + 10 + 18 + 21) positive for C. coli by culture (Table 2).
Estimates from Bayesian model
An important impact of season on the true prevalence estimates of C. coli was observed. The estimated proportion of batches containing C. coli (the sum of the C. coli-only and mixed batches, Table 4) increased from 13·7% in non-summer to 43·7% during the summer months for flocks that had not been previously partly depopulated and from 22·2% in non-summer to 43·7% during the summer months for previously depopulated flocks. For C. jejuni, the impact of season appeared to be important only for batches that had not been previously partly depopulated; in this case the prevalence decreased from 81·4% in the summer to 52·3% in non-summer. For batches that had been previously partly depopulated there was very little change in C. jejuni prevalence between non-summer (88·1%) and summer (94·7%). The increase in C. coli during the summer months was mainly due to an increase in mixed positive samples with the corresponding reduction in the proportion of samples that contained only C. jejuni in batches of chickens from the previously depopulated flocks.
Estimated prevalence of Campylobacter in batches of broilers from four populations, using a Bayesian model applied to caecal sampling data*
CrI, Credible interval.
* A threshold value of Ct < 32 was used for a positive designation by PCR.
There was also an impact of the depopulation status of the flock on the true prevalence estimates of Campylobacter (Table 4). In the non-summer months, there was a marked difference in the total of Campylobacter prevalence between batches that were the first to be removed from the flock (56·4%) and previously partly depopulated batches (88·9%). In the summer, the Campylobacter prevalence of first removed batches was very high (84·1%), and even though the prevalence increased to 98·4% for previously partly depopulated batches, the relative change was much smaller than for non-summer.
Given a batch with only one species of Campylobacter present, culture had high sensitivity (97·5% and 83·3% for C. jejuni and C. coli, respectively), with a lower sensitivity for PCR (81·4% and 86·51% for C. jejuni and C. coli, respectively) (Table 5). The model results suggested a lower specificity of PCR (3·4% and 2·9% probability of a negative sample being designated as C. jejuni and C. coli, respectively) compared to culture (2·2.% and 1·0% for C. jejuni and C. coli, respectively). When analysis used a threshold of Ct<36 to designate a PCR test positive, there was a small increase in sensitivity and a similar decrease in specificity (see Supplementary Table S2).
Bayesian model estimates (plus 95% credible intervals) of the true sensitivity and specificity of PCR (using a threshold of Ct < 32 for a positive designation) and culture for detection of Campylobacter by species (applied to caecal samples from broilers)
CrI, Credible interval.
For mixed samples, the results indicated a bias in favour of culture detecting C. coli in preference to C. jejuni, with a 59·9% likelihood of a mixed sample being positive for C. coli compared to 37·3% of being positive for C. jejuni (Table 5).
There were differences of the order of <1% between the sensitivity of culture to detect Campylobacter spp. for each of the four sub-populations 1–4 (Table 6). The sensitivity of culture to detect C. jejuni was equal to or higher than that to detect C. coli; when there was a difference, it was more marked in the non-summer months, where the sensitivity of culture was 20% higher for C. jejuni than for C. coli, whereas for the summer months the difference was 5–10%, both for ‘first batch removed’ and for ‘previously depopulated’ batches.
Sensitivity of culture to detect Campylobacter spp., C. jejuni, and C. coli in each of the four sub-populations sampled in the study (i.e. taking into account the proportion of mixed samples cultured)
CrI, Credible interval.
Sensitivity of posterior estimates to prior assumptions
There was little difference in the results between the model with informative priors for culture sensitivity and specificity and results with only informative priors for culture specificity (the baseline model), except for the sensitivity of culture for C. coli (i.e. non-mixed) samples. The estimate of sensitivity for C. coli increased from 0·83 to 0·92 (but both with wide credible intervals) when informative priors were used for its sensitivity (Supplementary Table S3).
When non-informative priors were used for all parameters, there were few differences between the parameters compared to the baseline model, except for: (i) the estimate of the sensitivity of PCR to detect C. jejuni (in non-mixed samples) increased from 0·81 to 0·87, and (ii) the estimate of the specificity of culture reduced from 0·98 to 0·74.
DISCUSSION
This study has estimated the sensitivity and specificity of direct culture on mCCDA and a real-time PCR for detection of C. jejuni and C. coli from broiler caecal samples. Direct culture on mCCDA was the diagnostic test used to determine Campylobacter prevalence in a recent 3-year survey [Reference Lawes13] and was also used in the EU baseline survey [21]. Results indicate imperfect sensitivity and specificity of both PCR and culture, with potentially important differences in sensitivity of culture by species. This imperfect sensitivity suggests a likely underestimation of the prevalence of Campylobacter in the UK survey (see Table 6) where an overall prevalence of 79·2% was observed with 74·8% and 25·1% of the positive broiler batches being contaminated with C. jejuni and C. coli, respectively [Reference Lawes13]. The overall sensitivity of culture for both C. jejuni and C. coli varies between the sub-populations considered in the present study, probably due to differences in the relative prevalence and contamination levels of C. jejuni/C. coli/mixed samples in each sub-population. Estimation of the overall sensitivity of culture for each species indicates that there will be a greater underestimation of the C. coli prevalence than that of C. jejuni (Table 6). For samples where both species are present, there will be greater underestimation of C. jejuni than C. coli (Table 6).
Specificity of culture was found to be close to that obtained by a previous study (~98%) [Reference Woldemarium20], although it was higher for C. coli than C. jejuni. One possible cause of false positives and thus imperfect specificity of culture is the misclassification of species due to the hippurate test. The sensitivity of culture estimated in the present study is higher (64%) than that reported in [18] but comparison between the two studies is difficult owing to markedly different methodologies used for collection, transportation and culture of the samples. Other studies [Reference Rosenquist, Bengtsson and Hansen10, Reference Rodgers19] have also reported lower sensitivity of culture, but used a different matrix and therefore the culture tests are not directly comparable with the present study.
The analysis of Campylobacter results by species prompts the need for a more complicated expression for the likelihood of the data compared to the standard two 2-test model [Reference Gardner14]. Due to the larger number of prevalence and test sensitivities that need to be estimated in the Bayesian model here, it was considered important to obtain informative parameters for some of the parameters, in order to assist with model identifiability. The sensitivity analysis indicated that models with informative priors for the sensitivity and specificity of culture produced estimates with reasonably close agreement to those obtained using non-informative priors. One of the main differences in using non-informative priors was that the estimate of the sensitivity of culture to detect C. coli in a non-mixed sample was lower (83%) than the estimate with an informative prior (92%). This sensitivity of the estimate of C. coli culture to the choice of prior is likely to be due to the relatively small number of Campylobacter-positive samples that contain only C. coli and not C. jejuni; out of 436 samples only 19 were identified as C. coli-only positives by PCR (at a cut-off of 32). This leads to difficulties in robust statistical inference for the sensitivity of culture to detect C. coli from the present study data alone. The estimate for the specificity of culture for C. jejuni when it was the only species present was also influenced by the choice of prior; with non-informative priors it dropped from 0·97 to 0·74. There is a potential lack of identifiability in the infection status of samples that were C. jejuni positive for culture but C. jejuni negative for PCR. However, the estimate of specificity of 0·74 is not credible in the light of previous work [Reference Woldemarium20] or the recent EU survey on Campylobacter in broilers, with apparent prevalence in pooled caecal samples being as low as 2% in one MS [21].
The aim of the inclusion of an informative prior scenario as part of the sensitivity analysis was to determine how sensitive the model results were to changes in the priors. It turned out that the priors generated for the sensitivity of culture for mixed samples from a previous study [Reference Vidal9] were very close to the posteriors when non-informative priors were used, indicating good agreement with the present study.
A very high proportion of flocks detected as having C. coli by culture were co-infected with C. jejuni by PCR testing. This finding may indicate that in most cases where C. coli has colonized a flock, it does so at a higher level than C. jejuni at time of slaughter, similar to the findings of [Reference El-Shibiny, Connerton and Connerton22], albeit with a limited number of C. coli and C. jejuni strains. When both species are present, C. coli is generally present in higher numbers and hence more likely to be the species detected by culture (J. Rodgers, AHVLA, unpublished data). This might suggest a different epidemiology for the two species such as different contamination or multiplication rates; once C. coli colonization has occurred, C. jejuni may also colonize the flock but does not reach such high levels when C. coli is present. Further difference in the epidemiology of the two species is reflected by the impact of whether a batch was the first to be removed from the flock, and the impact of seasonality on the prevalence of Campylobacter. For Campylobacter spp. and C. jejuni these were consistent with previous studies, where prevalence was higher in thinned flocks, and in the summer [Reference Lawes13]. However, for C. coli, there was little difference in prevalence between the first batch and previously partly depopulated flocks, although there was a clear seasonal difference. Thinning or partial depopulation does not seem to increase the prevalence of C. coli; this may imply that other factors affect flock colonization with this species such as environmental and climate conditions. It may also indicate different times of colonization as an earlier colonization would allow more time for C. coli to become established and predominate in the flock. Further investigations are required to explore this hypothesis.
Although high, the sensitivity of PCR was lower than the culture method. This is possibly due to the intrinsic higher limit of detection of the PCR (>105 c.f.u./g) [Reference Randall12] compared to culture (>102 c.f.u./g) [Reference Rodgers19] and therefore caecal samples with lower concentrations of bacteria will not be detected by PCR. By contrast, PCR exhibited higher specificity over culture than that of 80% reported previously [Reference Randall12], but it should be noted on a much smaller number (n = 52) of samples. Furthermore, the latter study only tested a single flock population and was not able to apply Bayesian methods in the absence of a gold standard, therefore the adoption of culture as the reference standard may have influenced their results. The higher specificity found here is consistent with our experience of applying PCR to a large number of samples in Campylobacter surveys, where there were fewer culture-negative samples testing positive by PCR (data not shown). Furthermore, test validation of the PCR with several bacterial species underlines its high specificity [Reference Best23]. A high specificity (96·2%) of a real-time PCR relative to culture as the reference standard was also reported when applied to faeces from experimentally infected pigs [Reference Leblanc-Maridor24], and, by extension from above, the actual specificity could be even higher if some samples were false-negative by culture.
It would interesting to explore further the optimal choice of Ct threshold at which to determine a sample as positive by PCR, since changing this did have an effect on the sensitivity and specificity of the test. One method of doing this would be to use recently developed Bayesian methods that are able to analyse the PCR data without using a specific cut-off, i.e. the actual Ct value is used in the analysis [Reference Choi, Johnson and Thurmond25]. Such an approach would result in a more powerful dataset, since the model will have information on definite and borderline positives for each species, and provide further strength to the model inference.
In conclusion, the season of sampling had an important impact, especially for C. coli, which was more prevalent in the summer while for C. jejuni the effect of season was only marked for the batches first removed from the flock. Previous partial depopulation of the flock also had an important impact, with lower prevalence of both Campylobacter species (with a larger change for C. jejuni) for batches that were the first birds removed from the flock than for previously partly depopulated batches. Culture was more sensitive than PCR for both species in samples derived from populations infected with a single species of Campylobacter but its sensitivity was reduced for C. jejuni or C. coli in mixed populations although it was less marked for the latter. The PCR method is potentially able to detect both species in mixed samples but the total probability of correctly identifying at least one species by this method was similar to culture.
SUPPLEMENTARY MATERIAL
For supplementary material accompanying this paper visit http://dx.doi.org/10.1017/S0950268814000454.
ACKNOWLEDGEMENTS
The authors thank the Department for Environment, Food and Rural Affairs (Defra) for funding the study (project OZ0613) and the abattoirs and the poultry companies for participating in this study. The authors also thank Meat Hygiene Service colleagues for collecting the samples and colleagues from the Epidemiology and Risk Group and the Department of Bacteriology at the AHVLA for their involvement with the survey.
DECLARATION OF INTEREST
None.
