Abstract
Diarrhoeal disease (scours) in piglets, often associated with enterotoxigenic Escherichia coli (ETEC), is a substantial financial burden to the pig industry worldwide. Previous research has not explicitly examined the relationships between farm, pen and microbiological factors. Here we present a state of the art analysis to reveal empirical indirect – as well as direct – associations between management factors as putative risks for scours in pre- and post-weaned piglets. A Bayesian Network is constructed to identify the optimal structural model describing the relationships between risk factors. An additive model is then built to estimate more epidemiologically familiar odds ratios. Farm-level variance dominates the model, making many pen-level associations null. However, there is evidence that pre-weaning scours are less likely on farms with <400 sows (0.14, 0.03–0.50). Our results strongly suggest that smaller production units (piglets/pen) could reduce the incidence of scours in piglets. There is also some evidence that ownership of other livestock is a potential risk factor for pre-weaning scours, although this was observed only at one farm. Future research should be directed at better understanding the role of herd size and investigating the relationship between managing other livestock and the occurrence of scours in pig herds.
Similar content being viewed by others
Introduction
Pre- and post-weaning diarrhoea (scours) is a major production-limiting disease in Australian pig farms costing an estimated $AU7 million each year1. Outbreaks of scours are common and result in reduced growth rates, higher medication costs, increased morbidity and mortality2. Despite the importance of scours, the management practices that may contribute to or affect disease outbreaks are not well understood in Australian pig production systems.
As a syndromic disease, the causes of scours are many and varied. Stressors such as alterations to the quality, type or quantity of the diet, poor air quality and other environmental factors can contribute to scours3,4,5,6. Enterotoxigenic Escherichia coli (ETEC) have the ability to colonise the lower intestine via fimbrial adhesins (F4 (K88), F5 (K99), F6 (987P), F18 and F41)7 and the production of enterotoxins (heat-stable STa and STb, heat-labile LT, and Stx2e)8 results in dehydration and scouring. At weaning, maternal antibodies wane and with limited acquired immunity, piglets are susceptible to pathogenic bacterial species and more vulnerable to developing scours. Once piglets are placed in a pen following weaning, bacteria are shed in the faeces and this can then act as a source of infection to other piglets in the same pen or via shared housing. The survival and spread of bacteria is largely due to management practices once the pathogen is present on the farm.
More than three decades ago a pivotal study of weaning disorders in 89 French pig herds with contemporary conventional husbandry characterised disease onset as an eco-managerial problem9. Through examination of 515 environmental variables, 10 primary risk factors for weaning disorders were identified5. This study crucially recognised that these risks were interlinked and context dependent. Later the predictive power of individual (putative risk) factors were shown to vary over time10.
Australian herd management strategies have evolved since the early studies of environmental drivers of scours. For example, housing has changed with weaner piglets generally housed in larger groups and raised in eco-shelters on deep litter11, 12. Similarly, sow stalls are being phased out in favour of grouped housing, there are improvements to diet and nutrition and management of diseases through enhanced biosecurity procedures, and increased vaccine availability13. These changes have increased productivity, but have not been matched by comparable research into the drivers of diarrhoeal disease and whether key risk factors have changed14.
Here we use state of the art analytical techniques to explicitly examine the causation of scours in the context of pig farm management. Using a survey of pig farms in southeastern Australia, we construct a Bayesian Network (BN) using structure discovery algorithms to reveal empirical relationships between production factors and pre- and post-weaning scours15. Traditional analytic approaches reduce disease systems to a single outcome and ignore the relationships between risk factors that can lead to indirect risk pathways. When translating such results into applied tools, ignoring indirect pathways can overestimate the success of interventions; furthermore, understanding the whole disease system can lead to novel interventions through the control of highly influential factors that act as junctions of risk pathways15. To address multi-causal disease syndrome, traditional approaches to understand disease causation are inadequate.
Results
Characteristics of the farms included in the analysis are shown in Table 1. Herd size ranged from 45 to 20,000 sows per farm (mean 721 sows). Most farms were observed to have either (or both) pre- and post-weaning scours (17% and 24%, respectively) at the time of sampling. Approximately 22% of pens had a history of diarrhoea, as reported in the questionnaire. The number of pens with <199 and ≥200 piglets (the median value) was similar, although the distribution was skewed by one herd with very large pens (1,000 piglets).
The identification of the BN structure assumed conditional probability tables for each node. In converting the model to an additive BN, it was necessary to remove intercept terms because in most cases these were collinear with other variables. Removal of the intercept terms made the model tractable. The model structure – averaged over the bootstrap samples – included 43 arcs with greater than 50% support (Fig. 1). Estimating odds ratios from the additive BN reduced the number of arcs with statistical support to 30. Up to 10% of data was missing due to producers declining to answer questions.
The risk of pre-weaning scours was a function of the number of sows kept, whereby ownership of <400 sows was protective against pre-weaning scours (median odds 0.14, and 95% credibility interval 0.03–0.50). The odds ratio of pre-weaning scours was perfectly predicted by ownership of other livestock because only one farm had neither other livestock nor pre-weaning scours. This arc was therefore removed after model fitting (a dashed line in the figure). Out of the 17 farms included in the final analysis, 14 (82%) had both ownership of other livestock and evidence of pre-weaning scours.
Pre-weaning scours were a risk factor for post-weaning scours (3.19, 1.99–5.34). There was strong evidence that the number of sows was associated with post-weaning scours; however, unlike pre-weaning scours, a smaller number of sows was a risk factor (1.97, 1.00–3.87). The odds of a history of diarrhoea, whilst related to the presence of bedding material, did not have any statistically supported associations once the additive BN was calculated.
Of note, there was no statistical support for associations between recent disease events (within the last twelve months) and either pre- or post-weaning scours. Also, an association between suspected beta haemolytic E. coli and either pre- or post-weaning scours was not supported by the data. Beta haemolytic E. coli identification was a statistical orphan node in the network (i.e. there was no statistical evidence to support associations between this and any other variable in the BN). Non-beta haemolytic E. coli was also an orphan variable, although it appeared (with weak evidence) to occupy a very different set of pathways.
The inclusion of farm-level random effects dramatically reduced the number of supported arcs (from 30 to 5, shown in Fig. 2). This is important because it suggests that many factors, perhaps unsurprisingly, co-occur based on management practices. Importantly, there was continued evidence for the associations between the number of sows and pre-weaning scours. This suggests that even after accounting for clustering within farms, there is strong evidence to support this association at the pen level. The association between livestock and pre-weaning scours was removed because it reflected a single farm and was therefore likely to be biased.
Curiously, the other arcs that were supported with the inclusion of farm-level random effects (i.e. the odds of suspecting >0% beta haemolytic E. coli) had not previously been supported by the credibility intervals of the additive BN. This is surprising because at the pen level, ignoring the clustering within farms, this factor was not statistically supported. However, the inclusion of farm clustering revealed stronger evidence that the odds of detecting beta haemolytic E. coli were reduced by weaning at >5 weeks of age (0.30, 0.08–0.97), which itself was reduced by the absence of litter. The odds of not having litter were strongly associated with not having intensive indoor production, suggesting that beta-haemolytic E. coli may be more likely to be seen in extensive production (although this association was not significant; χ2 = 0.88, p = 0.35, and hence no direct arc was supported in the BN).
Discussion
Scours in weaning pigs is an on-going challenge to pig production in Australia. Despite considerable investment in counter measures, little research has been done to untangle management factors that may reveal the eco-epidemiology of diarrhoeal diseases. Here we explicitly examine the relationships between farm and pen level factors that are putative risk factors for scours.
The risks of pre-weaning scours were robustly predicted by the number of sows. This implies a route of transmission of enteropathogens that is worthy of further investigation. On larger farms – indicated by a higher number of sows (and the arc from post-weaning scours to the number of piglets per pen and pen size) – pre-weaning scours was more likely to be observed and these farms, at least some, were at risk of post-weaning scours. The latter association is, however, more likely to be a farm management issue than a piglet-pen risk. In particular, it is interesting that the presence of beta-haemolytic E. coli and the use of vaccination – in either sows or pre-weaned piglets – had no direct associations with pre- or post-weaning scours. Beta-haemolytic E. coli can be detected in the faeces of both healthy and diarrhoea piglets but studies have observed higher levels in diarrhoeal piglets6, 16; vaccination is a common practice in pig production14.
Recently a Canadian study assessing biosecurity practices and diseases in pigs found strong associations between proximity to other livestock and disease transmission17. Although Cox et al. targeted Porcine reproductive and respiratory syndrome, Swine influenza, Mycoplasma pneumonia and Swine dysentery, and used a proximity measure (whereas we asked specifically “Do you run other livestock on the same property?”), this suggests that by minimising such risk factors associated with E. coli scours producers could also minimise other diseases known to have dramatic impacts on pig production. We found that only one farm had neither ownership of other livestock nor pre-weaning scours, making it difficult to draw robust conclusions. However, ownership of other livestock could be an indicator of inadequate biosecurity, or might increase the risk of pathogenic strains of E. coli being introduced to a pig herd. This risk factor is worthy of further investigation. It is a common practice for Australian pig producers to raise other livestock on their farm (van Breda et al.14).
Though statistically biased (due to separation in the data), the majority of farms analysed had both other livestock on the same farm and pre-weaning scours. This is suggestive of transmission between animals18. Once within a pen, smaller management units tend to be less prone to post-weaning scours, however this depends on the management context of a production system19. It would be useful to examine alternative causes of scours given that vaccination in general, and for E. coli in particular, was only tangentially associated with diarrhoeal disease in piglets. Given the plethora of infectious causes of enteropathy it is possible that less ubiquitous species are transmitted between hosts.
A key finding presented here is the dominance of farm-level variability. Transmission routes likely vary between farms, and given the associations at a pen level, are associated with pen and unit size. Fewer piglets per pen are likely to be associated with reduced risk of post-weaning scours. We hypothesise that this is related to the chance of identifying and responding to early signs of scours in a timely fashion, in addition to the differences in management between smaller and larger enterprises. For example, farms with fewer buildings (≤5) are less likely to have ventilation, and pens with ventilation are more likely to have fewer than 200 piglets per pen and to have bedding.
This study is the first to collect so much data on farm, pen and microbiologic aspects of scours from south eastern Australian pig production. As such is it uniquely positioned to explore the balance of management factors that may influence the risk of scours in weaning pigs. However, the dominance of farm-level variability in management factors limits extrapolation of findings to other farms that have different farm practices. The diversity of farm management systems necessitated some computational simplifications (such as binary categorisation of factors). The distribution of farm sizes reflects the breadth of production systems with 80% of pens containing fewer than 350 piglets, but the remaining 20% ranging up to 1000 piglets per pen.
The transient nature of scours makes identifying key risk factors difficult. Bayesian network analysis has been used previously to identify risk factors associated with pig diseases and climate20 and antimicrobial resistance in E. coli 21. Bayesian network techniques are an invaluable modern tool to understand and map risk factors and disease. Future work on scours to examine on-farm biosecurity procedures and potential routes and causes of infection would be useful. The effectiveness of control strategies – and particularly the use of antibiotics – for scours in pig production could be explored using Bayesian network techniques.
Materials and Methods
Data
Data on 58 variables were collected through a survey of 195 pig pens from 22 farms across southeastern Australia (New South Wales [n = 9], South Australia [n = 3] and Victoria [n = 10]). This survey has been described in detail by Van Breda et al.14. The number of herds included in this study was based on the available funding and the practicalities of sampling a large geographical area. Farms recruited were identified by swine veterinarians using prior knowledge of E. coli outbreaks; farms were also recruited at the 2013 Bendigo Pig Fair, Victoria and by contacting farmers directly via industry membership. A preliminary telephone questionnaire was conducted to establish the owner/manager’s willingness to participate and the suitability of their herd to be included in the study. Herds with less than 40 sows were excluded due to the limited number of piglets available for sampling at the time of the herd visit.
A cross-sectional survey was conducted between September 2013 and May 2014 and faecal samples to identify E. coli were collected from each herd. Samples were stratified to capture piglets one week prior to weaning (10 samples) and piglets that had most recently been weaned on the day of the herd visit (40 samples). Depending on the size of the pen, five samples were collected from pens of <20 piglets; and 10 samples were collected from pens of >20 piglets (an average 8.9 pens [range 1 to 14] were sampled per farm). A formal questionnaire, administered to the owner, manager or the leading farm hand, was conducted once at each farm. It included questions covering management, production and biosecurity practices. Details of the questionnaire, sample management and analysis is described elsewhere14. All animal sampling procedures and interactions were carried out in strict accordance with the recommendations made by The University of Sydney Animal Ethics Committee. The protocol was approved by The University of Sydney Animal Ethics Committee (Approval number: N00/7-2013/3/6002).
Statistical Analyses
BN analyses to address model structural uncertainty are computationally intensive. Therefore, to make the network tractable, all variables were screened to eliminate variables that would either reduce the sample size of the model or had no statistical evidence to justify inclusion in the network. First, the proportion of missing observations for each variable was counted. Fourteen variables with more than 5 per cent of observations missing were dropped from further analysis. This reduced the number of variables from 58 to 44. Second, the remaining variables were examined for univariate association with either pre- or post-weaning scours at the time of herd visit or with a history of diarrhoea using logistic regressions and an inclusive threshold of P ≤ 0.2. This was conducted with complete observations and, when the retained variables had also been screened for possible overlap, based on expert knowledge (e.g. the number of samples collected and the percentage E. coli positive). This resulted in 23 variables with some statistical evidence of association with scours that were retained for further analysis. To construct the BN, observations with missing data for any given variable were excluded, reducing the sample size from 195 to 174 pens and from 22 to 17 farms. The motivation for removing variables during analysis is summarised in Table 2.
A BN was constructed from the retained variables. Each variable was represented as a binary node and the structure identified using a modified (tabu) greedy hill-climbing search. Structural uncertainty was modelled with non-parametric bootstrapping (1000 samples) and only arcs with greater than 50% support were retained. The conditional-probability model was then translated into a system of mixed-effects logistic regressions. An additive BN has a more epidemiologically familiar interpretation (odds ratios)22 with clearer interpretation compared to a BN (with probabilistic contingency tables) and allows the inclusion of random effects representing the 17 farms included in analysis to account for clustering. The disadvantage of an additive BN is the computational limits on examining so many interconnected variables. The conditionally probabilistic model, though conceptually different, was computationally tractable to identify the model structure and allow parameter estimation. Random effects were fit with diffuse gamma distributions. The network structure was calculated by the R package bnlearn23 and the parameter estimates in JAGS 3.4.024.
Data availability
Survey data is available at https://doi.org/10.1371/journal.pone.0172528.s002 and https://doi.org/10.1371/journal.pone.0172528.s003.
References
Department of Agriculture and Fisheries. Piglet scours, https://www.daf.qld.gov.au/animal-industries/pigs/pig-health-and-diseases/a-z-pig-diseases/piglet-scours. Accessed 27/02/2017 (2010).
Fairbrother, J. M., Nadeau, E. & Gyles, C. L. Escherichia coli in postweaning diarrhea in pigs: an update on bacterial types, pathogenesis, and prevention strategies. Anim. Health Res. Rev. 6, 17–39 (2005).
Campbell, J. M., Crenshaw, J. D. & Polo, J. The biological stress of early weaned piglets. J Anim. Sci. Biotechnol. 4, doi:10.1186/2049-1891-4-19 (2013).
Lee, I. K., Kye, Y. C. & Kim, G. R. Stress, Nutrition, and intestinal immune responses in pigs - a review. Asian-Australas. J. Anim. Sci. 29, 1075–1082 (2016).
Madec, F. & Josse, J. Influence of environmental factors on the onset of digestive disorders of the weaned piglet. Ann. Rech. Vet. 14, 456–462 (1983).
Pluske, J. R., Pethick, D. W., Hopwood, D. E. & Hampson, D. J. Nutritional influences on some major enteric bacterial diseases of pig. Nutr. Res. Rev. 15, 333–371 (2002).
Nagy, B. & Fekete, P. Z. Enterotoxigenic Escherichia coli (ETEC) in farm animals. Vet. Res. 30, 259–284 (1999).
Zhang, W. P. et al. Significance of heat-stable and heat-labile enterotoxins in porcine colibacillosis in an additive model for pathogenicity studies. Infect. Immun. 74, 3107–3114 (2006).
Tillon, J. P., Meurier, C., Madec, F. & Josse, J. Mise en place d’un réseau d’observation sanitaire des éle- vages de porcs de la région de Bretagne. Bull. Off Int. Epiz. 92, 835–844 (1980).
Buddle, J. R., Mercy, A. R., Skirrow, S. Z., Madec, F. & Nicholls, R. R. Epidemiological studies of pig diseases: 1. Use of French protocols for risk factor assessment to predict the health status of Australian herds. Aust. Vet. J. 75, 274–281 (1997).
Barnett, J. L., Hemsworth, P. H., Cronin, G. M., Jongman, E. C. & Hutson, G. D. A review of the welfare issues for sows and piglets in relation to housing. Aust. J. Agric. Res. 52, 1–28 (2001).
Kruger, I., Taylor, G., Roese, G. & Payne, H. Deep-litter housing for pigs. In Primefacts profitable and sustainable primary industries: Primefact 68, https://www.scribd.com/document/11493551/Deep-Litter-Housing-for-Pigs-Primefact-68-Final. Accessed 09/05/2017 (2006).
Department of Primary Industries. NSW Pork Industry Overview 2015 - Intensive livestock production, http://www.dpi.nsw.gov.au/__data/assets/pdf_file/0011/578747/pork-industry-overview-2015.pdf. Accessed 09/05/2017 (2015).
Van Breda, L. K., Dhungyel, O. P., Ginn, A. N., Iredell, J. R. & Ward, M. P. Pre- and post-weaning scours in southeastern Australia: A survey of 22 commercial pig herds and characterisation of Escherichia coli isolates. PLoS One 12, doi:10.1371/journal.pone.0172528 (2017).
Ward, M. P. & Lewis, F. I. Bayesian Graphical modeling: Applications in veterinary epidemiology. Prev. Vet. Med. 110, 1–3 (2013).
Hampson, D. J., Woodward, J. M. & Connaughton, I. D. Genetic analysis of Escherichia coli from porcine postweaning diarrhoea. Epidemiol. Infect. 110, 575–581 (1993).
Cox, R., Revie, C. W., Hurnik, D. & Sanchez, J. Use of Bayesian Belief Network techniques to explore the interaction of biosecurity practices on the probability of porcine disease occurrence in Canada. Prev. Vet. Med. 131, 20–30 (2016).
Moon, H. W. & Bunn, T. O. Vaccines for preventing enterotoxigenic Escherichia coli infections in farm animals. Vaccine 11, 213–200 (1993).
Geenen, P. L., Döpfer, D., Van der Meulen, J. & De Jong, M. C. M. Transmission of F4+ E. coli in groups of early weaned piglets. Epidemiol. Infect. 133, 459–468 (2005).
McCormick, B. J. J., Sanchez-Vazquez, M. J. & Lewis, F. I. Using Bayesian networks to explore the role of weather as a potential determinant of disease in pigs. Prev. Vet. Med. 110, 54–63 (2013).
Ludwig, A. et al. Identifying associations in Escherichia coli antimicrobial resistance patterns using additive Bayesian networks. Prev. Vet. Med. 110, 64–75 (2013).
Lewis, F. & Ward, M. Improving epidemiologic data analyses through multivariate regression modelling. Emerg. Themes Epidemiol. 10, doi:10.1186/1742-7622-10-4 (2013).
Scutari, M. Learning Bayesian Networks with the bnlearn R Package. J. Stat. Software 35, doi:10.18637/jss.v035.i03 (2010).
Plummer, M. JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling, https://www.r-project.org/conferences/DSC-2003/Proceedings/Plummer.pdf. Accessed 09/05/2017 (2003).
Acknowledgements
This study was funded by the Cooperative Research Centre for High Integrity Pork (Project 2A-106). Dr. Om Dhungyel, The University of Sydney, is thanked for assistance in sample collection, questionnaires and laboratory analysis.
Author information
Authors and Affiliations
Contributions
L.v.B., M.P.W. designed the study and collected data. B.M.c.C. undertook the analysis. L.v.B., B.M.c.C., M.P.W. drafted the manuscript.
Corresponding author
Ethics declarations
Competing Interests
The authors declare that they have no competing interests.
Additional information
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
McCormick, B.J.J., Van Breda, L.K. & Ward, M.P. Bayesian Network analysis of piglet scours. Sci Rep 7, 6202 (2017). https://doi.org/10.1038/s41598-017-06399-2
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-017-06399-2
This article is cited by
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.