O, Salmonella, Where Art Thou? Modelling Salmonella infection in swine farms in Spain using Hamiltonian Monte Carlo methods
-
1
VISAVET Health Surveillance Centre (UCM), Spain
-
2
Center for Animal Health Research, National Institute of Agricultural and Food Research and Technology, Spain
-
3
Ministerio de Agricultura, Alimentación y Medio Ambiente (Spain), Spain
-
4
Department of Animal Health, Faculty of Veterinary Medicine, Complutense University of Madrid, Spain
Background
Salmonella infection is the second most prevalent foodborne zoonosis in the European Union (EU) with 91,662 confirmed human salmonellosis cases in the EU in 2017 [1]. In Spain, pork and pork products are a major source of human salmonellosis [2]. Two of the most common Salmonella serotypes in swine, S. Typhimurium and its monophasic variant (I 4,[5],12:i:-), were shown to be the dominant serotypes associated with human salmonellosis in Spain [1]. EFSA baseline reports on Salmonella infection in swine in Europe demonstrated that Spain had one of the highest levels of infection in fattening and breeder pigs [3, 4].
Spatial modelling techniques have been widely applied to understand the epidemiology of diseases; they can help to detect areas of higher risk of infection/disease, which could then be linked to potential risk factors. Conditional autoregressive (CAR) models that allow considering explicitly second order spatial effects have been widely used for this purpose. These models can be fitted in a Bayesian framework using Markov Chain Monte Carlo Gibbs sampling through popular software such as WinBUGS and OpenBUGS, although, depending on the model structure, the convergence of certain models can be challenging.
The current study explored the application of Stan, a Hamiltonian Monte Carlo-based framework, to fit Bayesian generalised linear regression CAR models using surveillance data of Salmonella infection at the pig-farm level in Spain. Stan is a state-of-the-art platform for full Bayesian statistical modelling [5]. It allows exceedingly flexible modelling with high-performance computation and has a highly supportive community that ensures information transparency. In this study, we first explored the spatial distribution and potential spatial trends in Salmonella infection at the pig-farm level in Spain by using multiple spatial analytical techniques and then examined the risk factors using multivariable models in Stan.
Methods
Data on samples collected for monitoring of antimicrobial resistance in Salmonella in swine from 2002 to 2013 and 2015 were derived from the database of Spanish Veterinary Antimicrobial Resistance Surveillance Network programme. Faecal samples were randomly collected annually that altogether added up to more than 50% of the slaughter capacity in Spain each year and that were located in no less than half of the autonomous communities of Spain. Each faecal sample, containing the faeces of two randomly sampled pigs from the same farm, was collected in a sterile container and stored at refrigeration (4°C) until it was sent to the laboratory within the next 24 hours. Salmonella isolation was performed according to ISO 6579:2002/Amd 1:2007, the method recommended by the European Union Reference Laboratory for Salmonella in faecal and environmental samples [6].
To examine potential risk factors for the risk of Salmonella infection at pig farm level in Spain, data related to pig farm practices in Spain were acquired. They included (a) the numbers and density of different type of farms (i.e., the combination of commercial, self-consumption, Intensive and mixed, and extensive farms) in each province (IEP) and (b) the numbers of different stages or types of pigs (i.e., piglets, weaners, fattening pigs, gilts, sows and boars) in different types of farms IEP. The ratios between the numbers of fattening pigs over other pigs IEP were calculated.
One sample was randomly selected to establish the farm Salmonella status for farms with more than one sample. Overall, the annual and provincial apparent prevalence was calculated as the number of positive farms over total farms sampled. Empirical Bayesian smoothing was performed to adjust provincial prevalence with Gabriel Graph describing the neighbouring relationships [7].
A Poisson model was fitted with the number of positive farms IEP as the outcome variable and the expected number of positive farms IEP as the offset. Global and local Morans’ I tests were run on the standardised residuals from the model to assess the presence of global and local spatial autocorrelation, respectively [7, 8]. The Poisson model of the spatial scan statistics was employed to detect the presence of areas with increased risk of Salmonella infection at the pig-farm level [9]. The pseudo-P-values for all these statistics were estimated using 999 iterations.
Bayesian modelling was performed in RStudio with ‘rstan’ and ‘rstanarm’ packages [10, 11]. Default weakly informative priors of ‘rstanarm’ package were used for the priors. Sampling was done with 4 Markov chains with 1000 iterations. Markov chain Monte Carlo diagnostics, model diagnostics and model selection were facilitated by ‘bayesplot’ and ‘loo’ packages [12, 13].
Poisson models with the same outcome and offset as the aforementioned model examined the potential spatial distribution of the risk of Salmonella infection and its association with pig farm practices in Spain. All covariates were standardised before their inclusion in the model. Two spatial effects, structured and unstructured, were examined in the model. The structured spatial effect, constructed using the CAR model, carried information about the neighbouring relationship between the provinces, and the unstructured consisted of independent province information.
Univariable models were fitted for all covariates, at first without considering the spatial effects. Correlation between covariates with a ratio of the mean to the standard deviation (MSD) of the posterior distribution >1.3 (i.e., pseudo-P value ≈ 0.2) was assessed using Spearman test. One of the two highly correlated covariates (i.e., r ≥0.85) along with covariates that had no high correlation with others (r <0.85) were then included in the multivariable model selection.
Model selection was performed in two stages. A backward model selection strategy using MSD ratio was first employed, starting with all selected covariates and biologically meaningful pair-wise interactions. The selection process stopped when all the MSD ratios were >1.7. The second stage of model selection was performed by comparing the predictive ability measured by (a) WAIC weights, (b) Pseudo-Bayesian model averaging (PBMA) weights without Bayesian bootstrap, (c) PBMA weights with Bayesian bootstrap, and (d) Bayesian stacking weights of the models. The final model had the highest values from most weighting methods. Models with all variables with MSD ratios >1.96 and with/without those with MSD ratios between 1.7 and 1.96 were considered in this second stage. The effect of adding (a) structured spatial effect, (b) unstructured spatial effect or (c) both were finally assessed in the model with the best predictive ability using the aforementioned weighting methods.
Results
Up to 3,027 samples over the 14 years were included in the study, with an average of 223 (range: 170-384) samples per year. A median of 60 (interquartile interval: 17.5-90, range: 4-383) samples were collected from 35 different abattoirs over the 13 years. Abattoirs were located in 11 out of 18 autonomous communities in Spain; 804 (26.6%), 652 (21.5%) and 403 (13.3%) samples were from Cataluña, Castilla-La Mancha and Murcia, respectively. The sampled farms were located in 43 out of 52 provinces in Spain; 424 (14.0%) were from Murcia, 261 (8.6%) were from Huesca and Barcelona, respectively, and 252 (8.3%) were from Toledo. Overall, Salmonella prevalence at the farm level was 34.3% (95% confidence interval [CI]: 32.6-36.0). The yearly, provincial and spatially adjusted prevalence are shown in Figure 1-3.
Neither global nor local spatial autocorrelation was detected in the residuals of the Poisson model. However, the spatial scan statistics identified local clusters with an increased risk in Salmonella infection at farm level in the northeast and the east of Spain (P<0.03; Figure 4).
The structured spatial effect in the univariable model suggested a West-East increasing risk of Salmonella infection in pig farms in Spain (Figure 5). In the final model, only the number of fattening pigs IEP was associated with an increased risk of Salmonella infection, with the risk of infection increased by 1.6% (95% credible interval [CrI]: 0.7−3.2%) each 10,000 increase of fattening pigs. Covariates associated with a decreased risk of Salmonella infection include the number of piglets IEP (1.3% [95% CrI: -0.2−2.7%] per 10,000 piglets), the number of weaners IEP (2.7% [95% CrI: 1.7−3.7%] per 10,000 weaners), and the ratio between the number of fattening pigs and other pigs IEP (15.3% [95% CrI: 4.6−24.9%] per unit). Furthermore, three interactions were associated with a decreased risk: (a) the interaction between the number of piglets and the number of weaners IEP (0.4% [95% CrI: 0.2−0.5] per 10,000 piglets or weaners), (b) the interaction between the number of fattening pigs and the ratio IEP (0.2% [95% CrI: 0.1−0.3%] per 10,000 piglets or per unit of the ratio), and (c) the interaction between the number of piglets and the ratio IEP (0.6% [95% CrI: 0.0−1.2] per 10,000 piglets or per unit of the ratio). None of the spatial effects could be retained in the final model.
Conclusion
We showed that there were more Salmonella-positive pig farms in eastern and north-eastern Spain than in the rest of the country. We also demonstrated that Stan can serve as an effective and efficient alternative for Bayesian spatial modelling in veterinary epidemiology. The number of fattening pigs IEP is associated with an increased risk of Salmonella infection in pig farms, and the number of piglets, weaners and the ratio between the fattening pigs are associated with a decrease in risk.
Acknowledgements
The authors acknowledge the funding from NOVA (Novel approaches for design and evaluation of cost-effective surveillance across the food chain) project.
References
1. European Food Safety Authority (EFSA) and European Centre for Disease Prevention and Control (ECDC). The European Union summary report on trends and sources of zoonoses, zoonotic agents and food-borne outbreaks in 2017. EFSA Journal 2018. 2018;16(12):5500.
2. European Food Safety Authority (ESFA) and European Centre for Disease Prevention and Control (ECDC). The European Union summary report on antimicrobial resistance in zoonotic and indicator bacteria from humans, animals and food in 2017. EFSA Journal 2019. 2019;17(2):5598.
3. European Food Safety Authority (EFSA). Report of the Task Force on Zoonoses Data Collection on the Analysis of the baseline survey on the prevalence of Salmonella in slaughter pigs, in the EU, 2006-2007, Part A: Salmonella prevalence estimates. The EFSA Journal 2008. 2008;135:1-111.
4. European Food Safety Authority (EFSA). Analysis of the baseline survey on the prevalence of Salmonella in holdings with breeding pigs, in the EU, 2008, Part A: Salmonella prevalence estimates. EFSA Journal 2009. 2009;7(12):1377. doi: 10.2903.1377.
5. Carpenter B, Gelman A, Hoffman MD, Lee D, Goodrich B, Betancourt M, et al. Stan: A probabilistic programming language. Journal of Statistical Software. 2017;76. doi: 10.18637/jss.v076.i01.
6. International Organization for Standardization. Microbiology of food and animal feeding stuffs - Horizontal method for the detection of Salmonella spp. - Amendment 1: Annex D: Detection of Salmonella spp. in animal faeces and in environmental samples from the primary production stage. Geneva, Switzerland 2007.
7. Roger S Bivand, Edzer Pebesma, Gomez-Rubio V. Applied spatial data analysis with {R}, Second edition: Springer, NY; 2013.
8. Gómez-Rubio V, Ferrándiz-Ferragud J, Lopez-Quílez A. Detecting clusters of disease with R. Journal of Geographical Systems. 2005;7:189-206.
9. Chen C, Kim AY, Ross M, Wakefield J. SpatialEpi: Methods and data for spatial epidemiology. 2018.
10. Stan Development Team. RStan: the R interface to Stan. R package version 2.17.3. 2018.
11. Stan Development Team. RStanArm: Bayesian applied regression modeling via Stan. R package version 2.17.4. 2018.
12. Gabry J, Mahr T, Bürkner P-C, Modrák M, Barrett M. bayesplot: Plotting for Bayesian Models. 1.7.0. 2019.
13. Vehtari A, Gabry J, Yao Y, Gelman A. loo: Efficient leave-one-out cross-validation and WAIC for Bayesian models. R package version 2.1.0. 2019.
Keywords:
spatial analysis,
Spatial clustering,
CAR model,
STAN,
Bayesian,
Salmonella,
Spain,
pigs
Conference:
GeoVet 2019. Novel spatio-temporal approaches in the era of Big Data, Davis, United States, 8 Oct - 10 Oct, 2019.
Presentation Type:
Regular oral presentation
Topic:
Spatio-temporal surveillance and modeling approaches
Citation:
Teng
K,
Martinez Aviles
M,
Ugarte
M,
Barcena
C,
De La Torre
A,
Lopez
G and
Alvarez
J
(2019). O, Salmonella, Where Art Thou? Modelling Salmonella infection in swine farms in Spain using Hamiltonian Monte Carlo methods.
Front. Vet. Sci.
Conference Abstract:
GeoVet 2019. Novel spatio-temporal approaches in the era of Big Data.
doi: 10.3389/conf.fvets.2019.05.00007
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
10 Jun 2019;
Published Online:
27 Sep 2019.
*
Correspondence:
Mx. Kendy Tzu-Yun Teng, VISAVET Health Surveillance Centre (UCM), Madrid, Spain, kendy.t.teng@gmail.com