Spatial epidemiology of porcine reproductive and respiratory syndrome in Thailand

Porcine reproductive and respiratory syndrome (PRRS) is a pig disease that emerged in the last part of the twentieth century and spread through pig production systems. The disease was first recognized in the United States in 1987 and in Europe in 1990 [1]. Today, the disease has become endemic in many countries throughout the world following an epidemic phase [1]. The PRRS viruses (PRRSV) mainly cause reproductive failure in sows and respiratory disease in piglets and therefore significantly impact the productivity of affected pig production systems [2]. The viruses can directly transmit through placenta and body fluids of infected animals as well as through contaminated fomites, vectors and aerosols [3]. So far, the viruses have not been found in other host species in natural conditions [1]. In experiment conditions, some avian species, mallard ducks in particular, have been found to be able to harbor and shed the viruses for several days after challenged [4]. Following infection, most pigs clear the virus in 3–4 months in vivo but some pigs have been reported to remain carriers of the virus for several months [5].

Previous epidemiological studies on PRRS herd-to-herd transmission risk identified the introduction of infected pigs or semen into the herds as an important risk factor, with subclinical and persistently infected pigs playing an important role in the spread of the disease [6]. Several other farm-level risk factors have been previously reported, including the purchase of animals from herds with PRRSV, the use of semen for artificial insemination from infected boars, high herd size [7],[8], poor farm management practices (e.g. not isolating animals after purchase [7]), or the use of modified-live vaccines [8],[9]. At a higher geographical level, infections from neighboring farm by aerosol spread [8], and high densities of pig farms in the immediate area have also been recognized as risk factors [9].

In April 2006, an atypical form of PRRS was reported in China, Jiangxi province [10]. The new strain caused high and continuous fever, red discolorations of the skin, blue ears in the late phase of the disease, and high mortality rate in all ages of pigs. The virus was subsequently identified as highly pathogenic porcine reproductive and respiratory syndrome (HP-PRRS) characterized by a 30 amino-acid deletion in the nsp2-encoding region of the PRRSV [10],[11]. The PRRSV genome is approximately 15 kb in length and composed of 9 open reading frames (ORFs), of which ORF1a and ORF1b encode 12 non-structural proteins (nsp). The nsp2-encoding region is genetically the most variable area and crucial for viral replication due to its protease activity [12]. When compared to the typical strain, the new one showed a higher tissue tropism in vivo, which may contribute to its higher virulence [13]. Since 2007, this new HP-PRRS has been reported in many other countries throughout Asia [14]–[17].

In Thailand, the mild form of PRRSV was first isolated in 1996 from piglets with chronic respiratory distress and the virus was subsequently identified as the US genotype [18]. Since this first documented occurrence, all PRRSV strains isolated in Thailand have been identified as belonging to the EU and US genotypes [12],[19]. The import of pigs or semen from European and North American countries [19], or the smuggling or unauthorized use of modified-live vaccine [20] are possible pathways of entry into the country. There was no evidence of introduction of the new HP-PRRS in Thailand [12],[21] until its first report in pig farms located in the northeast of Thailand in 2010 [16]. Since then, the import of PRRSV-positive animals in the country was banned by the Thai authorities [21], but it is likely that the virus may have been introduced to Thailand through the illegal imports of infected animals [16].

In order to deal with PRRS outbreaks, animal quarantine, movement control, disinfection of infected premises/establishments, treatment of affected animals, and surveillance in and out the affected area were implemented and coordinated by the Department of Livestock Development (DLD) staff [22]. In addition, a public awareness campaign was launched to disseminate information on prevention good practice measures and identification of early symptoms. The campaign targeted pig farmers and aimed to provide them with detailed information on PRRS. Moreover, the campaign encouraged them to enhance their biosecurity level and to apply basic precautionary management practices such as using disinfectant before entering farms and quarantining newly introduced pigs, among other measures [22].

The objective of this study was twofold. First, we aimed to provide a first description of the spatio-temporal pattern of all PRRS cases in Thailand in 2010. Second, by assembling a series of variables quantifying the distribution of different pig types in the country and anthropogenic factors, we also aimed to carry out a first investigation of the risk factors most strongly associated with all PRRS occurrences, possibly including HP-PRRS in Thailand.

The temporal pattern of PRRS in Thailand (Figure 1) demonstrated only sporadic occurrences in 2007, 2008, and 2009. In 2010, the outbreaks from January to July 2010 were more frequent than in the previous 3 years, followed by a gradual increase from August to December 2010. A simple trend analysis indicated that the weekly occurrence tended to increase over time in the period of interest (linear regression, F_1,363 = 102.7, p < 0.001). The date that maximized the F-statistic of the linear regression including the day number and the period as a dummy variable was the 1^st of August 2010 (F_2,362 = 19.6, p < 0.001). During the first period (January - July 2010) and second period (August to December 2010), there was 0.4% (n = 33) and 1.1% (n = 82) of sub-districts reporting at least one outbreak, respectively.

The cluster analysis identified six spatial clusters of PRRS outbreaks in 15 provinces of Thailand across the 2 periods in 2010 (Figure 2). The primary cluster occurring in the second period and was located in the South-East of region 3, followed by 5 secondary clusters located in 4 regions : the West of region 2 (in the first and second periods), the South-East of region 8 (in the first and second periods), the South-West of region 4 (in the first and second periods), the North-East of region 6 (in the second period), and the North of region 4 (in the first period), respectively. The details of the spatial clusters are provided in Table 1.

The risk factors found significant at the alpha level of 0.1 in the univariate logistic regression analysis were (i) the number of farms with fattening pigs (p-value = 0.062), (ii) the number of farrow-to-finish farms (farms which include breeding, producing piglets and fattening pigs) (p-value = 0.061), (iii) the human population (p-value = 0.005), (iv) the number of farms with breeding sows (p-value = 0.016), (v) the density of farms with breeding sows (p-value = 0.024), (vi) the number of farms with breeding piglets (p-value = 0.031), and (vii) the density of farms with breeding piglets (p-value = 0.071). In the multivariate logistic regression model (Model I), the only variables found simultaneously significant were the human population and the number of farms with breeding sows (Table 2). Both variables were positively associated with the presence of the virus, with the change in log-likelihood upon removal of the number of farms with breeding sows (7.20) being slightly greater than that of the human population (6.28). The predictive power of the model, as quantified by the area under the curve (AUC) of the receiver operating characteristics (ROC) plots (as showed in Figure 3), was moderate (the mean AUC was 0.733 of the model set and 0.723 of the test set).

Variables	Mean coefficient	Mean SE*	Mean odds ratio (OR)
			Adjusted odds ratio	95% confidence interval
Constant	−2.611	0.428
No. of human population	0.479	0.253	1.669**	1.024 -2.779
No. of farms having breeding sows	1.750 × 10-2	1.823 × 10-2	1.016	1.004-1.029
Autoregressive term	2.438	0.521

*SE stands for standard error.

**The odds of a sub-district being PRRS positive was increased by a factor of 1.669 (95% CI 1.024 to 2.779) for every 10,000 unit increases in the number of human population in a sub-district.

Results of the multivariate logistic regression model (Model I) with 100 bootstraps applied for analyzing PRRS presence/absence at the sub-district level in Thailand.

The BRT model (Model II) was run with a tree complexity of 3, a learning rate of 0.003 and a bag fraction of 0.5. The combination of two variables yielded the maximum predictive power with a mean AUC of 0.961 on the model set and 0.801 on the test set (Figure 3). The two variables selected by our forward entry procedure were the human population and the density of farms with breeding sows, with an average respective relative contribution of 51.64 and 48.36, respectively (Table 3). The fitted function of the BRT model shows the effect of the predictive variable on the predicted response [23], with both variables showing a positive association with the fitted function (Figure 4).

Variables	Rel. Con.* (mean)	Rel.Con. (SD)	Mean of AUC** (range)
			Model set	Test set
No. of human population	51.64	5.44	0.961 (0.838-1.000)	0.801 (0.692-0.908)
Density of farms having breeding sows	48.36	5.44

*Relative contribution.

**Area under the curve (AUC) of the receiver operating characteristics (ROC) plots.

Results of boosted regression trees (Model II) with 100 bootstraps applied to model PRRS presence/absence at the sub-district level in Thailand.

When crosschecking the significant variables from both models, it was found that both risk factors identified by the forward-entry BRT model were significantly associated with disease occurrence when analyzed using the multivariate logistic multiple regression, but yielded a lower AUC than the variable selected directly by the multivariate logistic regression. Conversely, variables selected by the backward removal multivariate logistic regression were also providing good predictions when used in the BRT, but here too, with a lower AUC compared to the variables selected by the BRT approach.

The probability of PRRS presence predicted by the two models (Model I and Model II) was mapped for all sub-districts (Figure 5). Isolated high risk areas were distributed in several parts of the country in both maps with a higher frequency in the Northeast (regions 3 and 4), the North (regions 5 and 6) and the South (region 8 and along the border of regions 8 and 9). The map predicted by Model II, with higher predictive power, we allows identifying 9 areas with higher risk than the rest of the country: (i) the Northern area of region 5; (ii) the Northeastern areas of region 6; (iii) the Western area of region 6; (iv) the Northern area of region 4; (v) the Southern areas of region 4; (vi) the Northwestern areas of region 3; (vii) the Southeastern area of region 3; (viii) the central area of region 2; and (ix) the border area between regions 8 and 9.

The year 2010 may have marked a change in the epidemiology of PPRS in Thailand, with the first notification of HP-PRRS in the country. A change is apparent, both in the spatial distribution of PRRS and in the average number of cases reported daily. Three main hypotheses may explain this change. First, the common PRRS may have started spreading more extensively than before. Given that the disease has been known to be in the country, and quite widespread, for a reasonable amount of time, it seems unlikely that it would have suddenly spread to new areas for no apparent reason (e.g. changes in production systems, trading patterns, etc.). Second, the new and more pathogenic HP-PRRS variant could have started spreading in the country [16], resulting in more distinct clinical symptoms that made the disease more apparent, resulting in an increase in reported cases. Third, the first detection of HP-PRRS in Thailand may have triggered an increased awareness of farmers and veterinary officers to PRRS symptoms, which may also have contributed to a higher number of reported cases. It is difficult to formally differentiate these hypotheses in the absence of empirical data making a virological distinction between infections caused by PRRS and HP-PRRS viruses. However, the spatial and temporal patterns support that the period ranging from August to December should be analysed separately, since one can assume a different level of reporting during the first period.

Both statistical approaches used to explain the distribution of PRRS identified human population as an important risk factor associated with the occurrences of PRRS at the sub-district level. Human population is often found as a risk factor in livestock infectious diseases, such as for example highly pathogenic avian influenza [24]–[26], and this can reflect two mechanisms. Transmission can be supported by farm-related activities such as transport of outputs (pigs), or inputs (feed, piglets) through contaminated equipment, vehicles, boots, and others [1],[3],[27], allowing the disease to be introduced to other farms (between farms) or other pigs (within farm). One can intuitively assume that an isolated farm in a sub-district with a low human population could have a lower number of possibilities of transmission through those activities than farms located in densely populated areas. However, the positive statistical association between PRRS and human population may also simply result from a reporting bias, with farms located in the most populated sub-districts being more likely to report disease problems, either because farmers were better informed through information campaigns, or because significant mortality would be more easily witnessed by other farmers.

Both statistical modeling approaches also identified an association between the number of farms with breeding sows, either expressed in absolute terms (logistic regression) or as density (BRT). Whilst this specific factor has not yet been reported elsewhere, the most similar result was obtained by Weigel et al. [7], who found an increased risk of PRRS diagnosis associated with a higher number of sows. These results support the hypothesis that sow farms may play a key role in the epidemiology of PRRS, at various steps along their particular production cycles. First, the introduction of replacement gilts is a first possibility of introduction of the virus to the herd. Thanawongnuwech and Suradhat [28] indicated that the sources of PRRS infection in Thailand mostly resulted from infected replacement gilts from different sources or from inadequate gilt acclimatization. In addition, risk factors associated with gilts have been reported in other risk factor studies, who found risk factors such as, for example, “not isolating gilts after purchase” [7], or “purchase of gilts from a PRRS positive herd” [29]. Second, after introduction, sows usually become involved in the mating program, where both natural mating by boar or artificial insemination can potentially transmit the virus to the sows [30]. The infection can occur through physical contact with the infected boar or through semen [3]. Upon infection, the sows can potentially transmit the disease both horizontally to other pigs on the farm, and vertically to the fetus through passage of the virus across the placenta [3]. So, the replacement of gilt, their involvement into mating programs, and their central role for feeding piglets may explain the pivotal role played by sows in terms of PRRS transmission.

Despite testing a wide set of host types, farm types and anthropogenic variables, only two variables were associated with PRRS presence at the sub-district level. Furthermore, the AUC obtained by the logistic regression was only moderate, suggesting that many other factors, probably operating at the farm level were not accounted for. The comparatively higher AUC obtained through the BRT modeling approach would need to be confirmed by external validation, because even if the AUC quantified using the test appears to be fairly high, it may still be overestimated due to spatial autocorrelation of PRRS presence/absence in the training and test set.

Further analyses, combining sub-district level factors, with factors observed at the farm level may help to clarify the respective role of risk factors identified in this study. In particular, we found no significant association between PRRS presence and the number or density of breeding boars, which was identified as a major carrier of PRRS viruses in many previous studies [7],[8],[30]. However, the role of breeding boars may be strongly influenced by the extent their movements between farms, or even, between sub-districts, and by the particular measures that are taken by the owner to prevent transmission. Moreover, the identification of several spatial clusters highlights areas where local-scale risk factors could be better investigated [31].

In the long term, risk factor analysis aims to improve PRRS surveillance strategy by implementation of risk-based surveillance strategies [32]. Indeed, one can expect a more optimal use of the available resources if active surveillance is well targeted within high-risk farms or areas [32]. One approach could be, for example, to maintain passive surveillance in all types of pigs and pig farms, triggered by visible manifestation of clinical signs. To complement this, active surveillance programs could be implemented more systematically by sampling pigs in farms found to be at higher risk of PRRS infection, or suspected to play a particular role in the persistence and spread of PRRS. The results of this study suggest that farms with breeding sows may be one of those target groups, but based on previously published papers, breeding boars may be an additional and complementary focus of active surveillance due to their potential role in farm-to-farm disease transmission reported elsewhere. Simultaneously, other measures such as enhanced communication on good farm management practices and a public awareness campaign should be reinforced.