Between-farm direct or indirect contact via movement of animals or biological materials (e.g., semen), or cross-contamination through inputs such as machinery or human workers, is among the most important factors contributing to disease spread in food animals (1). Farm-to-farm contacts spread diseases that affect the US swine industry, including porcine reproductive and respiratory syndrome (PRRS) and porcine epidemic diarrhea (PED). For both PRRS and PED, animal movements (e.g., gilts, boars, weaned pigs, feeder pigs, and cull animals) represent one of the most important disease transmission routes between farms (2–6).

Understanding the network structure of food animal industries is critical for efficient production and disease control. For example, the sharing of information among agents (e.g., farmers, suppliers, and brokers) within a network may result in an increase of economic efficiency due to the selection of strategies that can decrease production and/or transaction costs (7). Indeed, social network analysis (SNA) is an analytical tool that has been widely used in the field of veterinary medicine to design disease control plans (8). SNA has been used to quantify the nature of connections (referred to as edges or contacts) among elements (nodes or vertices) in a population (9). Nodes may be farms or other facilities (e.g., slaughter houses, truck wash disinfection stations, or feed plants) from, to, or through which, animal populations are connected, and contacts among nodes may be categorized as direct or indirect (8, 10). SNA enables researchers to better understand animal movement patterns and, consequently, provide insights on how diseases diffuse in a given industry (11–13). For example, in many livestock industries, a minority of farms typically account for the majority of animal movements (14–16). Identification of those few farms, often referred to as “hotspots” or “super-spreaders” for disease transmission, may help formulating contingency plans to control high impact diseases, as timely intervention to targeted farms may enhance the probability of such plans being successful (13–15, 17). Similarly, efforts to improve animal management and biosecurity in super-spreaders may also contribute to reducing disease risk and prevalence (17, 18).

The US swine industry is characterized by large numbers of documented pig movements within and between states and regions. From 1970 to 2001, the number of pigs moved from one state to another (or from Canada to the US) increased from 30 to 50 million (19). Increases in the number and distance of movements reflect growth in the number of farms specializing in specific phases of the production cycle. Indeed, a growing proportion of feeder and finishing swine farms are located in the Midwest in close proximity to the grain used to feed pigs (20). In contrast, breeding populations tend to be located in areas distant from the major growing pig regions, such as the southeastern US, where grain inputs are not as critical (20–22). Whereas the regional specialization of different industry components has undoubtedly improved efficiency, the necessary movement of animals between the two regions alters the risk of long-distance disease spread (1, 22–24).

Animal movements are only partially regulated in the US, and no source provides complete information on such movements. For example, the United State Department of Agriculture, through the animal disease traceability program, collects information on movements of cattle, bison, equines, sheep and goats, swine, and poultry, only when movements cross state boundaries, except when livestock are moved to slaughter facilities or chicks moved from hatcheries (25). The lack of movement data creates a particular problem for the control of diseases, such as PRRS. In that context, regional control programs (RCPs), voluntarily organized and coordinated by producers, serve as means to share sanitary status information among farmers located in a given area. Sharing information within an RCP in Minnesota (RCP-N212) has been correlated with a decrease in PRRS incidence (26), and thus, one may hypothesize that sharing additional information about pig movements would further improve control program effectiveness. Unfortunately, lack of information about between-farm movements hinders attempts to describe network structure, hence impairing ability to prevent and control disease.

To elucidate the role of network structure in the spread of swine diseases, the relation between PRRS manifestation and animal movements between farms (and other related sites, such as buyer stations or market sites) was assessed in two counties in Minnesota (27). A positive association between positive PRRS status and the number of direct and indirect suppliers (in-reach degree) was observed in one county, but no additional network measures were significantly correlated with positive PRRS status (27). Although that early study provided valuable insights about pig movements between sites and their potential contribution to disease spread, a more complete assessment of the structure of contacts is required to understand disease spread. We use data from Wayne (27) and more recent data collected by the RCP-N212 to build a predictive movement model between sites, which is then used to estimate a complete movement network for the RCP-N212 in Minnesota. The results may be incorporated into a disease-spread model to help explain disease dynamics and support disease prevention and control activities within the RCP-N212.

The aim of this research was to predict animal movements among sites located within a given RCP. Unfortunately, data of movement networks are often incomplete or unavailable for food animal industries characterized by a large number of animal movements between sites, such as the US swine industry (19, 25). Therefore, we employed machine-learning techniques to illustrate how models may be fitted by using a subset of the data to increase their completeness and accuracy. Specifically, using information available in only two counties, we studied the likelihood of possible movements among sites in a larger-scale swine disease RCP in Minnesota, referred to as RCP-N212. In general, networks predicted by the RF model were consistent with the observed data used for model training and testing in terms of both spatial and production-type connectivity patterns.

The SN and RN networks exhibited relatively similar centrality patterns and a marked flow of animal movements from upstream to downstream sites in the production chain. However, the network structure of both counties also indicated that finishing (and even market) sites might provide animal replacement (e.g., gilts and boars) to upstream sites. Because outbreaks of diseases, such as PRRS, are also common in downstream sites (26), movements from those sites to upstream sites could perpetuate disease in those areas. Indeed, previous research has shown that despite an overall decrease in the occurrence of PRRS, spatial and temporal aggregations of that disease allowed for continued hotspots throughout the period of study (26). These interactions merit further analysis to explain swine disease dynamics, especially for industry-persistent diseases such as PRRS (2, 6, 48).

While model results suggest that ownership and distance are strongly related to the probability of pig movement between sites, the production type of the origin and destination sites also influenced the probability of pig movements from one location to another. Moreover, if we consider that different types of farms might share transportation services, whereby farms may ship or receive different type of animals (e.g., feeder pigs and finishing pigs), such mixing might facilitate spread disease via contaminated vehicles (4, 49). Thus, we suggest that a complete evaluation of disease risks associated with transportation of pigs between facilities should take into account factors such as the commercial relationship between sites, including contractual agreements in the US swine industry (50), and site production type (6).

As mentioned previously, we used information available from two small, county-based networks to estimate parameters for predictions of animal movements that closely fit observed movements as judged by standard statistical tests. We were able to validate our predictions within SN and RN. The parameters generated by our model were used to predict animal movements between sites over a larger area, i.e., RCP-N212. This is essentially an out-of-sample prediction. As data on actual animal movements were not available for RCP-N212, we cannot directly test the accuracy of our predictions for the larger network. The results for larger network appear reasonable in that they are consistent with the topology of the SN and RN networks, and these results may be useful in helping understand actual (but unobservable) animal movements in Minnesota. We believe that such out-of-sample prediction is warranted for the scale of RCP-N212, given that farms within this program are similar to the farms in SN and RN in terms of geography, demography, and management. However, predictions that would rely on more extensive extrapolation (such as at the scale of multiple states) would not be appropriate given the scope of our sampling. In addition, the value of being able to assemble a full network may not be to target individual farms, but rather to capture possible regional patterns in connectivity.

Based on comparisons between observed, county-level data and regional-level model predictions, such as the distribution of movement distances and general attributes of the network, we believe that our predicted full network for RCP-N212 has structural features that are similar to the partial data used to estimate the probability of movement between farms. Thus, our findings appear reasonable and provide insight to better understand animal movement patterns within the RCP-N212. This, in turn, may help farmers design private strategies for sanitary management, as well as aid policy-makers in structure-based decisions. However, given inherent limitations to predictive modeling, we acknowledge that our results may provide only general insights about movement patterns, thus additional work must be done before strong conclusions can be made regarding the utility of the predictions achieved in this way.

The RCP-N212 covers 34 out of 87 counties in Minnesota, accounting for 38% of the total swine facilities with 100 or more heads in the state. Because farm distribution is heterogeneous within Minnesota, with a greater number of farms toward the south (28), it is reasonable to infer that the distribution and type of sites across the RCP-N212 should influence our network predictions. Furthermore, given that 65.1% of the sites are dedicated to the last stage of production, we acknowledge that a fraction of sites within the RCP-N212 must trade animals with sites located in neighboring states, such as Iowa or Illinois, or even more distant states such as North Carolina, where a high number of sow farms are located (20, 51, 52). This could have led to an overestimation of movements between some sites in our model, especially for those underrepresented in the RCP. To tackle this issue, we applied two constraints to our model: (1) we restricted prediction of movements by increasing the probability threshold to 85% when assigning a possible link between two sites and (2) we constrained the maximum number of links per site using maximum values of in- and out-degree. As a result our movement predictions within the RCP-N212 were conservative, showing a lower density in the expected network (0.4%) than in the two observed networks (1.4% and 1.3% for SN and RN, respectively), although some metrics might have been overestimated, such as out-degree for nurseries. This is because there are very few nursery farms in this region, and many of the finishing farms are actually sourcing for pigs from other states outside of RCP-N212. However, our algorithm restricted their choice of nurseries to those within RCP-N212, perhaps leading to a false inflation of their out-degree. Future work should expand to larger geographic regions that encompass all stages of production, capturing movements between states. We anticipate that we could improve model predictions by obtaining information regarding contract relationships and animal movements among Minnesota sites and suppliers located outside RCP-N212.

In the predicted RCP-N212 network, we found that sites with higher in- and out-degrees overlap with areas where spatial and temporal aggregations of PRRS have occurred (26). Therefore, it is possible that movements, in addition to farm density, might play an important role in the persistent circulation of disease in the area. The co-aggregation of animal movements and PRRS, a disease believed to be transmitted, at least in part, by animal movements (2, 6, 49), suggests that our predicted network might be capturing important features of the underlying industry structure, which indirectly supports the validity of our network predictions.

The characterization of network structures often may help for planning production and designing strategies to control animal disease (7, 8, 11, 18). The approach developed here is an early step for helping in design strategies to control swine diseases regionally. For example, the spread of swine pathogens within the full network can be simulated using computational models, which would be valuable for both predicting patterns of between-farm spread and for evaluating alternate intervention and control strategies. Among them, for instance, vaccination strategies that maximize the collective good could be quantitatively explored, including minimum levels of coverage that may prevent disease circulation in the network. Additionally, the approach developed here may reduce time and cost for data collection, as collection of movements among a partial set of sites might be sufficient to predict movements among a larger set of sites.

Among classification techniques, there are several approaches that might be used to predict possible outcomes, such as links between sites. While the focus of this paper is not to provide an exhaustive review of these techniques, here we offer some ground for further discussion and perhaps comparative studies. The RF approach has high accuracy without overfitting, it is also fairly stable to the presence of outliers and noise, and it may handle the correlation between predictors (40, 41, 53). This may be important in the context of this study, as some atypical movements between sites may occur, predictors may be correlated, and the probability of animal movement between two or more sites may often occur in a non-linear fashion. Alternatively, other supervised techniques might be used. For example, support vector machines, a vector function based technique that splits the data for classification purposes, might resolve non-linearity in the data by using a non-linear kernel function, though its performance sometimes might be compromised (40).

In conclusion, we present an approach to predict the network structure of contacts between and among farms in a region by using partial data. Our results, combined with information on the occurrence of disease in the area (i.e., outbreaks of PRRS within the RCP-N212), may be incorporated into a disease transmission model that will help to evaluate the effectiveness of prevention and control strategies in a region, with the ultimate objective of mitigating the impact of endemic disease and hypothetical epidemic incursions. The approach here may also be applied to other regions and production systems, where information on animal movements is only partially regulated, thus improving decision-makers’ ability to plan and implement disease surveillance and control activities.

All the authors have met the four criteria described at the guidelines: PVD designed and data interpretation, revised and approved the version to be published, and agreed to be accountable for all aspects of the work. KV and LJ designed, revised and approved the version to be published, and agreed to be accountable for all aspects of the work. SW data interpretation, revised and approved the version to be published, and agreed to be accountable for all aspects of the work. AP designed, revised and approved the version to be published, and agreed to be accountable for all aspects of the work.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.