Forecasting viral disease outbreaks at the farm-level for commercial sow farms in the U.S

Abstract

Porcine epidemic diarrhea virus (PEDv) was introduced to the U.S. in 2013 and is now considered to be endemic. Like many endemic diseases, it is challenging for producers to estimate and respond to spatial and temporal variation in risk. Utilizing a regional spatio-temporal dataset containing weekly PEDv infection status for ∼15 % of the U.S. sow herd, we present a machine learning platform developed to forecast the probability of PEDv infection in sow farms in the U.S. Participating stakeholders (swine production companies) in a swine-dense region of the U.S. shared weekly information on a) PEDv status of farms and b) animal movements for the past week and scheduled movements for the upcoming week. Environmental (average temperature, humidity, among others) and land use characteristics (hog density, proportion of area with different land uses) in a 5 km radius around each farm were summarized. Using the Extreme Gradient Boosting (XGBoost) machine learning model with Synthetic Minority Over-sampling Technique (SMOTE), we developed a near real-time tool that generates weekly PEDv predictions (pertaining to two-weeks in advance) to farms of participating stakeholders. Based on retrospective data collected between 2014 and 2017, the sensitivity, specificity, positive and negative predictive values of our model were 19.9, 99.9, 70.5 and 99.4 %, respectively. Overall accuracy was 99.3 %, although this metric is heavily biased by imbalance in the data (less than 0.7 % of farms had an outbreak each week). This platform has been used to deliver weekly real-time forecasts since December 2019. The forecast platform has a built-in feature to re-train the predictive model in order to remain as relevant as possible to current epidemiological situations, or to expand to a different disease. These dynamic forecasts, which account for recent animal movements, present disease distribution, and environmental factors, will promote data-informed and targeted disease management and prevention within the U.S. swine industry.

Keywords: Animal movement; Forecasting; Machine learning; Porcine epidemic diarrhea virus; Spatial epidemiology; Swine.