What if you could use sow data to predict the future? That’s what the SwinalytIQ team at Iowa State University is working to do. By evaluating information already collected at sow farms, they are estimating a group’s mortality rate in the nursery prior to visible signs of declining health.
The Nursery Mortality Challenge
Nursery mortality is costly and complex, explains Mateus Cardoso, a research assistant and master’s student in bioinformatics and computational biology at Iowa State University. Unfortunately, this is often recognized only after losses have already begun. In large production systems, those costs add up quickly.
“The goal is to estimate a nursery group’s mortality risk at or before weaning, or approximately 60 days before the end of the nursery period,” Cardoso says. “This estimate will help producers focus attention and resources on the groups most likely to experience problems.”
Using a Predictive Model to Determine the Future
To build a predictive model, researchers collected records from 2,025 commercial weaning groups. This data included sow farm health information such as porcine reproductive and respiratory syndrome (PRRS) status, Mycoplasma pneumonia and porcine epidemic diarrhea (PED) status, as well as reproductive and production measures including abortion rate, total pigs born, stillbirths, mummies and wean-to-service intervals.
“These variables were used to train machine learning models to search for combinations of factors associated with nursery mortality,” he says. “Predictions from multiple algorithms were combined into a single model.”
Different models identify different patterns and produce different types of errors, so Cardoso says combining them produced more reliable estimates than using the models individually. The final calibrated model achieved a correlation of 0.672 between predicted and observed mortality in an independent test and correctly identified approximately 82% of the groups.
How Can This Help Pork Producers?
“For producers, the benefit provided by the model is not simply knowing which groups may have higher mortality, but rather having that information early enough to act,” Cardoso says. “Instead of applying the same level of intervention to every group, farms could direct labor and veterinary resources toward groups of pigs predicted to have the greatest need.”
Cardoso says knowing the expected level of mortality could trigger closer observation after placement, additional diagnostic testing, review of vaccination or treatment plans, reinforcement of biosecurity practices, environmental checks, or adjustments in staffing and feed-management strategies.
“The financial benefits will vary among farms, but the principle is straightforward,” he points out. “Preventing even a small number of deaths conserves the production costs already invested in each pig through breeding, farrowing, feed, labor, transportation and healthcare. Earlier identification of groups predicted to experience greater mortality may also reduce emergency treatments, improve labor efficiency and limit the performance losses that often accompany disease challenges.”
Transition from Reaction to Anticipation
This technology is not intended to replace stockmanship or veterinary judgment, he says, but instead offers an additional layer of information.
“Machine learning models can help producers transition from reacting to nursery mortality to anticipating it, while utilizing data that they are already collecting,” Cardoso says.
As more data becomes available, Cardoso and the SwinalytIQ team will continue evaluating and refining the model to improve its reliability and determine how it can best support decision-making across production systems.
“Data from additional farms will be used to determine how consistently the model performs across different health conditions, management practices and levels of data quality,” he says.
