From reactive to proactive: impact of artificial intelligence on management and selection of livestock

Implications.

• Collection of automated measurements using sensors, cameras, and production data enables a shift from reactive to proactive management of livestock systems. Reducing time to intervention will mitigate the effect of health issues on the sustainability of farms and ranches.

• Machine learning and artificial intelligence can incorporate diverse data streams into a single model that yields a prediction for key performance indicators of animal health, well-being, and productivity.

• Farmers, ranchers, and managers can utilize real-time daily forecasts from artificial intelligence systems to devise intervention plans to embrace proactive management.

• Phenotypes measured by these sensors or predicted by artificial intelligence will allow selection for drivers of sustainability, rather than low-information indicator traits.

Introduction

As innovation and advances in technology have exponentially increased over the past several years, the amount of raw data that has been collected, stored, and analyzed has risen at a commensurate rate (Morota et al., 2018). Few, if any, industries are immune to these advances, as increases in the adoption of technology and innovation are positively related to the profitability and sustainability of most businesses, agricultural or otherwise (Geroski et al., 1993; Läpple and Thorne, 2019). Agriculture, especially in crop production, has been an early and rapid adopter of precision management techniques, which are usually powered by artificial intelligence (AI) tools to assist crop management and decision support (Lowenberg-DeBoer and Erickson, 2019). However, the livestock production sector faces additional challenges compared to precision crop farming in the on-farm application of technology, such as animal mobility and rearing conditions that are unfavorable for the use of electronic devices. As a consequence, much of the research in AI for livestock production has been in the academic setting, and a small portion of these technologies are currently deployed in the production setting (Fuentes et al., 2022).

There are several technologies that utilize AI in livestock production to quantify a current outcome in a production system. Examples of these technologies are cameras that estimate the current body weight of pigs (Vranken and Berckmans, 2017; Morota et al., 2018), microphones to assess cough incidence in wean-to-finish pig barns (Vranken and Berckmans, 2017), or accelerometers that track overall animal activity levels (Vázquez Diosdado et al., 2015). Within these sensors and cameras, machine learning models, typically based on artificial neural networks, take in signals (e.g., audio and video) captured by the device, process the input, and output a prediction for each respective outcome (e.g., body weight, cough incidence, and activity level). Livestock producers then use these predictions to decide whether to deviate from standard management protocols. While there is little doubt that the inclusion of AI-assisted support tools in livestock production systems will increase in future years, the sector is currently in an “early-adopter” phase, where only the most innovative farms and ranches have implemented monitoring systems based on AI (Vranken and Berckmans, 2017). Unfortunately, most of these technologies only currently provide a snapshot of the health and productivity of the animals at a given point in time, as opposed to a forecast of future metrics that would drive profitable decisions.

Historically, livestock producers have managed farms and ranches reactively as opposed to proactively. For example, rather than a proactive prevention or early treatment strategy, many diseases are treated reactively, often after most of the economic losses have been realized. AI can promote a behavioral change toward proactive livestock production. By combining sensors, cameras, and production data in a cohesive infrastructure, livestock producers will have predictions to anticipate detrimental factors that influence animal health, well-being, and overall economic sustainability.

Reactive Livestock Production

In worst-case scenarios, livestock managers have not evaluated the efficiency or profitability of their operation and have only reactively acted when issues arise. Others have used “close-out” reports to retrospectively evaluate the performance of their production systems. Regardless of species, these reports only consider historical data for the previously sold lot, finishing cohort, birth season, or lactation period. These reports, while still important, lead to a reactive decision-making process based on descriptive summaries rather than predictive models, where changes to management plans are enacted based on human perception of previous groups of animals. This reactive approach is generally not optimal for future groups of animals. While human intuition can be successful at times, part of the current inefficiencies in livestock industries is when this intuition fails. These historical records have tremendous value as training data for predictive models. However, the potential is not realized without a predictive, proactive approach.

Currently, in many genetic evaluations, traits have been chosen for predictions based on ease of measurement, rather than a strong relationship to biological processes or sustainability outcomes. The logical next step is to track production metrics daily as the animals are in the growing or lactation phase. However, this is also generally an inadequate approach, as daily time series from multiple health and productivity indicators are complex, and detrimental patterns are hard to recognize by manual inspection. Sensors can help farm staffs monitor barns or individual animals more closely. But, as the number of sensors used increases, the decision to change management strategies or intervene during adverse health events becomes more complex, especially if sensors provide conflicting assessments. Further, outcomes may result from interactions of the measurements from multiple sensors. Early warning systems using sensors have been previously proposed and evaluated for individual indicators (Stuth et al., 2003; Vranken and Berckmans, 2017). However, for a more complex variable such as mortality, a network of diverse data streams will be necessary to maximize forecasting accuracy and optimize intervention plans.

The few complications stated above, among many others, have contributed to the slow adoption of AI in the animal sector, which perpetuates reactive livestock production. The next frontier is proactive livestock production, and AI is integral to this required behavioral shift.

AI-Assisted Proactive Livestock Production

Sustainability, defined as profitability, social responsibility, and environmental impact, is key for livestock producers (Tedeschi et al., 2015), and animal health and well-being are key to sustainability. First and foremost, the healthiest animals will also be the most productive and economically efficient animals. Thus, prioritizing animal health and well-being in AI-assisted proactive livestock production systems benefits both farmers and the companies providing various sensors and decision support tools.

Proactive livestock production requires a system-level thought process. A farmer cannot manage what he or she cannot measure. In the example of a cough incidence sensor, the goal is to identify a respiratory illness many days before a manager would have noticed the health issue. Then, the farmer would intervene in a timelier manner to ultimately reduce mortality and improve animal performance. While in some cases increased coughing results in increased mortality, it is not a perfect indicator of the outcome variable mortality, which is most important to the farmer. Mortality in pigs is highly variable and influenced by hundreds of factors, such as temperature, age, and genetics, and cough incidence is just one piece of the puzzle (Gebhardt et al., 2020a, 2020b). To manage and quantify the direct effects of a factor on mortality rate, daily measurement of mortality is necessary. To accurately forecast changes in mortality, a diverse set of data streams comprising sensors, cameras, and production data is necessary. This holds true for many other key performance indicators outside mortality. However, as the number of data streams increases, the decision to modify management protocols becomes more complex. Machine learning forecasting models can incorporate diverse streams of data and efficiently provide decision support predictions based on highly complex relationships between indicator variables and the outcome variable.

Comprehensive, well-designed training, validation, and testing frameworks are necessary for machine learning models that have sufficient accuracy to enable proactive management and information-rich genetic prediction (Figure 1). Machine learning models that are successfully deployed can forecast deviations from normal for a given outcome variable, such as mortality, days in advance. With adequate warning, managers can devise intervention plans or consult veterinarians and anticipate increases in mortality or changes in other key outcomes. Continued research in sensors, cameras, and other automated high-throughput data streams is critical to provide a wider range of indicators for production outcomes. Further, a systems perspective allows key biological variables, such as basal metabolic rate from cameras or sensors, to be predicted in genetic evaluations, rather than indicators, such as body weights. In addition, time investment in the collection of daily production data remains important to maximize the impact of machine learning models. In the face of labor shortages, prioritizing the collection of the most impactful data that characterize animal health, system-wide efficiency, and direct economic impacts is critical to the application of these technologies in the coming years.

Figure 1.

Comprehensive illustration of a proposed machine learning framework consisting of four key phases for proactive livestock management. (A) Investigation phase: evaluation of sensors and production data for associations with key outcome variables. (B) Modeling phase: curation of training dataset and evaluation of machine learning models for predictive performance using cross-validation. (C) Testing phase: evaluation of best-performing machine learning model in the production setting for impact on key outcome variables and management staff behavior. (D) Operations phase: deployment of machine learning model across a production system and continuous monitoring of forecasting performance. Only models that have made measurable positive impacts in the testing phase should be deployed into production. Created with BioRender.com.

Conclusions

As innovation continues in livestock production, there will be an abundance of data and tools available to farmers and researchers to help manage livestock. Machine learning models and AI systems enable a shift from reactive to proactive livestock production. Additional investment of time and financial resources to collect high-quality data using sensors and cameras increases the impact of machine learning models on proactive livestock production. Adoption of AI on farms and ranches will have a measurable positive impact on animal health and well-being. This approach also allows for the measurement of more influential phenotypes to include in genetic evaluations. Ultimately, AI-enabled proactive management will improve the sustainability of livestock production in the face of a growing population and demand for high-quality animal protein across the globe.

About the Authors

Jared E. Decker is Wurdack Chair in Animal Genomics and an associate professor in the University of Missouri Division of Animal Science, Genetics Area Program, and Institute for Data Science and Informatics. Decker received his BS from New Mexico State University, where he majored in Animal Science with a minor in Biology. He earned his PhD at the University of Missouri in Genetics, with a PhD minor in Statistics. His research focuses on understanding the history of cattle breeds, improving the accuracy and breadth of genomic tests, and creating genomic tools to match cattle to their environment and management.

Corresponding author: deckerje@missouri.edu

Caleb J. Grohmann is the Production Data Scientist at Carthage Veterinary Service, Ltd. Caleb completed a PhD in Informatics and Data Science with an emphasis on Bioinformatics at the University of Missouri Informatics Institute. Grohmann received his BS at the University of Missouri–Columbia, where he majored in Animal Sciences with a minor in Agricultural Economics. He then earned his MS in Animal Sciences at the University of Illinois at Urbana–Champaign, where his research focused on estimating relationships between growth performance, carcass, and meat quality measurements and carcass and primal cut value of weaning-to-finishing pigs. His PhD research focused on the prediction of mortality episode occurrence in wean-to-finish pig farms using machine learning models.

Author Contributions

C.J.G.: conceptualization, funding acquisition, writing—original draft, writing—review & editing, visualization. J.E.D.: conceptualization, funding acquisition, writing—review & editing, supervision, resources.

Conflict of interest statement. The authors declare no real or perceived conflicts of interest.