Data challenges and practical aspects of machine learning-based statistical methods for the analyses of poultry data to improve food safety and production efficiency

Abstract

Leveraging data collected by commercial poultry requires a deep understanding of the data that are collected. Machine learning (ML)-based techniques are capable of "learning by finding" nonobvious associations and patterns in the data in order to create more reliable, accurate, explanatory, and predictive statistical models. This article provides practical definitions and examples of ML-based statistical approaches for the analysis of poultry production and poultry food safety-based data. In addition to summarizing the literature, two real examples of the supervised machine learning ensemble technique, random forest (RF), are provided with respect to predicting egg weights from a commercial layer farm and identifying the potential causes of a Salmonella outbreak from a commercial broiler facility. Specifically, as an example, for the prediction of egg weights, a training model and a test model were created, and a modification of RF was used to explore the ability to predict egg weights. Results identified multiple variables including Age, Farm Location, Body Weight, Total Eggs, Hens Housed, and House Style which were predictive of the continuous variable Egg Weight. With respect to the accuracy of the variable Egg Weight, the average error between the predicted and actual egg weight was determined to be less than 3%. With respect to broiler food safety, a relational database was constructed and a supervised RF model was developed to identify the predictors of Salmonella in a grow-out farm and associated broiler processing plant. Predictors of Salmonella that included livability, density of birds in the grow-out farm, and breeder age were identified. The task of choosing the most appropriate ML-based model(s) that accounts for the large number of variables common to the poultry industry and addresses the intricate interdependence between several production parameters and inputs while predicting multiple sequential outputs is complex. The use of ML techniques in combination with new data streams including sensors (e.g., visual and audio), IoT, and Web-scraping could offer a more comprehensive, efficient, and timely approach toward evaluating productivity, food safety, and profitability in commercial poultry.

Impact of the gastrointestinal microbiome and fermentation metabolites on broiler performance

“…For example, the production of germ-free broilers to separate host factors with a controlled introduction of specific GIT microbiota offers a means to distinguish specific GIT microbial factors ( Guitton et al, 2020 ). Advanced analytic tools such as machine learning have been proposed to identify complex associations and develop predictive models for improving food safety and production efficiency ( Pitesky et al, 2020 ). Such analytical approaches also offer an opportunity to integrate the rapidly increasing and complex GIT microbiome compositional database with microbial metabolic and fermentation activities into overall statistical modeling for broiler performance.…”

Effect of Environmental and Farm-Associated Factors on Live Performance Parameters of Broilers Raised under Commercial Tropical Conditions

“…The use of commercial data is always challenging [34,35] due to the large volumes of data, the high variability and multidimensionality, repeated measurements over time (time series data), high correlation and multicollinearity, and the lack of structure. The classical objective of statistical models is to identify the exact effect of a set of variables on another.…”

“…In contrast, using modern data analytics and machine learning (ML) tools, it is possible to explore environmental sensor data, cluster using farm traits and categories, and identify causality and variable importance that cannot be determined with traditional statistics [34][35][36]. The ML modeling methods tolerate non-linearity, non-normality, and multicollinearity because little to no assumptions are being made.…”

“…The ML modeling methods tolerate non-linearity, non-normality, and multicollinearity because little to no assumptions are being made. The ML models are designed to make the most accurate predictions, and methodologies include cross-validation for future utilization of models [34,35]. Confounding effects may not have the same impact in ML because the objective is not to identify the exact effect of a variable on another.…”

Application of Visual Radiographic Analysis of Quality Grade of Table Eggs

“…A variety of imaging techniques were used in the research for egg quality grading. Most of them use spectroscopy capturing and analysis for quality grading [21][22][23], which can only determine the superficial visual difference but cannot quantitatively analyze the entire egg. X-ray radiographic imaging techniques are gaining popularity nowadays in various fields of agriculture and food-quality evaluation [24,25].…”