An Initial Study on the Use of Machine Learning and Radio Frequency Identification Data for Predicting Health Outcomes in Free-Range Laying Hens

Abstract

Maintaining the health and welfare of laying hens is key to achieving peak productivity and has become significant for assuring consumer confidence in the industry. Free-range egg production systems represent diverse environments, with a range of challenges that undermine flock performance not experienced in more conventional production systems. These challenges can include increased exposure to parasites and bacterial or viral infection, along with injuries and plumage damage resulting from increased freedom of movement and interaction with flock-mates. The ability to forecast the incidence of these health challenges across the production lifecycle for individual laying hens could result in an opportunity to make significant economic savings. By delivering the opportunity to reduce mortality rates and increase egg laying rates, the implementation of flock monitoring systems can be a viable solution. This study investigates the use of Radio Frequency Identification technologies (RFID) and machine learning to identify production system usage patterns and to forecast the health status for individual hens. Analysis of the underpinning data is presented that focuses on identifying correlations and structure that are significant for explaining the performance of predictive models that are trained on these challenging, highly unbalanced, datasets. A machine learning workflow was developed that incorporates data resampling to overcome the dataset imbalance and the identification/refinement of important data features. The results demonstrate promising performance, with an average 28% of Spotty Liver Disease, 33% round worm, and 33% of tape worm infections correctly predicted at the end of production. The analysis showed that monitoring hens during the early stages of egg production shows similar performance to models trained with data obtained at later periods of egg production. Future work could improve on these initial predictions by incorporating additional data streams to create a more complete view of flock health.

Keywords: aviary; big data; eggs; poultry; smart farming; welfare.

Figure 1

The outline for the machine learning workflow adopted demonstrating the data partitioning and minority class re-sampling process. The blue areas denote processes in the workflow. The Kernel-Smoothed bootstrap sampling was performed after the testing set was partitioned from the train/validate set to ensure a complete out-of-sample test.

Figure 2

(Left) Principal Component Analysis (PCA) across all 16 features visualized using a biplot normalized to arbitrary units. UF, LF, RNG, and NB notation refers to the mean durations spent at the upper feeder, lower feeder, range, and nest box, respectively, for each of the denoted productions periods (pre-lay = PRL, peak lay = PL, late lay = LL, and end of lay = EL). (Right) A Pareto chart plotting the proportion of explained variance within each component. The blue line plots the cumulative variance across the components from left to right.

Figure 3

(Left) the ratio of minority-to-majority class samples for each target. (Right) the proportion of data points in the minority class that fall into the outlier, rare, borderline, and safe neighbor categories.

Figure 4

Two-dimensional representation of the dataset constructed using t-SNE for (left) the most balanced data set, keel bone damage, and (right) least balanced dataset, the beak damage, response with the minority class point denoted with red crosses and majority class points designated with blue circles.

Figure 5

Performance distributions plotted as boxplots for each response variable. Part (A) plots the area under the ROC curve (AUC), with the red dotted line marking the performance of a random-chance classifier. Part (B) plots the accuracy for each model. Part (C) plots the sensitivity, and part (D) plots the precision for the minority class within each response. The responses are ordered from left to right by descending mean AUC.

Figure 6

Feature importance measured through the mean decrease in accuracy for the three responses with the highest AUC values—(left) Spotty Liver Disease, (middle) Ascaridia galli, and (right) Cestode infections. The responses have been classified according to the location of the hen on either the range (RNG), upper feeder (UF), lower feeder (LF), or nest box (NB) during pre-lay (PRL), peak lay (PL), late lay (LL), or end of lay (EL).

Figure 7

Performance distributions plotted as boxplots for each model using only the PRL features. Part (A) plots the area under the ROC curve (AUC), with the red dotted line marking the performance of a random-chance classifier. Part (B) plots the accuracy for each model. Part (C) plots the sensitivity, and part (D) plots the precision for the minority class within each response. The responses are ordered from left to right by descending mean AUC.

Figure 8

Feature importance measure through the mean decrease in accuracy for the three responses with the highest AUC values—(left) Spotty Liver, (middle) Ascaridia galli, and (right) Cestode infestation. The responses have been classified according to the location of the hen on either the range (RNG), upper feeder (UF), lower feeder (LF), or nest box (NB) during pre-lay (PRL), peak lay (PL), late lay (LL), or end of lay (EL).