Predicting Egg Production Losses Using Peak Analysis and a Random Forest Model

Early warning systems and decision-making tools have the potential to forecast and simulate egg production losses, allowing producers to implement corrective actions pre-emptively. The aim of this study was to create an egg loss forecasting model for free-range egg producers. Commercial farm records comprising of 281 Australian free-range flocks dating from January 2010 until November 2021 were used. Flocks were located in 4 states, with sizes ranging from 10,000 - 40,000 laying hens. Relevant data used in the model included hen-day production, mortality, feed intake, in-house temperatures, and weather conditions. The egg production drops within the egg production curves were identified using a change detection algorithm or peak analysis. The approach outlined in figure 1 was used to train a random forest model designed to forecast egg production losses using a window size of 7,14, 21 and 28 days before egg production loss onset. The forecasting periods ranged from 0 - 5 days. The performance of each model was evaluated using the Area Under Curve (AUC), sensitivity, specificity and positive predictive values.

While the 7-day window size produced the least predictive results (77% accuracy) across all performance metrics, 14, 21 and 28 windows predicted comparable outcomes (87%, 88%, and 90% accuracies, respectively). The sensitivity, specificity, and positive predictive values showed a likewise pattern. Using the 28-day data to further train the algorithm, the mean laying rate, hen age and the laying rate change over two consecutive days were the best predictors with a mean decrease accuracy of 0.75, 0.6 and 0.6 respectively. When analysing the odds ratio, the low daily laying rate standard deviation which is an indicator of good consistent flock performance was negatively correlated to the lower risk of egg production drop. We conclude that the AUC value of the random forest model was 0.90 using a 28-day window, indicating the random forest model had an outstanding classification performance. These prediction outputs which indicate the likelihood of egg production losses may be used to implement corrective action to avoid egg production losses. Further research is required on the application of machine learning models in the development of egg production loss alert system and diagnostic tool.