Location: Egg and Poultry Production Safety Research Unit
Title: Machine learning algorithms predict Listeria prevalence and Campylobacter species on pastured poultry farms based on management practices and farm environment variablesAuthor
|
Submitted to: International Poultry Scientific Forum
Publication Type: Abstract Only Publication Acceptance Date: 11/29/2023 Publication Date: 1/29/2024 Citation: Li, X., Rothrock Jr, M.J., Oladeinde, A.A. 2024. Machine learning algorithms predict Listeria prevalence and Campylobacter species on pastured poultry farms based on management practices and farm environment variables. International Poultry Scientific Forum. IPSF:2024. Interpretive Summary: Technical Abstract: Listeria and Campylobacter infections from the consumption of poultry have significant public health and economic impacts, and there is a potential for these risks to increase as poultry management systems increase the exposure of broilers to the environment. To investigate how broiler management practices and the farm environment effect foodborne pathogen ecology, management, environmental, and microbiological data collected from 11 pastured poultry farms were analyzed using three machine learning (ML) algorithms: (1) random forest classification, (2) logistic regression classification, and (3) XGBoost. The goal of this work was to apply these ML methods using broiler management farm environment variables to: (1) Predict Listeria prevalence and compare the performance of these new ML models with the previously used predictive microbiological model, and (2) predict Campylobacter species recovered throughout pastured poultry production using a novel XGBoot algorithm. The random forest and logistic regression classification ML models identified the number of years of farming, broiler flock age, and sample types (e.g., feces, soil, whole carcass rinse) as the most significant parameters in predicting Listeria prevalence. In comparison to our previous non-ML based predictive models, we have primarily addressed the imbalanced sample types and applied newly developed transformation and normalization methods to improve the model predictive performance. Moreover, the logistic regression classification model confirms statistically significant management practice variables, which complement the tree-based random forest classification models, indicating the necessity of applying multiple ML methods simultaneously. Furthermore, our novel XGBoost model suggests that farm environment variables such as carbon-to-nitrogen ratio, moisture, conductivity, and zinc concentration are most predictive of Campylobacter species recovered within these pastured poultry farms. This work showcases how ML can be applied to pre-harvest poultry food safety research, and how these types of algorithms can be useful for stakeholders and researchers in developing new strategies to mitigate poultry-related zoonotic pathogens. |
