An ensemble learning approach to identify pastured poultry farm practice variables and soil constituents that promote Salmonella prevalence

Abstract

Animal sourced foods including contaminated poultry meat and eggs contribute to human non-typhoidal salmonellosis, a foodborne zoonosis. Prevalence of Salmonella in pastured poultry production systems can lead to contamination of the final product. Identification of farm practices that affect Salmonella prevalence is critical for implementing control measures to ensure the safety of these products. In this study, we developed predictive models based predominantly on deep learning approaches to identify key pre-harvest management variables (using soil and feces samples) in pastured poultry farms that contribute to Salmonella prevalence. Our ensemble approach utilizing five different machine learning techniques predicts that physicochemical parameters of the soil and feces (elements such as sodium (Na), zinc (Zn), potassium (K), copper (Cu)), electrical conductivity (EC), the number of years that the farms have been in use, and flock size significantly influence pre-harvest Salmonella prevalence. Egg source, feed type, breed, and manganese (Mn) levels in the soil/feces are other important variables identified to contribute to Salmonella prevalence on larger (≥3 flocks reared per year) farms, while pasture feed and soil carbon-to-nitrogen ratio are predicted to be important for smaller/hobby (<3 flocks reared per year) farms. Predictive models such as the ones described here are important for developing science-based control measures for Salmonella to reduce the environmental, animal, and public health impacts from these types of poultry production systems.

Keywords

• Pastured poultry farm
• Salmonella
• Ensemble learning
• Food safety