Prediction of Salmonella presence and absence in agricultural surface waters by artificial intelligence approaches

Abstract

The purpose of this study was to evaluate the performance of artificial intelligence tools for the prediction of Salmonella presence and absence in agricultural surface waters based on the population of microbiological indicators (total coliform, generic Escherichia coli, and enterococci) and physicochemical attributes of water (air and water temperature, conductivity, ORP, pH, and turbidity). Previously collected data set from six agricultural ponds monitored for two growing seasons were used for analysis. Classification algorithms including artificial neural networks (ANNs), the nearest neighborhood algorithm (kNN), and support vector machines (SVM) were trained and tested with a 539-point data set for optimum prediction accuracy. Classification accuracy performances were validated with data set (400 samples) collected from different agricultural surface water sources. All tested algorithms yielded the highest accuracy around 75 ± 1% for generic E. coli followed by enterococci (65 ± 5%) and total coliform (60 ± 10%). Classifiers calculated 6-15% higher accuracy ranging from 62 to 66% for turbidity than all other tested physicochemical attributes.Based on E. coli populations measured in other water sources, trained algorithms predicted the presence and absence of Salmonella with an accuracy between 58.15 and 59.23%. The classification performance of ANN, kNN, and SVM algorithms are encouraging for the prediction of Salmonella in agricultural surface waters.

Predictive Models May Complement or Provide an Alternative to Existing Strategies for Assessing the Enteric Pathogen Contamination Status of Northeastern Streams Used to Provide Water for Produce Production

“…However, like the present study, the Florida study (Polat et al, 2019) noted that the models were only as good as the data used to train them, and that models built using larger, more representative datasets are needed. Since collecting water quality data can be costly, the Florida study (Polat et al, 2019) suggested that a cost-effective way to generate a dataset of sufficient size would be to pool existing datasets from different water sources (e.g., streams, ponds) and regions (e.g., Northeast, Southeast, Southwest). We agree with this recommendation, and think that such multi-regional and multi-year datasets are key to the development of realistic models that can be integrated into on-farm food safety management plans.…”

“…below, future studies will need to weigh trade-offs between model interpretability and model accuracy. Despite the aforementioned limitations of the current study and the Florida study (e.g., small sample size, small number of water sources in either the training or test datasets; Polat et al, 2019), these studies suggest that predictive models may be useful for identifying and managing food safety hazards associated with preharvest water use.…”

“…In addition to identifying learners and data types that should be the focus of future studies, this study also sought to determine if machine learning-based models are likely to be useful for managing Salmonella and pathogenic E. coli risks associated with preharvest water use. Based on the findings of this, and the only other study (Polat et al, 2019), to the author's knowledge, that used machine learners to predict enteric pathogen presence in irrigation water, predictive models may be useful for identifying when and where pathogens are likely to be present in agricultural water. Although predictive models could be an alternative to E. coli water quality testing, they could also be used to supplement ongoing E. coli-based water quality monitoring (e.g., to target sampling to sites and times with a higher risk of pathogen presence).…”

“…However, a common refrain within the food safety community is that we cannot “test our way to food safety,” and that alternatives to test-based management strategies are needed. A previous study that sampled six Florida ponds used to source water for produce production, used support vector machines (SVM), k-nearest neighbor (kNN) and neural net learners to develop univariable models to predict Salmonella presence or absence (i.e., nine models per algorithm; Polat et al, 2019 ). This study demonstrated the potential utility of predictive models for managing enteric pathogen contamination of agricultural water in the Southeast ( Polat et al, 2019 ).…”

“…A previous study that sampled six Florida ponds used to source water for produce production, used support vector machines (SVM), k-nearest neighbor (kNN) and neural net learners to develop univariable models to predict Salmonella presence or absence (i.e., nine models per algorithm; Polat et al, 2019 ). This study demonstrated the potential utility of predictive models for managing enteric pathogen contamination of agricultural water in the Southeast ( Polat et al, 2019 ). Similarly, studies conducted in the Northeastern United States have developed and validated classification tree-based models for predicting Listeria monocytogenes , and Listeria spp.…”