agricultural water (e.g., rain events were associated with L. monocytogenes isolation from samples collected downstream of dairy farms; P = 0.002). The information presented here will enable growers to comanage their operations to mitigate the produce safety risks associated with preharvest surface water use.

While the Food Safety Modernization Act established standards for the use of surface water for produce production, water quality is known to vary over space and time. Targeted approaches for identifying hazards in water that account for this variation may improve growers' ability to address pre-harvest food safety risks. Models that utilize publicly-available data (e.g., land-use, real-time weather) may be useful for developing these approaches. The objective of this study was to use pre-existing datasets collected in 2017 (N = 181 samples) and 2018 (N = 191 samples) to train and test models that predict the likelihood of detecting Salmonella and pathogenic E. coli markers (eaeA, stx) in agricultural water. Four types of features were used to train the models: microbial, physicochemical, spatial and weather. "Full models" were built using all four features types, while "nested models" were built using between one and three types. Twenty learners were used to develop separate full models for each pathogen. Separately, to assess information gain associated with using different feature types, six learners were randomly selected and used to develop nine, nested models each. Performance measures for each model were then calculated and compared against baseline models where E. coli concentration was the sole covariate. In the methods, we outline the advantages and disadvantages of each learner. Overall, full models built using ensemble (e.g., Node Harvest) and "blackbox" (e.g., SVMs) learners out-performed full models built using more interpretable learners (e.g., tree-and rule-based learners) for both outcomes. However, nested eaeAstx models built using interpretable learners and microbial data performed almost as well as these full models. While none of the nested Salmonella models performed as well as the full models, nested models built using spatial data consistently outperformed models that excluded spatial data. These findings demonstrate that machine learning approaches can be used to predict when and where pathogens are likely to be present in agricultural water. This study serves as a proof-of-concept that can be built upon once larger datasets become available and provides guidance on Weller et al. Predicting Foodborne Pathogen Presence in Water the learner-data combinations that should be the foci of future efforts (e.g., tree-based microbial models for pathogenic E. coli).

No abstract

The Food Safety Modernization Act (FSMA) includes a time-to-harvest interval following the application of non-compliant water to pre-harvest produce to allow for microbial die-off. However, additional scientific evidence is needed to support this rule. This study aimed to determine the impact of weather on the die-off rate of E. coli and Salmonella on spinach and lettuce under field conditions. Standardized, replicated field trials were conducted in California, New York, and Spain over two years. Baby spinach and lettuce were grown and inoculated with a ∼104 CFU/mL cocktail of E. coli and attenuated Salmonella. Leaf samples were collected at 7 timepoints (0-96h) following inoculation; E. coli and Salmonella were enumerated. The associations of die-off with study design factors (location, produce type, and bacteria) and weather were assessed using log-linear and biphasic segmented log-linear regression. A segmented log-linear model best fit die-off on inoculated leaves in most cases, with a greater variation in the segment 1 die-off rate across trials [-0.46 (95% confidence interval (95% CI): -0.52, -0.41) to -6.99 (95% CI: -7.38, -6.59) log10 die-off/day] compared to the segment 2 die-off rate [0.28 (95% CI: -0.20, 0.77) to -1.00 (95% CI: -1.16, -0.85) log10 die-off/day]. A lower relative humidity was associated with a faster segment 1 die-off and an earlier breakpoint (the time when segment 1 die-off rate switches to the segment 2 rate). Relative humidity was also found to be associated with whether die-off would comply with FSMA's specified die-off rate of -0.5 log10 die-off/day. Importance The log-linear die-off rate proposed by FSMA is not always appropriate, as the die-off of foodborne bacterial pathogens and specified agricultural water quality indicator organisms appear to commonly follow a biphasic die-off pattern with an initial rapid decline followed by a period of tailing. While we observed substantial variation in the net culturable population levels of Salmonella and E. coli at each time point, die-off rate and FSMA compliance (i.e., at least a 2 log10 die-off over 4 days) appear to be impacted by produce type, bacteria, and weather; die-off on lettuce tended to be faster than that on spinach, die-off of E. coli tended to be faster than that of attenuated Salmonella, and die-off tended to become faster as relative humidity decreased. As such, the use of a single die-off rate for estimating time-to-harvest intervals across different weather conditions, produce types, and bacteria should be revised.

Pathogen contamination of agricultural water has been identified as a probable cause of recalls and outbreaks. However, variability in pathogen presence and concentration complicates the reliable identification of agricultural water at elevated risk of pathogen presence. In this study, we collected data on the presence of Salmonella and genetic markers for enterohemorrhagic E. coli (EHEC; PCR-based detection of stx and eaeA) in southwestern US canal water, which is used as agricultural water for produce. We developed and assessed the accuracy of models to predict the likelihood of pathogen contamination of southwestern US canal water. Based on 169 samples from 60 surface water canals (each sampled 1–3 times), 36% (60/169) and 21% (36/169) of samples were positive for Salmonella presence and EHEC markers, respectively. Water quality parameters (e.g., generic E. coli level, turbidity), surrounding land-use (e.g., natural cover, cropland cover), weather conditions (e.g., temperature), and sampling site characteristics (e.g., canal type) data were collected as predictor variables. Separate conditional forest models were trained for Salmonella isolation and EHEC marker detection, and cross-validated to assess predictive performance. For Salmonella, turbidity, day of year, generic E. coli level, and % natural cover in a 500–1,000 ft (~150–300 m) buffer around the sampling site were the top 4 predictors identified by the conditional forest model. For EHEC markers, generic E. coli level, day of year, % natural cover in a 250–500 ft (~75–150 m) buffer, and % natural cover in a 500–1,000 ft (~150–300 m) buffer were the top 4 predictors. Predictive performance measures (e.g., area under the curve [AUC]) indicated predictive modeling shows potential as an alternative method for assessing the likelihood of pathogen presence in agricultural water. Secondary conditional forest models with generic E. coli level excluded as a predictor showed < 0.01 difference in AUC as compared to the AUC values for the original models (i.e., with generic E. coli level included as a predictor) for both Salmonella (AUC = 0.84) and EHEC markers (AUC = 0.92). Our data suggests models that do not require the inclusion of microbiological data (e.g., indicator organism) show promise for real-time prediction of pathogen contamination of agricultural water (e.g., in surface water canals).