Performance of Conditional Random Forest and Regression\nModels at Predicting Human Fecal Contamination of Produce Irrigation\nPonds in the Southeastern United States

Irrigating\nfresh produce with contaminated water contributes to\nthe burden of foodborne illness. Identifying fecal contamination of\nirrigation waters and characterizing fecal sources and associated\nenvironmental factors can help inform fresh produce safety and health\nhazard management. Using two previously collected data sets, we developed\nand evaluated the performance of logistic regression and conditional\nrandom forest models for predicting general and human-specific fecal\ncontamination of ponds in southwest Georgia used for fresh produce\nirrigation. Generic Escherichia coli served as a general fecal indicator, and human-associated <i>Bacteroides</i> (HF183), crAssphage, and F+ coliphage genogroup\nII were used as indicators of human fecal contamination. Increased\nrainfall in the previous 7 days and the presence of a building within\n152 m (a proxy for proximity to septic systems) were associated with\nincreased odds of human fecal contamination in the training data set.\nHowever, the models did not accurately predict the presence of human-associated\nfecal indicators in a second data set collected from nearby irrigation\nponds in different years. Predictive statistical models should be\nused with caution to assess produce irrigation water quality as models\nmay not reliably predict fecal contamination at other locations and\ntimes, even within the same growing region.