Summary
Effective preharvest, field-level produce sampling is challenging because current practices typically yield few positive samples with fields rarely re-testing positive. Statistical theory suggests one reason is that detecting rare contamination events would require 100s to 1,000s of random samples, or targeting sampling towards higher risk locations in the fields. This project will develop and validate tools for the produce industry to evaluate exiting and improved produce field sampling plans. Specifically, we will develop a simulation model of contaminated in-field produce and the process of picking samples. We will simulate fields in four regions of the US, contaminated by fecal deposits, irrigation water, or low-level soil bacteria. We will simulate convenience, generic, and risk-based sampling plans. Results will be used to communicate to growers the number and location of samples needed to achieve a known power to detect contamination. We will validate these simulations against academic literature, industry partner data, and field-trials of controlled contamination of spinach. Our goal is to provide tools for growers to (i) develop improved sampling plans, (ii) customize those plans to their individual fields, and (iii) quantify the performance and costs of the plan – all to better identify and manage preharvest food safety risks.
Technical Abstract
Effective preharvest, field-level produce sampling is challenging because current practices (such as compositing 60 samples collected in a Z-pattern) typically yield few positive samples and only rarely do fields re-test positive. Statistical theory suggests one reason is that powerful sampling to detect rare contamination would require 100s to 1,000s of random samples and that not all sampling plans are equally effective at detecting point-source, systematic, or sporadic contamination. An alternative to current industry practice is risk-based sampling, which could incorporate recently developed and validated geospatial models that identify field areas with predicted high or low pathogen prevalence. Thus, the produce industry has a research need for tools to improve preharvest sampling plans, including risk-based sampling. This project will meet this need by developing and validating a produce field simulation model to evaluate sampling plans. These fields will be representative of (i) the Central Valley, CA; (ii) Yuma, AZ; (iii) the Delmarva Peninsula; and (iv) Upstate NY. For each field, we will simulate one of three types of contamination: (i) point sources, e.g. fecal deposits; (ii) systematic sources, e.g. contaminated irrigation water; and (iii) sporadic contamination, e.g. low-level contamination by endemic soil bacteria. We will simulate, in silico, sampling produce from these fields according to (i) convenience sampling plans, e.g. n60 composites of Z-pattern samples; (ii) improved generic plans, e.g. random, systematic, or stratified random sampling; and (iii) risk-based plans, targeting field-specific, high-risk locations. Results will be used to communicate to growers the number and location of samples needed to detect the hazard for a target proportion of contamination events (e.g., 95%). We will validate these simulations against academic literature and industry partner data, including data from n60 composite sampling of fields in the Central Valley, CA and Yuma, AZ. Finally, we will validate simulations against fieldtrials of spinach subject to controlled contamination events, by taking 450 individual samples covering entire fields and evaluating multiple samplings of those fields in silico. The goal of this project is to help growers (i) develop improved sampling plans for in-field produce, (ii) customize those plans to their individual operations, and (iii) quantify the power and costs of the new plans. These results will impact the industry by enabling proactive risk management through efficient detection of pathogens in fields.
Research Objectives
1. Simulate contamination of produce fields that are representative of commercial fields in four produce-growing regions of the United States.
2. Evaluate convenience, improved generic, and risk-based sampling plans.
3. Validate simulations against data from industry partners and academic literature. 4. Validate simulations against field-trials of controlled contamination events.
Findings & Recommendations
Simulation can provide valid results for in-field microbiological sampling. The model was considered valid because it was used to simulate experiments in both existing academic literature and newly collected inoculated field trial data, and in both cases the model results gave confidence intervals that included experimental results. This validation justified use of the model for studying the power of low prevalence, low concentration contamination scenarios in industryrelevant 1-acre and larger fields. Results confirmed that randomized sampling plans such as simple random sampling (SRS) and stratified random sampling (STRS) outperform convenience sampling plans (here we evaluated systematic, serpentine sampling) by having lower variance in the power to detect point source hazards. Current industry sampling practices are likely powerful enough to reliably detect large-scale, systematic contamination events. A generic 1-acre field simulation found that a 60 random sample composite of 3-g grabs would almost always detect uniform contamination at −1 log(CFU/g) (equivalent to 1 CFU per 100 g of product).
• Conversely, a lower-concentration systematic contamination of −4 log(CFU/g) (equivalent to 1 CFU per 10 kg of product) was rarely detected.
• Representative systematic contamination events, including contaminated irrigation water and fertilization with improperly composted manure, were modeled with intermediate levels of −3 log(CFU/g) (0.5 CFU per lb) and −0.7 log(CFU/g) (1 CFU per 5 g), with intermediate detection probabilities. Current industry sampling practices are not likely powerful enough to reliably detect point-source contamination events. A generic 1-acre field simulation found that a 60 random sample composite of 3-g grabs would fail 80% of the time to detect contamination from a fecal pellet with a 1.9-m radius spread.
• Representative field simulation for CA, AZ, and NY confirm this failure to detect.
• Representative field simulation for VA tomatoes, where point sources of contamination have been shown to have little spread, found very little detection of point-source contamination. Current industry sampling practices are not likely to detect sporadic contamination from endemic soil bacteria at very low levels (defined as −5 log(CFU/g) (1 CFU per 100 kg). Therefore, preharvest false positives are not likely. These three points, taken together, suggest that currently sampling practices are primarily useful to verify lack of major food safety failures.
• Further, it is not likely possible to test individual fields for safety without significant, likely impractical, increases in sample sizes, OR alternative sampling technologies.