Modeling and Risk Assessment of Food Safety Risks to the Food Supply

NON-TECHNICAL SUMMARY: Use of predictive modeling and quantitative microbial risk assessment tools are gaining increased acceptance both by the food industry and by regulatory agencies. Despite this increased acceptance, the number of academic researchers actively involved in pioneering the use of these tools is very limited. Dr. Schaffner and his team of graduate and undergraduate students regularly use these tools to solve a wide variety of food safety problems. Examples of the sort of problems currently under investigation in Dr Schaffner\'s lab include: modeling and assessing the risk of salmonellosis posed by microwavable entrees containing raw poultry; modeling and assessing the risk posed by the growth of Salmonella in cut tomatoes; simulating the transmission and risk posed by norovirus in foodservice settings; assessing the risk of low levels of Salmonella in peanut butter; modeling and assessing the risk E. coli O157:H7 in leafy greens from field to fork; and the development of a computer-based tool for assessing the impact of statistical sampling protocols on the probability of detecting agents of bioterrorism.

<P>

APPROACH: Our standard procedure in developing predictive models and quantitative risk assessments is to start by surveying the literature, including models available in the USDA ARS Pathogen Modeling Program, and models and data available in ComBase. Available data are then converted into comparable units and combined into statistical distributions. Examples of the approach as applied to predictive modeling are illustrated in Dominguez and Schaffner (2008), but briefly, the first step is to conduct a literature search and identify studies as sources for growth rates for the relevant at different temperatures. Studies presenting growth data on the commodity of interest, as well studies presenting growth data for commodity-associated pathogen strains on laboratory media are selected for further analysis. Growth parameters are extracted directly from tables or growth curves (by superimposing a regression line over the exponential phase of growth); calculated from generation time values (Growth Rate = Ln(2) /Generation Time); or from specific growth rate values (Growth Rate = 1/Ln(10)*Specific Growth Rate). Variations in methodologies and systematic deviations are corrected as needed. Baranyi (personal communication) has also noticed systematic deviations during the creation of the ComBase database, where a large number of observations were extracted from the literature and compared. Where needed, a correction factor of Ln(2)/Log(2) is applied to reported generation times (13,19,20,37), before being converted to growth rate values. Data are typically modeled using a square-root or Ratkowsky equation relating the square-root of the bacterial growth rate and storage temperature (T). Although purely empirical, the Ratkowsky equation, and variations of it, has been used with good results to model microbial growth in many foods. Examples of the approach as applied to quantitative microbial risk assessment can be found in Montville and Schaffner (2005) but briefly literature data is collected by searching medical and biological databases for documents related to the topic under study. Ungraph software (Biosoft, Ferguson, MO) is used to convert graphical data to numerical form. Numerical data are combined wherever appropriate (i.e. where data had approximately the same range and peak). Data are translated into appropriate discrete or probability distribution functions. Numerical data are log transformed using Excel (Microsoft corporation, Redmond, WA) and histograms were generated for both literature and experimental data. The appropriate statistical distribution for each set of numerical data was determined using BestFit software (Palisade Corp, Ithaca, NY). Quantitative risk assessment models are created using Analytica (Lumina Decision Systems, Los Gatos, CA) or @risk software (Palisade Corp). Results for simulated input distributions as well as final results are obtained by running from 1,000 to 1,000,000+ iterations of the simulations. Tornado analysis can be used to determine the relative significance of the input variables.