Modeling for Quantitative Microbial Risk Assessment Thomas P. Oscar, PhD USDA, ARS Princess Anne, MD, USA

Risk Assessment Hazard Identification Hazard Characterization Food Safety Information Predictive Microbiology Hazard Identification Hazard Characterization Risk assessment is a holistic approach to food safety that is the umbrella under which all food safety information can be organized to protect public health. Risk assessment consists of the four steps shown here. As discussed in this presentation, predictive microbiology has an important role in the first three steps of risk assessment. Exposure Assessment Risk Characterization

Hazards Chemical Physical Microbial Although many types of hazards are present in food, the focus in this presentation will be on modeling methods that have been applied to assessing the risk of microbial hazards with a specific focus on Salmonella and chicken. Microbial

Pathogen Events (growth, death, survival, removal, cross-contamination) Rare Random Variable Uncertain In contrast to chemical and physical hazards, microbial hazards in food are dynamic because of pathogen events that change the number of microbial hazards as the food moves through the risk pathway. Pathogen events in food are often rare events – meaning that they occur less than 100% of the time. They are also random events that are variable and uncertain. To properly model these pathogen events,

=Round(IF(Discrete=0,0,Pert),0) Rare Events’ Modeling Iteration 1 2 3 : 100 Discrete 1 : Pert (0,1,4) 1.8 1.2 0.2 : 2.2 Power 63.1 : Round 63 : a rare events modeling approach is needed. This involves linking distributions for incidence of the event with continuous distributions for extent of the events and then using Monte Carlo simulation to randomly sample the distributions. For microbial hazards, it is important to round the results to whole numbers because it is not possible to have a fraction of a microbe. An important feature of rare events’ modeling approach is that it allows risk assessors to properly simulate changes in pathogen incidence and number as food moves through the risk pathway. =RiskDiscrete({90,10},{0,1}) =RiskPert(0,1,4) =Power(10,Pert) =Round(IF(Discrete=0,0,Pert),0)

Risk Pathway (Unit operations and pathogen events) Packaging (Contamination) Distribution (Growth) Cooking (Death) Serving (Cross-contamination) The risk pathway is a series of unit operations and associated pathogen events. This is an example of a risk pathway that was published in the scientific literature. This study used the rare events’ modeling approach to assess the impact of post-process risk factors on the safety of chicken contaminated with Salmonella. Consumption (Dose-response) J. Food Safety (1998) 18:371-381

Initial Contamination Unit Operation Pathogen Event Incidence Extent Packaging Initial Contamination 20% 1 (0 – 3) log/bird Distribution Growth 0.5 (0.1-3.0) logs R A E V N T S M O D L I G 20% R I S K A E M N T This slide shows the input settings for the first two unit operations and pathogen events in the risk assessment model for Salmonella and chicken. The model was constructed in an Excel spreadsheet and was simulated using @Risk. The output graph pictured here shows the level of contamination after distribution versus the initial contamination at packaging. Each symbol represents an individual chicken. Chickens along the diagonal are those on which Salmonella did not grow during distribution, whereas those above the diagonal are those chickens on which Salmonella grew during distribution. Which chickens experienced growth of Salmonella and the extent of that growth was a random chance event that was properly simulated using the rare events’ modeling approach. The purple number in the upper right hand corner of this output graph shows the incidence of Salmonella contamination after distribution. J. Food Safety (1998) 18:371-381

0.9% Unit Operation Pathogen Event Incidence Extent Cooking Survival 20% -1.5 (-2 to -1) logs 0.9% The next unit operation in the risk pathway in this study was cooking. In this case, incidence refers to the % of chickens that were not cooked properly. When a chicken was cooked properly, no Salmonella survived, whereas when a chicken was not properly cooked, there was a chance that some of the Salmonella could survive the cooking process. In this simulation, Salmonella survived on only 9 of the 10,000 chickens simulated. Again, the survival of Salmonella during cooking was a random chance event. The incidence of Salmonella contamination after cooking was < 1%. J. Food Safety (1998) 18:371-381

7.0% Unit Operation Pathogen Event Incidence Extent Serving Cross-contamination 25% 2 (1 to 5)% transfer 7.0% After cooking, the next unit operation and pathogen event was cross-contamination during serving. As shown here, 25% of the chickens experienced food handling practices that resulted in potential cross-contamination of the cooked chicken with Salmonella from the uncooked chicken. These results demonstrate that there is no apparent correlation between the dose consumed and the initial level of contamination at packaging. This occurred because it was by random chance which chickens experienced Salmonella growth during distribution, Salmonella survival during cooking and Salmonella cross-contaminated during serving. The incidence of Salmonella among cooked chickens at consumption was 7.0% in this simulation. Thus, 700 consumers were exposed to Salmonella at levels that ranged from 1 cell to 310 cells. J. Food Safety (1998) 18:371-381

Unit Operation Pathogen Event Incidence Extent Normal Risk High Risk Consumption Normal Risk 80% 750 (500-1000) cells High Risk 20% 200 (50 to 350) cells Normal Risk To assess consumer response to Salmonella exposure, the dose that caused a Salmonella infection was modeled using distributions for normal risk and high risk. During simulation of the model, @Risk randomly assigned an infection dose to each chicken where 80% of the infection doses were from the distribution for normal risk and 20% were from the distribution for high risk. As shown here, it was by random chance which chickens were associated with infection doses of normal or high risk. In fact, the most highly contaminated chicken at packaging was by random chance associated with an infection dose that was of normal risk. Examples of factors that would result in an infection dose from the high risk population are: ingestion of a highly virulent strain of Salmonella, consumption of anti-acid pills that would increase the virulence potential of the ingested Salmonella and consumption of the chicken by someone with an underlying health problem. High Risk J. Food Safety (1998) 18:371-381

Relative risk of infection = (Dose Consumed ÷ Infection Dose) * 100 Higher risk! In the final step of the risk assessment, the probability of a Salmonella infection was calculated by dividing the dose consumed by the infection dose. The results shown here demonstrate that chickens with lower levels of contamination at packaging can pose a higher risk of infection than chickens with higher levels of contamination at packaging when they are temperature abused, under-cooked, and consumed by someone from the high risk population. J. Food Safety (1998) 18:371-381

Hazard Identification Cornerstone Expensive Number and Subtype Packaging The cornerstone of any risk assessment is knowing the initial distribution of the pathogen in the food. Determining the level of a microbial hazard in samples of food is time consuming and expensive and thus, realistically can only be done at one point in the risk pathway. A good place for a food company to apply hazard identification is at packaging.

Microbial Ecology Minority Unattached Attached Entrapped When determining the initial level of a microbial hazard in a food, it is important to consider the microbial ecology of the food. For example, it is important to recognize that microbial hazards are minority members of the microbial community and as a result they are not uniformly distributed in the food. In fact, most samples of a food will not contain the pathogen. It is also important to consider that the pathogen can be in various states of attachment within the food matrix and thus, the methods used must be capable of quantifying the hazard regardless of how it is associated with the food. Sampling methods such as swabbing, rinsing and sponging do not accurately quantify levels of pathogens in food.

Predictive Model (Initial Contamination) Standard incubation conditions Detection limit = 102 cells/ml Detection Time One approach for quantifying all forms of a microbial hazard in a food for risk assessment is to perform a whole sample enrichment and then develop a model that predicts the initial contamination as a function of detection time. As illustrated here, at the beginning of the enrichment period, the target pathogen will be below the detection limit of the assay. However, during incubation under standard conditions for pathogen growth, the target pathogen will multiply and eventually reach the detection limit of the assay. This method is based on the concept that there is an expected mathematical relationship between the initial number of pathogens in the food sample and the detection time during enrichment. This method enumerates unattached, attached, and entrapped pathogens in the food sample and has a range of detection from 1 viable cell to +infinity. Target pathogen (< 1/ml) J. Food Prot. (2004) 67(6):1201-1208

Final Standard Curve 95% Prediction Interval This slide shows the results of a study in which the initial number of Salmonella inoculated into 25 g of chicken meat was found to be related, in a non-linear manner, to the PCR detection time score. The range of this standard curve is from 1 viable cell to one million viable cells per 25 g of chicken meat. After its construction, this standard curve was used to construct a predictive model for use in risk assessment. This was accomplished by first defining a pert distribution for each PCR detection time score, as shown here for a PCR detection time score of 10. These distributions were then J. Food Prot. (2004) 67(6):1201-1208

Predictive Model J. Food Prot. (2004) 67(6):1201-1208 Rare Events assembled in an Excel spreadsheet. Next, a discrete distribution for incidence of PCR detection time scores was added. The model was then simulated with @Risk to determine the initial contamination for any size sample that was a multiple of the original sample size, which was 25 g. In this example, the model predicted the incidence and extent of Salmonella contamination among 100 g servings of chicken meat. The output distributions from this model can be used as input distributions in a risk assessment model that uses the rare events modeling method. J. Food Prot. (2004) 67(6):1201-1208

Exposure Assessment Develop predictive models for hazard events from hazard identification to consumption Growth Survival As mentioned before, hazard identification is too expensive to apply at more than one point in the risk pathway. Consequently, after the initial distribution of hazard in the food is determined by microbiological testing, predictive models can be developed and used to predict how the initial distribution of the microbial hazard changes from hazard identification to consumption. Cross-contamination Physical Removal

General Regression Neural Network (GRNN) Model Rare Events Model This is an example of a predictive model that uses a rare events modeling approach in combination with neural network and Monte Carlo simulation modeling methods to predict the growth and survival of Salmonella from a low initial dose on raw chicken skin with native microflora and as a function of the prevalence of three different Salmonella serotypes. The output from this model is a probability distribution that can be used directly in risk assessment to simulate the effect of temperature abuse on changes in Salmonella levels as chicken meat moves through the risk pathway. J. Food Prot. (2009) 72(10):2078-2087

Hazard Characterization Mild Illness Severe Illness Chronic Disability Infected Illness Death Severity of Illness When a food serving that is contaminated with a hazard is consumed, the response of the host falls on a continuum from no response to death. To model the response, criteria are used to classify the response into one of the responses shown on this slide. Doctor Hospital

Hazard Characterization Uniform Pathogen Food Host Human feeding trials are no longer ethical! To determine the dose of a hazard that causes a specific response one can conduct a controlled feeding trial. Typically, these trials are done with a uniform pathogen, uniform food and uniform host population and typically such feeding trials yield a sigmoid-shaped dose-response curve, such as the one shown here. Although human feeding trials are no longer ethical,

J. Infect. Dis. (1951) 88:278-289; Risk Anal. (2004) 24(1):41-49. a large human feeding trial was conducted in 1949 in which healthy male prisoners in Chicago were fed different doses of 13 strains of Salmonella in a glass of eggnog after their mid-day meal. The prisoners were then closely monitored for signs and symptoms of salmonellosis. Illness data from this feeding trial were recently modeled using the rare events method and the resulting predictive model for dose-response is shown here. In the scenario pictured here, a feeding trial involving four of the 13 strains was simulated using @Risk. Rare Events Model J. Infect. Dis. (1951) 88:278-289; Risk Anal. (2004) 24(1):41-49.

The results of this simulation indicated that when a food is contaminated with multiple hazards of different virulence that a non-sigmoid dose-response curve is obtained and thus, a different approach to modeling hazard characterization is needed. One that accounts for differences in virulence among strains of Salmonella. Risk Anal. (2004) 24(1):41-49.

Disease Triangle Modeling -2 log -1 log Pathogen Host Very young Very old Cancer Diabetes HIV Pregnant : Top clinical isolate Acid resistant : High Risk In addition, the new method needs to account for differences in resistance among consumers and effects of the food on pathogen virulence and host resistance. In this new approach to dose-response modeling, which we call disease triangle modeling, pathogen, food and host factors are classified as normal or high risk and when the factor is classified as high risk, the probability distribution for illness dose is shifted to the left by 0.5, 1 or 2 logs depending on the risk factor. High fat Anti-acid : Food -0.5 log Oscar, book chapter, in press

Disease Triangle Model The disease triangle model was created in an Excel spreadsheet and was simulated with @Risk. During simulation of the model, @Risk randomly assigns a response dose to each food serving based on the incidence of the 8 classes of hazard, food and host factors. When the illness dose is greater than the dose consumed a response does not occur otherwise a response occurs. Rare Events Model Disease Triangle Model

Relative versus Absolute Risk 100% Uncertainty 0% Uncertainty There will always be data gaps! 0% Absolute 100% Absolute Despite our best efforts now and in the future, there will always be knowledge and data gaps that will prevent us from being able to make absolute predictions of risk. Thus, the best we can do is to make relative predictions of the risk of foodborne illness. One of the best tools we have for making relative assessments of risk is the process of

Oscar, book chapter, in press Scenario Analysis What if ? Plant A Plant B 25% 10% scenario analysis. A scenario is by definition a unique set of input settings in a risk assessment model. By comparing two scenarios we can make a relative assessment of risk and use this relative assessment of risk to better inform our food safety decisions aimed at protecting public health. To illustrate this concept, lets take the fictional example of a food company that has two processing plants that are located in different regions of a country but produce the same food product that is contaminated with the same microbial hazard. Food from Plant A is more highly contaminated than food from Plant B but only food from Plant B has caused illness in consumers. Oscar, book chapter, in press

I see only one risk pathway Packaging (Contamination) Plant A Plant B I see only one risk pathway Distribution (Growth) Washing (Removal) To determine why this is so, the company hired a famous risk assessor who then created a risk assessment model to assess the situation. The model consisted of a series of unit operations and associated pathogen events. The risk assessor made the assumption that the risk pathway after packaging was the same for food from Plant A and Plant B. Cooking (Survival) Consumption (Dose-response) Serving (Contamination)

Oscar, book chapter, in press 90% 10% Plant B He then created a risk assessment model. The input settings for the hazard identification and exposure assessment module were the same for Plants A and B except that the settings for hazard incidence at packaging were changed when the scenario of Plant B was simulated. Rare Events Model Module A Oscar, book chapter, in press

Oscar, book chapter, in press Module B was used for hazard characterization and risk characterization and used the disease triangle modeling method for dose-response. The settings for Plants A and B were the same. To determine the illness rate the model was simulated for 100,000 servings of food. Rare Events Model Module B Oscar, book chapter, in press

Risk Assessment Results n = 200 replicate simulations per scenario To characterize the uncertainty of the illness rate, 200 replicate simulations for each scenario were conducted using a different random number generator seed to initiate each replicate simulation. Results of the simulations indicated that the food from Plant A was more likely to cause illness than the food from Plant B. However, this did not explain why only consumers of food from Plant B were getting sick.

Packaging (Contamination) I see two risk pathways I see data gaps! Plant A Plant B Hazard strain Time & Temp Predictive Models Consumer Surveys Distribution (Growth) Washing (Removal) So the food company hired a new risk assessor with a different vision ☼. This risk assessor, using the risk assessment model developed by the first risk assessor as a guide, had the company collect data and develop predictive models to better assess the risk of illness from Plants A and B. Cooking (Survival) Consumption (Dose-response) Serving (Contamination)

Initial Contamination Research Results Plant A Plant B Initial Contamination 25% 10% Temperature Abuse 20% 40% Washing 15% 30% Proper Cooking 90% Cross-contamination High Risk Food High Risk Pathogen 60% High Risk Host Results of this research indicated that there were important difference between Plants A and B after the food left the plant. It was discovered that food from Plant B was more often subjected to temperature abuse and cross-contamination than food from Plant A. In addition, food from Plant B was more often contaminated with a high risk strain of the hazard and was more often consumed by someone from the high risk population.

After simulating the model using the new data, the risk assessor filtered the results to remove the non-contaminated servings. He then used the filtered results to prepare summary graphs for the risk managers. Filtered Results

Oscar, book chapter, in press Exposure Assessment The first set of graphs presented to the risk managers showed the change in hazard incidence and number as a function of unit operations. Although the total hazard load was lower for Plant B at packaging, at serving, the hazard load was similar for Plants A and B because of the higher incidence of temperature abuse and cross-contamination for food from Plant B. Oscar, book chapter, in press

Hazard Characterization The second graph presented to the risk managers showed the population dose-response curves for Plants A and B and indicated that the RD50 was lower for food from Plant B than food from plant A. This occurred because food from plant B was more often contaminated with a high risk strain of the hazard and was more often consumed by someone from the high-risk population. Oscar, book chapter, in press

Risk Characterization Single Risk Pathway Multiple Risk Pathways Finally, the graph showing the relative risk of illness was presented to the risk managers and showed that the risk of illness was higher for food from Plant B than for food from Plant A. These results, which considered differences in post-process risk factors among plants, differed from those of the previous risk assessment that did not consider differences in post-process risk factors among plants. The food company was happy with the results of the second risk assessment because it provided an explanation for why food from Plant B was causing illness whereas food from Plant A was not.

Single Risk Pathway Unsafe Safe In our current approach to food safety, microbial performance standards are applied at the processing plant to identify safe and unsafe food. This approach to food safety is supported by our current approach to risk assessment that assumes a single risk pathway after food processing. However, as illustrated here, these approaches to food safety are not very good at identifying safe and unsafe food because they fail to consider differences in post-process risk factors among processing plants.

Multiple Risk Pathways Unsafe Safe In contrast, by implementing an approach to risk assessment that considers post-process risk factor differences among processing plants and that employs the rare events’ modeling approach described in this presentation, we can, as illustrated here, increase our rate of success at properly identifying safe and unsafe food at the processing plant. Safe

What is the goal ? Packaging Consumption Unsafe Cooking Safe Distribution Channel The goal of our food safety system of the future should be to maximize the public health benefit of food by ensuring both its safety and consumption. This can be accomplished by properly applying quantitative microbial risk assessment at the level of the processing plant to properly assess and manage the ‘true’ risk posed by our food supply. Consumption To maximize the public health benefit of food by ensuring its safety & consumption

Thank you for your attention!