In order to describe a food system, the model generated could include a number of stages. These stages are associated with the scope of the system we wish to describe and can include the events that begin prior to the product being harvested and all the events until the product is consumed. These can include contamination pathways that may exist in the environment where the product naturally exists, contamination events that could occur while the product is being processed, or, for instance, the effect that consumer behaviour may have on the product. The entire system can be described using a modular approach that focuses on specific sections of the food chain system.

In this section we provide an overview of the various stages or modules that could be developed. A convenient separation of the harvest to consumption chain can be performed along the following lines: pre-harvest models; harvest models; handling and processing models; storage and distribution models; and preparation and consumption models. The separation into these stages is done for convenience and logic purposes, but does not always have to be done in this way.

The starting point along the chain for the modelling exercise is dictated by the decision support information that is being sought by the modelling exercise. For instance, assume the goal of the exercise is to gain a better understanding of mitigations strategies. Further, assume that the system is such that the events that occur prior to harvest are beyond the control of any available technology or system. In this case, developing a pre-harvest model would be non-informative, and as a result, starting at the harvest stage would be a more reasonable approach.

Pre-harvest models are concerned with the estimation or description of the introduction or propagation of a hazard in the products within their rearing environments. In the case of seafood products, this stage would include, for instance, all the events prior to the fish, or shellfish being caught and loaded onto the shipping vessel. In the pre-harvest stage, the modelling that is done is primarily concerned with generating an estimate at the point of harvest, for the level of the hazard present on or in the product, and the frequency, or how often, the product is contaminated.

The type of information required by the modelling exercise should dictate the complexity of the modelling, similar to the decision on the starting point for the model. To illustrate, let’s assume that a hazard present in a fish species is the result of contamination occurring somewhere inland. This contamination is then transported to the watershed via some subterranean transport processes. The contamination transported to the watershed eventually gets into the fish species that are then harvested and sent to market. Modelling the contamination of fish could proceed along two potential directions: First, the contamination at the original source could be used and incorporated with complex subterranean transport modelling to arrive at an estimate of the concentration in the watershed, which would then be combined with additional modelling to estimate the level of contamination in the fish. Alternatively, the model could begin with the level in the watershed and skip the original source level and subterranean modelling.

These two approaches illustrate the logic process that should be considered prior to commencing on the development of the model. If there is no interest at the decision-making level, or no action that can be taken prior to the watershed contamination levels, then the second approach (starting at the watershed contamination level) should be used, because this approach requires the least additional modelling and associated assumptions to be taken. In some cases the data, and not the decision-making criteria, will decide on the starting level and complexity of the modelling process. For instance, if there is no data on watershed contamination levels, and there is data on original source contamination levels, then the more complex approach may be employed, even though there are no actions that can be taken prior to the watershed level.

It is not possible to provide a catalogue of all the different models that could be used in a pre-harvest module. The key concept that needs to be kept in mind, is that the ultimate goal of the module is to attempt to derive an estimate of the level and frequency of contamination at the point that the product is harvested. As long as the key outputs we are attempting to estimate are remembered, the specific models that are employed are essentially limitless.

To illustrate the type of modelling that
might be used in a pre-harvest model, the following example from the United States Food and Drug Administration (FDA) risk assessment for Vibrio
*parahaemolyticus* in shellfish is summarized (FDA,
2001).

The FDA pre-harvest module focuses on
estimating the likelihood that shellfish in a growing area will contain
disease-causing (pathogenic) strains of *V. parahaemolyticus* and the levels at
which they would be found.

The FDA pre-harvest
module starts with a review of the various pathways of introduction of V.
*parahaemolyticus* into shellfish growing areas and in shellfish. These pathways
of introduction are described and include the release of ballast water or
natural introduction by terrestrial and aquatic animals. In addition to the
introduction of the pathogen into the growing areas, the assessment also
describes and characterizes the factors that have an influence on the survival
and establishment of the pathogen in the growing areas. Issues such as
temperature, weather patterns, salinity, tidal flushing and other predictive
parameters are identified.

As an illustration of how the complexity
of the model is adjusted based on the information requirements and the data
availability, the FDA risk assessment does not model all the
parameters identified that could have an effect on the introduction or
establishment of the pathogen in the growing areas. The FDA risk assessment acknowledges that although there are a number
of factors that have been identified as potentially affecting the levels of
pathogenic *V. parahaemolyticus* in oysters at time of harvest, there is
insufficient quantitative data available to incorporate all of these factors
into a predictive model.

The FDA risk
assessment notes that in order to incorporate an environmental factor into the
simulation as a predictor of *V. parahaemolyticus* densities at harvest, it is
necessary to identify both the relationship of *V. parahaemolyticus* densities to
the parameter of interest, and the regional and temporal variation of the
parameter within the environment. As an example, if we were to incorporate the
effects of weather on *V. parahaemolyticus* densities, we would have to determine
the actual relationship between weather and *V. parahaemolyticus* density, as
well as a predictive model for the changes in weather in different locations
and at different times. Obviously, this could in some cases be
very useful; however, given the complexity that would be required and the lack
of any quantitative information upon which to base the estimates, it is not
very fruitful to pursue this modelling pathway.

The FDA risk
assessment generated a model that considers two primary components as
determinants of the level of *V. parahaemolyticus* in oysters at harvest.
According to research highlighted in the risk assessment, the effects of water
temperature and water salinity are considered to be the most important
parameters. Figure 3.1 shows the structure of the model and the parameters
considered.

Figure 3.1

**Schematic for FDA** *V. parahaemolyticus* **pre-harvest model.**

The pre-harvest module reviewed the best
available data on the relationship of total *V. parahaemolyticus* densities in
oysters (and water) versus water temperature and salinity and determined that a
study by DePaola *et al.* (1990) to be the most
appropriate. This study examined seasonal changes and collected samples from
all four regions of the United States (i.e. Northeast, Gulf Coast, Mid-Atlantic, and Pacific Northwest). The risk assessment looked
at other studies and noted that while there had been several other surveys of
V.*parahaemolyticus* between 1982 and 1995, these studies were typically
limited to specific regions and/or seasons, and few had reported quantitative
data. Typically, data on the presence or absence of detectable V.
*parahaemolyticus* is of limited value in developing a quantitative risk
assessment.

The pre-harvest FDA module generated a model that characterized the effect of
temperature on the mean log_{10} total *V. parahaemolyticus* densities.
This relationship was found to be approximately linear over the range of
environmental water temperatures. With regard to salinity, a quadratic effect
was found to be significant, suggesting that V.*parahaemolyticus* increase
with increasing salinity up to an optimal level and then decrease with increasing
salinity thereafter. There was no significant interaction between temperature
and salinity that was evident based on the data. As a result, the model used to
describe the concentration of *V. parahaemolyticus* as a function of salinity and
temperature was of this form:

log(*Vp / g*) = α + β * *TEMP* + γ_{1} * *SAL* + γ_{2} * *SAL*^{2} + ε

where TEMP
denotes temperature in °C; SAL denotes salinity in parts per thousand (ppt); ∀, ∃, (_{1}, and (_{2} are regression parameters and , is a
random normal deviate with zero mean and variance φ^{2}.

The resulting parameter estimates were reported as:

α = -2.6

β = 0.12

γ_{1} = 0.18

γ_{2} = -0.004

σ^{2} = 1.0

Figures 3.2 and 3.3 show the estimated
relationships between total *V. parahaemolyticus* densities in oysters versus
water temperature and salinity.

FIGURE 3.2

**Observed log10 V. parahaemolyticus (Vp) densities in
oysters versus water temperatureat different salinities**

FIGURE 3.3

**Observed log10 V. parahaemolyticus (Vp) densities in oysters versus salinity at different
temperatures**

The FDA *V. parahaemolyticus* risk
assessment is a good example of the critical analysis of the data and the use
of a logical thought process in order to determine the level of complexity
required by the modelling exercise. The FDA risk
assessment analysed the data and, together with experience and expert opinions,
found that the extremes of salinity (below 5 ppt) were detrimental to the
survival of *V. parahaemolyticus*. However, the influence of salinity within a
range of moderate environmental salinities (i.e. 5–35 ppt) was not as clear.
Based on the regression analysis, a quadratic relationship for V.
*parahaemolyticus* densities versus salinity within the 5–35 ppt range was found
to be consistent with the data. However, this projected effect of salinity was
not as strong as that of temperature. Within a broad range around the optimal
salinity of 22 ppt, the results of the regression suggested that the
differences in salinity actually encountered in oyster harvesting had
relatively little effect on the *V. parahaemolyticus* population.

Two considerations suggested that
neglecting the effect of salinity did not adversely affect the predictive value
of a model based on temperature alone. First, as shown in Figure 3.3, predicted
mean *V. parahaemolyticus* densities vary by less than 10 percent from the
optimal (maximum) density while salinity varies from 15 to 30 parts per
thousand (ppt). Secondly, measurements of oyster liquor
salinity at the retail level, which are strongly correlated with salinity of
harvest water, suggested that oysters are harvested from the more saline areas
of the estuaries year round. The mean oyster liquor salinity was found to be 24
ppt with a standard deviation of 6.5 ppt based on 249 samples. This study was
conducted year round with samples obtained from all regions of the United States. These two considerations suggest that the effect of
variation of salinity on predicted distributions of *V. parahaemolyticus*
densities would be minor, and modelling proceeded with temperature as the only
predictive variable.

The prediction of *V. parahaemolyticus*
densities was based on a regression analysis of the data with water temperature
as the only effect in the model. The resulting regression equation was reported
as:

log(*Vp / g*) = α + β * *TEMP* + ε

where TEMP
denotes temperature in °C, ∀ and ∃ are
regression parameters for temperature effect on mean log_{10} densities, and , is a random normal deviate with zero mean and variance φ^{2.}

Parameter estimates obtained for this equation were reported to be:

α = -1.03

β = 0.12

σ^{2} = 1.1

The predicted mean log *V. parahaemolyticus*
level versus temperature for the temperature only regression is shown in Figure
3.4. Clearly, this relationship is comparable to that which would be obtained
by fixing the salinity to a near optimal value (22 ppt) in the prediction
equation based on both water temperature and salinity. The temperature only
regression was used to model the relationship between temperature and density
of total *V. parahaemolyticus* at the time of harvest.

FIGURE 3.4

**Observed log10 V. parahaemolyticus (Vp) densities in oysters versus
water temperature**

The models that might be employed during harvest are those that describe the effect that specific harvesting practices might have on a hazard or in initiating the hazard. Given the scope of the events that could take place during harvest, it is often reasonable to collapse the harvest and pre-harvest models into one overall modelling stage. Clearly, the decision to separate or combine modelling stages is not a rigid guideline, but rather needs to be left to the discretion of the risk assessor. The risk assessor can combine the steps or separate them, taking into consideration such things as data availability and the needs for decision information and for complexity.

Similar to models developed for the harvest stage, the handling and processing models that are developed will primarily describe activities taking place during this stage that could have an effect on either the level or frequency of contamination. The level or frequency of contamination could be impacted by handling and processing steps that either kill the pathogen (either through thermal effects or some other process step), allow growth (through time-temperature conditions being made available), or allow additional contamination to be introduced (through cross-contamination). In general, it is difficult to provide specific models for handling and processing issues that are likely to be extremely varied.

Similar to other stages, the storage and distribution models are concerned with characterizing the effect that events occurring during these stages will have on the hazard. During storage and distribution, depending on the product we are assessing, contamination could occur during this stage with the hazard being introduced onto the product. This could occur, for instance, if the product is stored in an unpackaged state and environmental contamination is then allowed to enter the storage environment contaminating the product. Although the introduction of contamination onto the product is a possibility, a primary concern during this stage is the potential for microbial hazards to increase as a result of growth due to favourable time and temperature conditions, or to die off as a result of unfavourable time and temperature combinations. It is important to recognize that viruses and protozoa are generally unable to multiply in food products, and as a result, modelling growth in the post-harvest stages is usually limited to bacteria.

The modelling of contamination being introduced onto the product needs to be determined on the basis of the type of data available, or the mechanism by which the hazard is introduced. It is not possible to prescribe a model to use in order to characterize the introduction of a hazard onto a stored product. If data is available that allows a purely empirical relationship to be estimated between, for instance, storage conditions and hazard contamination level, then that empirical approach could be used. For instance, a linear regression using data on storage condition and hazard level could be used. In the absence of usable data, it may be necessary to generate models that attempt to mechanistically describe how contamination occurs during storage. For instance, a model that describes the process through which contamination moves from the storage environment onto the product. Overall, the risk assessor needs to use good modelling practice and judgment as well as the essential consideration of the need for the step to be modelled based on the importance of the event, and the decision making information that could be derived from modelling the process.

The modelling of microbial growth and
death is a field of study that has undergone extensive development over the
past decade. The field is known as predictive microbiology and uses
experimental data from sources including growth experiments conducted in the
lab on broth or real food substrates, and mathematical equations to describe
the behaviour observed in these experiments. McMeekin
*et al.*, (1993) is a good resource for a comprehensive treatment of the field of predictive
microbiology. In general, predictive microbiology estimates the growth and
death of microbial populations as a function of time, temperature, and other
environmental conditions. Typically, temperature has been the primary variable
used in these models to dictate the amount of growth that could occur; however,
researchers have developed models with numerous factors including pH, salt
concentration, lactate and other components. In developing and using models to
describe the growth and death of pathogens, the use of an increasing number of
variables to predict the behaviour of the pathogens needs to be balanced
according to the increased accuracy and the added complexity and data needed to
use the model. If the addition of several variables into the equation produces
results that are slightly more accurate but require a great deal of new data
and information that may not be immediately available for the specific product
being investigated, then obviously it would inadvisable to incorporate all
these variables in the model.

The field of predictive microbiology, expanded as a result of two large research programmes funded by the US and the UK. The US programme resulted in the release of a software package that is freely available called The Pathogen Modelling Program, which includes growth and inactivation models for several bacterial species. Food MicroModel is the software package that resulted from the UK programme and is a proprietary package for which licences need to be purchased for usage. In addition, there are other published models and data in the international literature that can be found readily using literature-searching software. A summary of the various packages and their features is given in Table 3.1.

In order to get a comprehensive
treatment of growth modelling, refer to McMeekin *et al.* (1993). We present a simple hypothetical
illustration of growth modelling below.

This example will illustrate an approach to modelling the growth
of bacteria with only temperature as a determining factor. The inclusion of
additional factors, as stated earlier, can be incorporated into the growth rate
equation if necessary. A commonly used equation to describe the growth rate of
bacteria is the expanded square root model of Ratkowsky *et al.* (1983). This
equation is shown below:

where “k” is the growth rate in generations
per unit time, T is the temperature, T_{min} and T_{max} are notational minimum and maximum
temperature for growth respectively, and b and c are regression parameters.

Figure 3.5 shows the change in the generation rate as a function of temperature. The parameters of the square root model would be estimated based on experimental data to which the model is fit. In this hypothetical case we are using hypothetical parameters, and would estimate that the optimum growth temperature for this pathogen would be estimated as approximately 32 °C, at which the generation rate is approximately 0.17 generations/minute.

Table 3.1

**Models available in software packages
(adapted from McMeekin et al., 2002 and Ross, McMeekin and Baranyi, 2000)**

Food micromodel | ||

Type of model | Organism | Factors modelled |

growth rate, lag time | Aeromonas hydrophila | Temp, pH, NaCl |

Bacillus cereus | Temp, pH, NaCl, CO_{2} | |

Bacillus licheniformis | Temp, pH, NaCl | |

Bacillus subtilis | Temp, pH, NaCl | |

Bacillus thermosphacta | Temp, pH, NaCl | |

Clostridium botulinum | Temp, pH, NaCl | |

Clostridium perfringens | Temp, pH, NaCl | |

Escherichia coli | Temp, pH, NaCl CO_{2} | |

Listeria monocytogenes | temp, pH, NaCl CO_{2}, nitrite lactate, | |

Staphylococcus aureus | Temp, pH, NaCl | |

Salmonellae | Temp, pH, NaCl nitrite | |

Yersinia enterocolitica | Temp, pH, NaCl | |

Pathogen modelling programme | ||

Type of model | Organism | Factors modeled |

growth rate, lag time, non-thermal death rate | Escherichia coli O157:H7 | temp, pH, NaCl nitrite, lactate, anaerobic, |

Listeria monocytogenes | temp, pH, NaCl nitrite, anaerobic, lactate | |

Staphylococcus aureus | Temp, pH, NaCl, nitrite, lactate | |

Salmonellae | Temp, pH, NaCl, nitrite | |

growth rate, lag time | Aeromonas hydrophila | temp, pH, NaCl nitrite, anaerobic |

Bacillus cereus | temp, pH, NaCl nitrite, anaerobic | |

Shigella flexneri | temp, pH, NaCl nitrite, anaerobic | |

Yersinia enterocolitica | temp, pH, NaCl nitrite, anaerobic | |

time to toxigenesis | Clostridium botulinum | temperature, pH, NaCl |

Food spoilage predictor | ||

Type of model | Organism | Factors modelled |

growth under fluctuating conditions, remaining shelf life | psychrotrophic pseudomonads | temperature, water activity |

Delphi | ||

Type of model | Organism | Factors modelled |

growth under fluctuating conditions | “generic” Escherichia coli | temperature, anaerobic |

Seafood spoilage predictor | ||

Type of model | Organism | Factors modelled |

Given the estimates for the generations per unit time (k), this can then be translated to the total number of generations formed at a particular temperature if the duration of time that the bacteria spend at that temperature is known. In this particular case, if we assume that the product containing the bacteria is stored at 25 °C for one hour, we can estimate the total amount of growth as follows:

- Generation rate at 25 °C = 0.09 generations/minute
- Time spent at this temperature = 1 hour = 60 minutes
- Generations formed = 0.09 gen/min × 60 min = 5.4 generations
Log growth = log (2

^{generations}) = 1.6 log

Therefore, in this case we would estimate that if the product was stored for one hour at 25 °C, there could be approximately 1.6 logs of growth. Hence, if the original contamination level was 1.0 log and the product was stored for one hour at 25 °C, we would expect the level of contamination on the product to approach approximately 2.6 logs after the storage period.

FIGURE 3.5

**Hypothetical generation rate per minute, using expanded square root
model with b = 0.02, c = 1.0, T _{min} = 10, and T_{max} = 35.**

It is important to recognize that this is a simplification of bacterial growth, and commonly a lag time occurs before growth begins. Subsequently, the growth is likely to be less than that estimated here since a portion of the one hour will be used up resolving the lag phase prior to the growth commencing. Similar to the models used to describe the generation rate as a function of temperature or other environmental factors, there are also models available that describe the rate at which the lag phase is resolved.

In the preparation and consumption section of a model, the goal is to first account for the effect that various practices during preparation may have on the hazard, and secondly, to determine how much of the hazard is consumed via the consumption of the food product.

The key components that need to be considered during preparation in order to estimate risk from the consumption of the product, is the effect that preparation practices, whether at home or at a commercial enterprise, will have on the hazard. Cooking is an important preparation practice that can have an impact on the hazard. Similarly, cross-contamination is also an important event that can occur during preparation that can have an impact on how the consumer is exposed to the hazard. There could be other specific preparation-associated events that would need to be considered based upon the hazard-product combination being studied. Evaluating these events would entail accounting for the population consuming the product and their preparation preference for the product. In Japan, for example, the consumption of raw fish is common practice; however, in many other countries fish is always cooked in some form. Similarly, the application of certain spices or fermentation practices could have an impact on the hazard and would need to be considered. Overall, the primary concern during preparation is to account for whatever steps are taken in order to prepare the product for consumption, and the effect that these steps might have on the exposure of the consumer to the hazard. The preparation steps taken could increase, decrease, or have no effect on the hazard, but the steps need to be considered and then appropriately handled (i.e. model and estimate effect or ignore if unnecessary).

Cross-contamination is a key exposure pathway that could occur during the preparation of the product. Unlike cooking, which is deliberately performed in order to prepare the product for consumption, cross-contamination is an unintentional result of the preparation process. Cross-contamination could allow the hazard to be transferred from the original product to the consumer either through contamination of other products, hands or any number of combinations. Fully describing and modelling all the possible events that could occur as a result of cross-contamination can be difficult since there can be many possible pathways. If a complete characterization is not possible in cross-contamination, one should still estimate the magnitude of the problem. In essence, if cross-contamination is estimated to be the dominant exposure pathway, then the cooking model may not need to be modelled in detail. Alternatively, if cross-contamination is estimated to be of small magnitude, then great effort may not need to be expended on fully characterizing or collecting more data on the cross-contamination issue.

Cooking is one of the most common preparation practices, and an example of how it might be modelled is illustrated here. Other preparation practices should be handled appropriately.

The primary output desired from the preparation module is the concentration of the hazard in the food product at the point of consumption. If the risk to the consumer is the result of toxins that are already in the food product prior to cooking, then as long as the toxin is heat-stable, the effect of cooking is likely to be minimal and can probably be ignored. If the toxin is not heat stable and gets de-natured as a result of the cooking process, then the degree of reduction could be estimated. This example illustrates how the reduction in the number of pathogens as a result of cooking might be estimated.

A common approach used to estimate the effects of cooking on bacterial numbers is through the use of “D” and ‘z” values.

The D-value is the time required at a specific temperature to destroy 90 percent (1 log decrease) of the population. The z-value is the temperature increase required to reduce the D-value by 90 percent, or a factor of 10. As an illustration, a D-value of 5min at 55 °C means that in order to reduce the population by 1 log, the population has to be held at 55 °C for a period of 5 minutes. If the z-value for this population is defined as 8 °C, then if the exposure temperature is raised by 8 °C (55 + 8 = 63 °C), the D-value will be reduced by 90 percent, or a factor of 10, so it will require only 0.5 minutes at 63 °C to reduce the population by 1 log.

The following illustrates an approach
used to estimate the reduction of *Campylobacter jejuni* during cooking. Although
C. *jejuni* is not a pathogen commonly associated with seafood, the approach
illustrated here can be translated to any pathogen, provided the appropriate
data is available.

The log reductions from cooking in this example were modelled based on the effects of temperature on the organism using experimentally determined D- and z-values. As described, the D-value is the time required at a specific temperature to destroy 90 percent (1 log decrease) of the population, while the z-value is the temperature increase required to reduce the D-value by 90 percent, or a factor of 10.

Blankenship and Craven (1982) studied the thermal sensitivity of C. jejuni in poultry meat. The thermal death times for a five-strain composite and strain H-840 in autoclaved ground chicken were determined (Table 3.2). The z-values for the five-strain composite and strain H-840 were reported as 6,35 °C and 5,91 °C, respectively.

Table 3.2

**Thermal death times for 5-strain C. jejuni composite (Blankenship and Craven, 1982)**

H-840z-value = 5.91 C | Five-strain compositez-value = 6.35 C | |

Temperature (deg C) | D-value (min) | D-value (min) |

49 | 20.5 | ND |

51 | 8.77 | 9.27 |

53 | 4.85 | 4.89 |

55 | 2.12 | 2.25 |

57 | 0.79 | 0.98 |

In order to estimate the log reductions at different times and temperatures, a linear regression was performed on the data. The regression used the log-transformed D-values, using an equation of the form shown in Equation 1:

Log(D) = (- a × Temp) + b | Equation 1 |

“a” and “b” are constants estimated through the regression procedure. However, within this equation the term “a” is equivalent to the inverse of the z-value. Therefore, the published z-value for the study was used and fixed while adjusting the “b” coefficient in order to provide a “least squares fit” to the data. In the current analysis only, the data for the five-strain composite in chicken meat was used in the linear regression; however, the data could be pooled and a linear regression performed on this data set as well. The results of both analyses (composite and pooled) are shown in Figure 3.6 and Figure 3.7.

FIGURE 3.6

**Linear regression using given Z-value and composite sample in chicken.
(Data from Blankenship and Craven, 1982)**

FIGURE 3.7

**Linear regression using composite and H-840 in chicken meat
(Data from Blankenship and Craven, 1982)**

The next step is to develop an estimate of the temperature of the product during cooking using the best available technique. The cooking temperature can be estimated based on experimental studies measuring the temperature in the product during cooking, or using thermodynamic equations that estimate the temperature reached in the product based on its material properties.

Once the temperature has been determined, Equation 2 and Equation 3 with parameters based on C. jejuni data can be used to estimate the D-value. Equation 2 is estimated as a result of the linear regression performed on the experimental data (see Figure 3–6).

Log(D) = (- 0.1575 × Temp) + 9.004 | Equation 2 |

The D-value at the temperature is, then, simply the log transform of the value (Equation 3).

D = 10^{(-0.1575 ×Temp)+9.004} | Equation 3 |

Finally, given the D-value, and recalling the definition of the D-value given earlier, the log reduction that would occur at that temperature for a given period of time (t) could be estimated using Equation 4.

Equation 4 |

The final step needed to estimate the amount of the hazard to which the consumer is exposed, is to obtain information on the consumption of the food product under consideration. This data includes serving sizes, amount consumed on a daily or annual basis, and the frequency that the product is consumed. The specific type of information will depend on the question addressed by the assessment and the type of data actually available.

If we assume that a certain fish species has a toxin concentration of 5 mg/100 g, and an estimate of the amount of toxin consumed by a population per year is required, then the need for the consumption information can be seen from this simple illustration.”

Concentration = 5 mg/100 grams, individual serving size = 150 grams, number of times consumers eat product annually = 24. From this information, the dose to which the consumer is exposed a per serving basis can be estimated as:

Per serving dose = (5 mg/100 grams) × (150 grams) = 7,5 mg/serving.

The annual exposure can then be estimated by multiplying the per serving exposure by the number of servings consumed annually:

Annual dose = (7,5 mg/serving) × (24 servings) = 180 mg

Obviously, the form of the data could be different than that shown above, for instance, rather than a per serving consumption size, the data may be reported on an annual consumption amount basis. It is often the case that data is collected for purposes other than risk estimation; as a result, the data has to be modified appropriately to extract the necessary information.

There are various sources of food consumption data that differ in terms of how the information is collected, reported and the stage along the production chain for which data is reported (raw products, retail products, consumed product). Generally, there are two types of food consumption data available and frequently used for characterizing food consumption patterns for microbiological risk assessment: food production statistics and food consumption surveys.

The most commonly available data type is food production statistics, which provide an estimate of the amount of food available to the total population. These reports are usually produced for raw or semi-processed agricultural commodities and represent the total annual amount of a commodity available for domestic consumption. The total amount available for consumption is divided by the total population in the country, which estimates the total annual quantity of food available for each person in the total population (per capita amount). The daily per capita amount can then crudely be estimated by dividing the annual amount by 365. Examples of this type of data include the FAO Food Balance Sheets and other national statistics on food production, disappearance or utilization. Because these data are available for most countries and are compiled and reported fairly consistently across countries, they can be useful in conducting exposure assessments at the international level. However, per-capita consumption statistics actually represent the food that is available for consumption and not actual quantities consumed. The losses in stores, households, private institutions or restaurants are not accounted for. Per capita consumption statistics are indirect measures of actual consumption and may overstate what is actually eaten.

The ideal data for consumption-related information comes from food intake survey studies such as those reported in the United States Department of Agriculture (USDA) Continuing Survey of Food Intakes by Individuals. This data tends to capture the amount of a specific type of food consumed on an eating occasion, and even separates the data into consumption statistics according to sex and age, which can be important depending on the hazard. These surveys usually include a representative sample of individuals from which consumption for the total population or specific population subgroups may be extrapolated. Typically, the surveys are short in duration (one to several days for each survey participant), but they provide much more detailed and specific information about the types of food consumed. Unfortunately, the surveys tend to be expensive endeavours, and as a result, food consumption surveys are conducted by a limited number of countries.

In addition to these two types of data sets, another source of data comes from retail food purchase reports. This data provides detailed information about specific food products that is often lacking from food consumption surveys and can complement the other data types.