6.2.5 Retail, distribution and storage

The aim of the retail, distribution and storage module is to estimate the change in numbers of Salmonella on broilers after processing and before preparation and consumption by the consumer.

Retail, distribution and storage steps

When considering distribution and storage of broilers, it is assumed that the broilers are already dressed, chilled or frozen, and ready for supply. Storage can mean storage at the processing plant prior to distribution, storage at the retail outlet or central distribution centre, and storage in the home.

The distribution and storage of processed broilers can influence the bacterial load on the meat. If broiler chickens are not packaged individually, cross-contamination can occur, increasing the prevalence of salmonellae within a batch. These bacteria can also multiply as a function of the temperature, the nutrient conditions, moisture content and pH of their environment. Hence there are several variables that can influence the contamination of an individual broiler by the time it is cooked in the home, including:

• The prevalence and numbers of salmonellae on finished broiler chickens.

• The conditions of storage, including:

  - storage temperature;
  - relative humidity and broiler moisture;
  - muscle pH;
  - whether pre-packed or unpacked; and
  - storage density.

• The conditions of distribution, especially

  - external temperature during:

  loading,
  transport, and
  delivery.

Data requirements and models available

There are several variables that may influence the prevalence and level of salmonellae on broiler chickens during retail, distribution and storage. For a general risk assessment framework, it is important to recognize the potential consequences of these variables in the production-to-consumer food chain. Factors such as likely temperature abuse conditions at any one stage can be utilized to model potential growth. For this, it is necessary to use predictive models that estimate the likely outcome of changes in the environmental conditions that the Salmonella experience. Data requirements for this purpose can be split into two main areas: choice of suitable predictive models, and the measurement of environmental changes during the retail, distribution and storage chain. In addition, studies that provide data on prevalence or numbers of organisms at retail are important in validating predictive modelling of the food chain.

Microbiological models can differ in mathematical complexity, but a complex model may not necessarily be the best choice to answer a particular risk management question (van Gerwen, 2000). The need for an accurate prediction needs to be offset by a consideration of whether the model is easy to use, whether it is robust and precise, and whether it has been validated against independent data. For example, if the objective of a risk assessment exercise is to demonstrate the most significant risk factors in a process, a simple model may have advantages over a complex model. However, if an accurate prediction of bacterial numbers is necessary, a more complex and accurate model may be preferable. In the choice of a suitable model, one must also consider the quality of the data that is going to be used to generate a prediction. If the temperature data on a process are poor, it may not be appropriate to use a complex model for the predictions. Often this can lead to a misinterpretation of the accuracy of the final prediction. The most appropriate model would be the simplest model possible for a given purpose and the given data quality, providing that it is validated and precise. A good model should also be subjected to an analysis method that quantifies the accuracy and bias of its predictions (Buchanan and Cygnarowicz, 1990). Ideally, a model should be both accurate and unbiased.

Models used in risk assessment must adequately reflect reality (Ross, Baranyi and McMeekin, 1999; Ross, Dalgaard and Tienungoon, 2000). Before predictive models are used in exposure assessment, their appropriateness to that exposure assessment and overall reliability should be assessed.

It is always possible to create a model that perfectly describes the data, simply by having a sufficiently complex model (Zwietering et al., 1991), but such models lack generality and would be less useful for predicting responses in other situations.

Two complementary measures of model performance can be used to assess the ‘validity’ of models (Ross, Baranyi and McMeekin, 1999; Ross, Dalgaard and Tienungoon, 2000). These measures have the advantage of being readily interpretable. The ‘bias factor’ (B_f) is a multiplicative factor by which the model, on average, over- or under-predicts the response time. Thus, a bias factor of 1.1 indicates that the prediction response exceeds the observed, on average, by 10%. Conversely, a bias factor less than unity indicates that a growth time model would, in general, over-predict risk, but a bias factor of 0.5 indicates a poor model that is overly conservative because it predicts generation times, on average, half of that actually observed. Perfect agreement between predictions and observations would lead to a bias factor of 1.

The ‘accuracy factor’ (A_f) is also a simple multiplicative factor indicating the spread of observations about the model’s predictions. An accuracy factor of two, for example, indicates that the prediction, on average, differs by a factor of 2 from the observed value, i.e. either half as large or twice as large. The bias and accuracy factors can equally well be used for any time-based response, including lag time, time to an n-fold increase, death rate and D value. Modifications to the factors were proposed by Baranyi, Pin and Ross (1999). As discussed above, typically, the accuracy factor will increase by 0.10-0.15 for every variable in the model. Thus, an acceptable model that predicts the effect of temperature, pH and water activity on growth rate could be expected to have A_f = 1.3-1.5. Satisfactory B_f limits are more difficult to specify because limits of acceptability are related to the specific application of the model. Armas, Wynn and Sutherland (1996) considered that B_f values in the range 0.6-3.99 were acceptable for the growth rates of pathogens and spoilage organisms when compared with independently published data. te Giffel and Zwietering (1999) assessed the performance of many models for Listeria monocytogenes against seven datasets and found bias factors of 2-4, which they considered to be acceptable, allowing predictions of the order of magnitude of changes to be made. Other workers have adopted higher standards. Dalgaard (2000) suggested that B_f values for successful validations of seafood spoilage models should be in the range 0.8 to 1.3. Ross (1999) considered that, for pathogens, less tolerance should be allowed for B_f >1 because that corresponds to under-predictions of the extent of growth and could lead to unsafe predictions. That author recommended that for models describing pathogen growth rate, B_f in the range 0.9 to 1.05 could be considered good; be considered acceptable in the range 0.7 to 0.9 or 1.06 to 1.15; and be considered unacceptable if <~0.7 or >1.15.

In another approach to assessing model performance, the group of researchers involved in the development of the predictive modelling program Food MicroModelä proposed that validation could be split into two components: first, the model’s mathematical performance (error₁), and second, its ability to reflect reality in foodstuffs (error₂) (Anon., 1998). They found that the error of a single microbiological concentration record was about 0.1-0.3 log₁₀ CFU/ml. Therefore, this could be considered the standard error obtained by fitting the model. If, during comparison of the predicted data with the measured data used to generate the primary model, the standard error was greater than 0.3-0.4 log₁₀ CFU/ml, then the authors suggested that the curve should only be used with caution for any secondary modelling stage. They went on to suggest that when a quadratic response surface was fitted to predicted kinetic parameters from the primary model to create the secondary model, the statistical tests should include a measure of goodness of fit. They suggested that the aim of a good model would be to achieve a standard error of no greater than 15-20%. Other suggested statistical tests were measures of parsimony (e.g. t-test), errors of prediction (e.g. least squares) and measures of robustness (e.g. bootstrap methods). The ability of a model to reflect reality in foodstuffs (error₂) is often assessed by conducting a review of the literature for measured data describing the kinetic parameter for prediction by the model. These data must not be the data used to generate the model. Ross (1999) suggested that validation data could be subdivided into sets that reflected the level of experimental control. Hence, data generated in a highly controlled broth system would be separated from data generated in a less controlled foodstuff. In this way, he argued that the performance of the model would not be undermined by evaluation against poor quality data or unrepresentative data. For examples of the limitations and difficulties of using validation data from the literature, see McClure et al., (1997); Sutherland and Bayliss, (1994); Sutherland, Bayliss and Roberts (1994); Sutherland, Bayliss and Braxton (1995); and Walls et al. (1996). The multiplicative factors of bias and accuracy discussed previously could be equally applied to quantification of both error₁ and error₂.

The selection of a model for a microbiological phenomenon must go further than the mathematics. It is all too easy to forget that a model is only as good as the data on which it is based. Bacteria are biological cells and as such the methodology used to enumerate their numbers greatly affects the count obtained. For this reason the predictive model should be based on replicate data using recognized enumeration methods. The use of resuscitation procedures for enumeration is particularly important when the organism has been growing near its physiological limits. Here, bacteria are often in a state of environmental stress and recovery is necessary to prevent the artificial depression of bacterial numbers. The method used to generate the data must be free from experimental artefacts that might artificially increase or decrease the bacterial count.

Growth

Bacteria multiply by a simple process of cell division, known as binary fission. A single bacterial cell reaches a stage in its growth when it undergoes a process that results in the single cell dividing into two daughter cells. The growth of bacterial populations therefore follows a predictable cycle that involves a period of assimilation - called the lag phase; a period of exponential growth - called the exponential phase; and a period of growth deceleration and stasis - called the stationary phase. Growth curves are often described kinetically by three variables: initial cell number (N₀), lag time (l) and specific growth rate (m), which can also be used to determine the generation time or doubling time of the population. Note that this simple description does not take the stationary phase into account. Prediction of the stationary phase is not always necessary for risk assessment, although a maximum population density parameter is often useful as an endpoint for the prediction of the exponential phase of growth. The values of these variables change with environmental conditions, including temperature, pH, water activity (a_w), nutrient state and the presence and concentration of preservatives. Studies of the growth of bacteria can generate different types of data. Kinetic data, involving the enumeration of bacteria during the growth cycle, describe the shape of the population growth curve in response to a specific set of growth conditions. Probabilistic data, involving measurement of simple growth or no-growth characteristics of the bacterial population, describe whether or not the bacteria will grow under certain growth conditions.

Growth Models

Microbiologists recognize that not all equations that are applied to bacterial processes can be considered models. A kinetic model should have a sound physiological basis (Baranyi and Roberts, 1995). This distinction has not always been made in the literature, and the word "model" has been invariably used to describe empirically-based curve fitting exercises.

Growth models increase in complexity from primary models that describe a population response, e.g. growth rate and lag time, to secondary models that describe the effect of environmental factors on the primary response, e.g. temperature and pH.

For the growth process of bacteria, an example of a simple primary model is shown in Equation 6.1.

Where N = number of bacteria; N₀ = initial number of bacteria; m = specific growth rate; and l = lag time.

This type of model could be applied to growth data to determine the primary kinetic parameters for specific growth rate and lag time for the given set of environmental growth conditions under which the data was generated.

There are several primary models that have been used routinely to describe the growth of bacteria. Examples are the Gompertz equation (Gibson, Bratchell and Roberts, 1988; Garthright, 1991), which is an empirical sigmoidal function; the Baranyi model (Baranyi and Roberts, 1994), which is a differential equation; and the three-phase linear model (Buchanan, Whiting and Damart, 1997), which is a simplification of the growth curve into three linear components.

Secondary growth models based on primary models have been created by replacing the term for specific growth rate and the term for lag time with a function that described the change of these response variables with respect to environmental factors such as temperature, water activity and pH. Examples are the non-linear Arrhenius model - where the square root model relates the square root of the growth rate to growth temperature (Ratkowsky et al., 1982) - and the response surface model. In the case of the simple model example in Equation 6.1, an example secondary model can be used to describe the growth of a bacterial population when temperature changes (Equation 6.2).

Where N = number of bacteria; N₀ · = initial number of bacteria; m = specific growth rate; l = lag time; and ¦_TEMP = mathematical function for the effect of temperature, such as a quadratic equation.

This type of model could be applied to growth data at different temperatures and would allow the calculation of the number of bacteria after a given growth period when temperature changes during that growth period. Secondary models developed from primary models are more useful than primary models alone for the quantification of risk, providing that the environmental factors influencing growth can be measured dynamically.

Growth Models for Salmonella in Chicken Meat

An ideal growth model for Salmonella should take into account the general issues raised previously about model selection, but, in addition, it should be tailored for the product under study. The ideal growth model would aim to encompass the variable limits for temperature, pH and a_w shown in Table 6.14, for which Salmonella are estimated to grow.

In the case of Salmonella in broilers, the model either should have been developed using data describing Salmonella growth in chicken meat, or at least be validated against real product data.

Table 6.14. Limits for growth of Salmonella (ICMSF, 1996)

Table 6.15. Growth models for Salmonella

Published growth models for Salmonella predict growth as a function of temperature, pH, water activity (a_w) and previous growth conditions. Table 6.15 summarizes the basis of several models.

The models of Broughall and Brown (1984) and Thayer et al. (1987) do not appear to have been validated by the authors. Validation is included for the other four models. Gibson, Bratchell and Roberts (1988) validated their model against growth data generated using pork slurry and data published in the literature. The model predictions were in good agreement with the observed data. The greatest variance was found at the extremes of the model, with low temperature or high salt concentration. This model has the advantage of being based on a considerable quantity of experimental observations and covers a wide selection of environmental growth conditions. However, the authors did not validate the work against observed data in chicken meat. The work reported by Oscar (1999a, b and c) concluded that previous growth temperature, pH and salt concentration had little effect on the estimates of specific growth rate and lag time for Salmonella Typhimurium. The author also demonstrated that it was possible to develop models in a food matrix including chicken meat, and hence these are useful for the purposes of this exposure assessment.

Survival

Under stress conditions, bacteria will either remain in a state of extended lag or may die slowly. Studies on the survival of Salmonella under stress conditions are limited. The number of S. Enteritidis was shown to remain constant during the storage of chicken breast at 3°C under a range of modified atmospheres over a 12-day study period (Nychas and Tassou, 1996). However, growth of enterobacteriaceae, including Salmonella, on naturally-contaminated chicken meat occurred at 2°C after 3 days in 30% CO₂, and after 5 days in 70% CO₂, with numbers increasing by 3 log cycles after 15 and 23 days, respectively (Sawaya et al, 1995). These investigators noted that Salmonella composed about 12% of the total enterobacteriaceae microflora, and the proportion remained constant throughout storage. It is possible that Salmonella growth is enhanced by the presence of competitive microflora. Hall and Slade (1981) carried out an extensive study of the effect of frozen storage on Salmonella in meat. In chicken substrate, the numbers of S. Typhimurium declined by 99.99% (4 log cycles) at -15°C over 168 days, and by 99.4% (2-3 log cycles) at -25°C over 336 days. Survival data for Salmonella have been summarized by ICMSF (1996).

Model selection for exposure assessment model

When considering broiler meat as a media for growth and survival of Salmonella, several factors can be simplified. At the surface of the meat, water activity might vary as a function of air moisture, chilling conditions and packaging method, but generally falls between a_w 0.98 and 0.99. The pH varies among muscle types, but is between pH 5.7 and 5.9 for breast meat and pH 6.4-6.7 for leg meat. The skin averages pH 6.6 for 25-week-old chickens (ICMSF, 1996). Poultry meat is also a rich source of nutrients such as protein, carbohydrate and fat, with essential minerals and vitamins. Consequently, it can be assumed that the growth of Salmonella will not be limited by the lack of available nutrients and hence the growth rate will be optimal for a given temperature within the pH and a_w limits of the poultry meat.

For the purposes of a simple exposure assessment model, the change in environmental conditions could be considered solely as a change in external temperature and chicken carcass temperature. It can be assumed that the pH of a broiler chicken will be pH 6.0 and that the water activity will be 0.99. Some appropriate models that could be used to predict changes in growth rate during retail, distribution and storage are:

• For temperatures between 10°C and 30°C, the growth model of Gibson, Bratchell and Roberts, 1988.

• For temperatures between 16°C and 34°C, the growth model of Oscar, 1999b.

• For temperatures between 4°C and 9°C, the survival model of Whiting, 1993.

• There are no appropriate models for temperatures below 4°C.

For the purposes of the current exposure assessment, the model developed by Oscar (1999b) was selected. The model was developed in chicken meat slurry and therefore took account of the interactions between the bacteria and the food matrix. In addition, the model was simple and easily applied. The author also assessed the accuracy and bias of the model by measuring the relative error of predictions against:

(i) the data used to generate the model; and

(ii) new data measured using the same strain and experimental conditions, but at intermediate temperatures not used in the data set used to develop the original model.

Median relative errors for lag time were given as 0.9% and -3% for comparisons (i) and (ii), respectively, and the median relative errors for growth rate were given as 0.3% and 6.8% for comparison (i) and (ii), respectively. The predictions for either parameter were unbiased. The accuracy of the model was deemed to be within accepted guidelines, as discussed above.

Temperature data characterizing retail, distribution and storage

Providing that suitable secondary kinetic models are available, it is necessary to examine the change in the environmental conditions with time during the retail, distribution and storage chain. The most common studies involve the use of temperature probes to measure the changes in product temperature during a process. For broiler chickens, the measurement of external surface and deep muscle temperatures may be used to characterize the growth or survival of Salmonella at these locations. Sampling can be used to measure pH and water activity changes with time, but these types of study are rarely conducted. Alternatively, thermodynamic models can be used to predict the temperature of a product given the external temperature and time. To ensure the predictions are consistent with measured data, caution must be exercised when using this approach.

Temperatures in the retail, distribution and storage chain tend to become less controlled from processor to consumer. Temperature and time studies of storage at the processing plant, distribution to the retailer and storage at the retailer often remain the unpublished property of the broiler industry or retailers. Few studies, if published, carry detailed data. Temperature and time studies of transport and storage by the consumer tend to be carried out by food safety organizations and are also largely unpublished. This presents problems for risk assessment unless access to these data can be arranged. Even with access to data in commercial organizations, it is often unlikely that data will be released that characterizes poor practice.

Data requirements and the data available

Growth modelling

Calibrated equipment should always be used for measuring time and temperature profiles of processes. Studies can be of a single step, such as storage at the retail stage, or be of multiple steps. In both cases, it is important to measure the environmental temperature, the external product temperature and the internal product temperature. Profiles should be measured in more than one product and, in the case of multi-step measurements, careful notes on the start and end times of the individual steps must be kept. It is important, where possible, to follow the same product throughout a multi-step process so that measurements from one step to the other can be related. Wherever possible, data should be analysed statistically to determine the within-step and step-to-step variability. If continuous measurement is not possible using a temperature data logger, then as many real-time measurements as possible should be made using a temperature probe.

Few thermal profile data for retail, storage and distribution were provided by FAO/WHO member countries as a result of the call for data. No actual data were found in the literature, although profiles were shown in graphic form in some studies. As an example, time and temperature data were kindly provided on whole broilers by Christina Farnan (Carton Group, Cavan, Republic of Ireland). These data are summarized in Tables 6.16 and 6.17.

When carrying out a quantitative exposure assessment, it is important to access national data. Data should be requested from national broiler processors and retailers.

Table 6.16. Summary of chilled chain data from Carton Group.

SOURCE: Data supplied by Christina Farnan, Carton Group, Cavan, Republic of Ireland.

Table 6.17. Summary of frozen chain data from Carton Group.

SOURCE: Data supplied by Christina Farnan, Carton Group, Cavan, Republic of Ireland.

Transport and storage temperatures during consumer handling of products can vary greatly. In the United States of America, a study was carried out in 1999 to quantify this process (Audits International, 1999). This work is a good template for carrying out similar research in other nations. Data were generated on retail backroom storage temperature, display case temperature, transit temperature, ambient temperature in the home, home temperature and home temperature after 24 hours. Tables 6.18 and 6.19 summarize the data. These example data were not generated in chicken but may be used as a guide.

These data can be useful to estimate growth or survival, or both, in a deterministic assessment, or as a basis for probability distributions for time and temperature in stochastic modelling.

Table 6.18. Summary of consumer transport and storage study on chilled products including meat

NOTES: N/A = Not available. SOURCE: Audits International, 1999.

Table 6.19: Summary of consumer transport and storage study on frozen dairy products

NOTES: N/A =Not available. SOURCE: Audits International, 1999.

Figure 6.4. Relationship of lag time and growth rate with increasing temperature as a function of time.

To illustrate a deterministic approach, the data in Table 6.19 can be used to demonstrate the predicted effect on the growth of Salmonella in a product during transport from the retail store to the consumer’s home. For this example, let the number of salmonellae on the product be 1000 CFU at the start and assume that the temperature increases linearly over the transport period. It is also assumed that the growth of the organism starts at the beginning of the transport period rather than in the store. The Oscar growth model (1999b) can be used to calculate the predicted growth pattern. The model calculates the lag time and specific growth rate for salmonellae as a function of time and temperature. The organism cannot grow until the elapsed time exceeds the lag period. As temperature increases, the lag period decreases and the specific growth rate increases. This is shown in Figure 6.4. Until the elapsed time is equal to the lag period the numbers of bacteria are fixed at the starting number (in this case 1000 CFU). Figure 6.4 shows that after 2.5 hours the lag period has been exceeded and the organism is allowed to grow at a rate set by the specific growth rate.

To calculate the relationship shown in Figure 6.4, the steps followed were:

• The thermal profile was divided into equal time and temperature blocks of 15 minutes.

• For each block, the model was used to calculate the lag time and specific growth rate.

• The growth curve was fixed at the starting cell number until the elapsed time was greater than the lag period (2.5 hours).

• After completion of the lag period, the growth at each time and temperature block was calculated by dividing the specific growth rate by the growth period.

• The increases in bacterial numbers predicted at each time and temperature block were summed to give the final increase in numbers after completion of the thermal profile.

Figure 6.5. The predicted effect on the growth of Salmonella of temperature increase during consumer transport of product to home.

Data in Table 6.18 suggest that in a worst case scenario, a product at 14.4°C in the store could reach 36.6°C during transport over a period greater than 2 hours. Using the same approach, the effect of journey time on the growth of salmonellae can be demonstrated. Figure 6.5 shows the predicted consequences of a journey that results in a product at 14.4°C reaching 36.6°C over a 2-, 3- or 4-hour journey time.

The Oscar model (1999b) has a temperature range of 16°C to 34°C and calculations were only performed within this temperature range. It must be emphasized that predictive models should not be extrapolated beyond their boundaries.

Retail level prevalence and concentration data

Data on concentration and prevalence at the retail level could be useful as a starting point for an exposure assessment. Tables 6.20a, 6.20b and 6.20c summarize the data reported and collected to date. It is important to note, however, that study design details are lacking and the future collation of such details should be recommended.

Table 6.20a. Reported prevalence of Salmonella in poultry at retail.

NOTES: NS = not stated.

Table 6.20b. Prevalence and concentration.

Notes: (1) Immersion, no chlorine. (2) Immersion, unspecified level of chlorine present in chill water.

Table 6.20c. Numbers of Salmonella on whole carcasses at retail.

Notes: (1) MPN = Most probable number per carcass. Source: Dufrenne et al., 2001.