6.2 Review of literature, data and existing models

6.2.1 Introduction

Purpose

This section describes the information available to develop a production-to-consumption exposure assessment of Salmonella in poultry, specifically broiler chicken. To date, no complete quantitative exposure assessments have been developed for this pathogen-commodity combination. This discussion considers the way in which such assessments could be developed, focusing on data requirements and possible methodologies. In addition, this report presents summaries of some of the available data and discusses the utility and limitations of existing data. This discussion is followed by a description of the exposure assessment model developed for the current FAO/WHO risk assessment of Salmonella in broiler chicken (Section 6.3). The assessment focuses on home preparation and consumption of the product.

Organization

A general model framework for conducting an exposure assessment for this pathogen-commodity combination is outlined. The framework covers the various stages on the production to consumption pathways that can be analysed as individual modules.

Each module identified is discussed in detail with respect to data requirements, possible modelling approaches and data availability. The discussions on data availability are followed by a presentation of data that has been collected for each module, together with an assessment of its use in conducting a full exposure assessment. Some of these data will be country specific, while the remainder will be general and can thus be used for the majority of countries. Collection and presentation of the data serves to illustrate the type of information that is currently available to individual member countries, and simultaneously demonstrates where information is lacking, and thus highlights critical data gaps.

The data summarized in the following sections have been collected from the literature, through the FAO/WHO calls for data, from discussions with Salmonella experts (microbiologists, veterinarians and epidemiologists) and other sources. Therefore the database is current up to the point of writing this report, but it is acknowledged that additional information may become available in the future.

Although no complete quantitative exposure assessments, from production to point of consumption, have been developed to date for Salmonella in poultry products, there are models that describe segments of poultry production and processing. These are also reviewed, together with a model for Campylobacter spp. in fresh broiler products.

6.2.2 Production-to-consumption pathways

Overall model pathway

A general aim of microbiological exposure assessment for any pathogen-commodity combination is to provide estimates of the extent of food contamination by the particular pathogen, in terms of both prevalence and numbers of organisms, together with information on commodity consumption patterns for the population of interest. Estimation of these outputs can involve consideration of a number of complex and interrelated processes that relate to all stages of the production-to-consumption pathway. Throughout this pathway, process-specific factors will influence both prevalence and numbers of organisms on the product, and hence final exposure. Such effects will be both inherently variable, due to, for example, differences in production and processing methods, and uncertain because some aspects lack appropriate information.

Given this complexity, it is often necessary to split the overall pathway into a number of distinct modules, each representing a particular stage from production to consumption (Lammerding and Fazil, 2000). Such an approach has previously been used for S. Enteritidis in eggs (USDA-FSIS, 1998), Campylobacter jejuni in fresh poultry (Fazil et al., unpublished; A.M. Fazil, personal communication) and Escherichia coli O157 in ground beef hamburgers (Cassin et al., 1998). The resulting exposure model is then integrated with a dose-response assessment to yield the risk characterization outcomes. This type of an approach has also been described as a Process Risk Model (Cassin et al., 1998).

A modular framework for an exposure assessment of Salmonella in fresh broilers is outlined in Figure 6.1. Outputs from one module are used as inputs to the subsequent module. In particular, the variables that are likely to flow from one module to the next are the prevalence of contaminated birds, carcasses or products (P) and the probable numbers of organisms per contaminated unit (N). Each module should describe, quantitatively, the changes in prevalence and numbers that occur within that step, attributable to specific factors, including, for example, the extent of cross-contamination, processing effects, the opportunity for temperature abuse, and the organism’s ability to survive or grow under the conditions described.

Figure 6.1. Modular pathway to describe the production-to-consumption pathway. Each step describes the changes to prevalence (P) and numbers of Salmonella (N) that occur within that specific module.

Individual modules of the overall pathway

The first module shown in Figure 6.1 relates to on-farm production of broilers. Here the aim is to estimate prevalence of Salmonella-positive birds (intestinal carriage of Salmonella) and the probable number of organisms per bird at the time of transportation for primary processing. This can involve taking into account various epidemiological and farm management factors that may influence these parameters.

Following farm production, the second module of the overall pathway considers transport and processing of broilers. This module models the effects of transport and the sequential processing steps on the prevalence and numbers of organisms. Important considerations are changes because of the type of transport facilities, processing methods and conditions, including changes in prevalence because of cross-contamination between negative and positive birds.

In the third module, the effects of retail distribution and storage in the home of the consumer are modelled. With respect to retail, both transportation and "on-shelf" storage are considered. Similarly, home storage includes transportation from retail source.

Preparation of the broiler chicken product is considered in the fourth module. Changes in prevalence and numbers of Salmonella present for the specific product purchased is determined by handling and cooking practices, and may include estimating impacts of cross-contamination. The outputs from this module - the estimated prevalence of contaminated products and number of organisms present in the food at time of consumption - are used in the calculation of exposure.

The amount of chicken consumed during a meal by various members of the population, and over a period, is quantified in the fifth module. This information is combined with the outputs from the previous module - i.e. the predicted likelihood that the pathogen will be in the food, and the predicted numbers of organisms present - to yield an estimate of the total number of Salmonella ingested. This information, together with the dose-response (i.e. the likelihood of illness associated with the number of Salmonella the consumer ingests), is then used to calculate the risk estimate in the risk characterization.

Data needs

Quantitative modelling of the individual exposure steps requires quantitative information. Data can be collected from a number of sources including, but not limited to:

• national surveillance data;

• epidemiological surveys;

• industrial surveys;

• research publications;

• unpublished research work; and

• government reports.

Often these data are publicly available, appearing, for example, in the published literature. However, other data, such as those collected through industry surveys, are often confidential and thus access becomes difficult. It is vital that confidence be built up between the risk managers, the assessors and those who can provide valuable data for risk assessment. Confidence building requires discussions and meetings (interactive risk communication) to discuss the type of data needed and what the data are being used for (the risk management activity). In addition, discussions provide insight into the data and how they were generated, with regards to sampling strategy, testing methods, etc. Such insight can be important for correct modelling, and thus the final results. Overall, good communications among all parties is essential.

In certain cases, adequate data may not be available. One way of dealing with this is to use expert opinion. Use of expert opinion introduces several considerations, such as how to choose experts, how to avoid biased judgement, how to elicit information and how to combine information from different experts. This area of study has been discussed by Kahneman, Slovic and Tversky (1982) and by Vose (2000).

In risk assessment, and particularly in the development of generic models (i.e. for application in general commodity production, processing, distribution and consumption management decision-making), data often come from many different sources. Two issues arise from this: first, what data to include within the model, and, second, how to combine such information. Determining what data to include involves consideration of applicability, such as whether the data are relevant for a particular country; whether the data are representative of the existing situation; and whether scientifically and statistically sound sampling and testing methods were used in the collection of the data. Furthermore, regardless of the data selection criteria, the rationale and process for selection must be transparent. The importance of transparency is also emphasized for combining data. Thus various methodologies could be used, such as weighting of information, but the assessor must clearly set out the methodology to ensure clarity and reproducibility.

Overall, data collection is probably the most resource-intensive part of modelling exposure and involves many issues that influence the quality of the risk assessment outcome.

Modelling approaches

The modelling approach used for individual stages of the overall pathway will necessarily depend on the data available to quantify input parameters and, in certain cases, the simplifying assumptions made until further data becomes available. Approaches are likely to differ from one exposure module to the next, depending on the parameters being described. Moreover, the risk management question will also determine the overall approach followed.

Static and dynamic approaches

Mathematical models can be described as either static or dynamic in nature. Dynamic models describe a process over time while static models consider the state of a process at one particular point in time. Dynamic models are generally constructed in terms of differential or difference equations that describe the rate of change of model variables over time. This approach has been used for several years to describe the spread of infectious diseases in both humans and animals (see Anderson and May, 1991). In contrast, static models consider the probability of an event happening at a certain time, such as the probability of infection from consumption of a chicken product, or over a period of time, such as the probability of introduction of infection in a year.

To date, most full quantitative risk assessments have been driven by static risk management questions and thus the output estimates of risk can usually also be termed static. However, many of the sub-modules of the assessment may involve dynamic modelling to some extent. In particular, in a microbial exposure assessment, the retail and storage step may involve dynamic modelling of the growth of the organism under conditions of temperature abuse (for an overview of bacterial growth modelling, see McMeekin et al., 1993; Baranyi and Roberts, 1995). Some modules of the pathway may require a combination of static and dynamic modelling; thus, preparation may involve a description of both growth (dynamic component) and cross-contamination (static component).

Uncertainty and variability

Modelling of each stage will have to account for the inherent variability of the specific process. The level of variability may be country or region specific, although it may be possible to generalize. Variability will arise due to causes that include seasonal effects, different procedures followed by different producers, differences in primary processing facilities, characteristics of the distribution chain, and consumption patterns. Variability cannot be reduced within a model because it describes the natural process.

In addition to variability, it will be necessary to model the uncertainty surrounding these processes. Such uncertainty will relate to the level of knowledge concerning a process and is usually reflective of the amount of available data.

Ideally, risk assessment models will explicitly separate uncertainty and variability; in essence, not separating means that one is neglected, and this can be a critical assumption with regard to further analysis. Various methods for such separation have been proposed (such as Vose, 2000), but, in reality, this often becomes complex. Ideally, factors that may be variable or uncertain, or both, should be identified and their influence on the risk assessment outcome described.

Deterministic versus stochastic models

Consideration of variability and uncertainty within exposure assessments leads to discussion of deterministic versus stochastic modelling. Deterministic models use point values (e.g. the mean of a data set) to describe inputs and thus to determine outputs. Stochastic models modify the data inputs to represent variability, uncertainty or both, using probability distributions. Probability distributions describe the relative weightings of each possible value and are characterized by a number of parameters that determine their shape, such as the mean and standard deviation or the most likely, minimum and maximum.

Consider the situation where prevalence of Salmonella infection in broilers is unknown in two countries and an expert has provided the following opinion.

In this situation, in order to capture the expert’s opinion, a triangular distribution could be used to describe the uncertainty about prevalence for each country (Figure 6.2):

Figure 6.2. Probability distributions for P1 and P2

P1 = Triangular(0.1,0.4,0.6)
P2 = Triangular(0.1,0.15,0.25)

Stochastic models are most easily implemented on a computer using Monte-Carlo simulation. The technique of Monte-Carlo simulation involves repetition of the following events a large number of times (iterations):

1. Select a value for each input from its associated probability distribution (selection is determined by the shape of the distribution) to give one combination of input values.

2. Calculate the estimate of exposure for this combination of values.

3. Store the calculated value.

The stored values are then combined to give a probability distribution for the estimate of exposure. There are numerous references in the literature (e.g. Haas, Rose and Gerba, 1999; Vose, 2000) explaining Monte Carlo techniques, and the uses of different probability distributions.

Consideration of the risk management question

Production-to-consumption exposure assessments require considerable time, data and other resources. The inherent uncertainty and variability associated with modelling individual exposure steps in a production-to-consumption exposure pathway increases its complexity. However, this type of an assessment provides the most information for risk managers when implementation of intervention strategies may be considered at any point of the food chain, and, perhaps more importantly, for identifying important information gaps. However, alternative approaches can also be useful, depending on the risk management information needs for decision-making, and the availability of adequate data. For example, the exposure assessment can begin at the point of retail sale of poultry products, using contamination data collected at that point. This approach, in effect, disregards the effects of individual factors occurring prior to retail sale that contributed to the microbiological status of the product. A similar approach has been taken to model exposure to chicken contaminated with fluoroquinolone-resistant Campylobacter (CVM, 2001). This methodology is useful when data are limited, or when the complexity of the process and associated uncertainties means that modelling becomes difficult and resource intensive, but it does not facilitate the investigation of specific control measures. In particular, the effects of mitigation at different stages throughout the exposure pathway cannot be quantified. Of course, in certain cases, the investigation of specific control strategies may not be required and thus the importance of the risk question is highlighted.

Defining the correct question is the most important part of any risk assessment. The risk question drives the model and hence the approach followed in any one module. As such, it must be stressed that this report does not present a prescribed formula for model development. Rather, general approaches are presented.

6.2.3 Primary production

The overall aim of the production module is to estimate, first, the prevalence of live broiler chickens contaminated with Salmonella at the time of leaving the farm for processing, and, second, the number of Salmonella per contaminated bird.

Sources of infection

Ideally, control of Salmonella within broiler flocks relies on knowledge of the source of infection. Possible sources include water, feed, litter, farm staff and the environment both inside and outside the broiler house (Mead, 1992). Furthermore, hatcheries are possible sources of infection, as is vertical transmission.

Many studies associated with the production of broilers have investigated factors that increase the prevalence of Salmonella. Rose et al. (1999) summarize the literature into five groups of risk factors:

• Inadequate level of hygiene, Salmonella contamination of the previous flock, with persistence inside the house.

• Contaminated day-old chicks and contaminated feed.

• The farm structure (>3 houses on the farm).

• Wet and cold seasonal conditions.

• Litter-beetle infestation of the house.

Several of the studies included within this summary focus on broiler-breeder farms rather than broiler chicken production farms. However, it may be assumed that the risk factors identified above are applicable to all poultry flocks. Of the above-listed factors, feed and hatcheries are regarded as principle sources of infection.

An ideal exposure assessment of Salmonella in broilers would include the calculation of the probability of infection from a number of possible sources. Such calculation could be based on, for example, the numbers of salmonellae a chicken is exposed to from each source and the subsequent consequences of exposure. Results from epidemiological studies could assist in this type of calculation. Given such a model, possible control strategies could be investigated in a quantitative manner.

In reality, data relating to the numbers of Salmonella organisms within feed, litter, etc., and the numbers to which a bird has been exposed, is extremely limited or simply unknown. Due to this limitation, previous microbial exposure assessments have started from the point of estimating the prevalence of contaminated, Salmonella-positive birds (Fazil et al., unnpublished; A.M. Fazil, personal communication; Hartnett et al., 2001). Although this approach inhibits the investigation of on-farm control strategies, it is currently the most likely approach that can be used for developing an exposure assessment of Salmonella in broilers.

Prevalence of Salmonella-positive birds

Prevalence in this document is defined to be the probability of a bird being infected with Salmonella. To estimate prevalence, data are required on positive (infected) birds at the point of leaving the farm for slaughter. Such data should be representative of the population of broilers and hence should cover a number of producers, flocks and seasons. Often, this type of information is not available (Hartnett et al., 2001), and, in this case, flock prevalence and within-flock prevalence can be estimated and used to generate an estimate of bird-level prevalence.

Flock prevalence

Flock prevalence is the proportion of flocks containing one or more infected - Salmonella-positive - birds. Flock prevalence is a national estimate, hence country-specific data are required. Estimation of flock prevalence requires consideration of the broiler production methods used. Differences in production practices occur not only between countries, but also within countries. For example, within the United Kingdom (and therefore probably in many other industrialized countries), many poultry companies may have their own feed mills, breeder flocks and hatcheries, thus differences between companies may exist. In addition, different breeds of birds may be used, both within a country and worldwide. Further, flock sizes, densities and the conditions under which a bird lives can also vary, such as free-range and organic birds versus mass-produced commercial birds. Many of these factors may influence the Salmonella status of a flock.

In addition to production methods, it is possible that climatic conditions may also influence flock prevalence. There is distinct seasonal effect in the outbreak of human Salmonella cases, which peak in the summer months. However, Angen et al. (1996) have showed a significant increase in prevalence of Salmonella in broiler chickens in Denmark during the autumn months of September-November, and Soerjadi-Liem and Cumming (1984) demonstrated a higher probability of Salmonella infection in Australian flocks during the cold and wet season. Climatic effects may in turn produce variation in flock prevalence between different geographical locations of a particular country.

Consequently, it is likely that flock prevalence may vary from region to region, from producer to producer, from season to season, and even from year to year. Testing all poultry before leaving the farm is impractical, and hence, data from sampling a portion of flocks are used to estimate the flock prevalence distribution, and should be defined by the associated uncertainty.

Within-flock prevalence

Within-flock prevalence refers to the proportion of birds within a single flock that are infected with Salmonella. Within-flock prevalence of Salmonella is very likely to vary from flock to flock for a number of reasons. Factors influencing such variability include the virulence of the Salmonella strain, levels of stress within the broiler house, and the occurrence of other avian diseases that may concurrently weaken resistance to Salmonella. As with flock prevalence, this variability should be represented within the exposure assessment model.

Ideally, the prevalence of Salmonella within flocks would be determined by sampling all broilers within all flocks just before leaving the farm for slaughter, but such comprehensive data collection is impractical. Therefore, as with flock prevalence, sample data could be utilized to obtain an estimate of the distribution for within-flock prevalence, together with a description of its associated uncertainty.

Note that intermittent shedding may affect the detection of Salmonella and thus birds and flocks testing negative by cloacal swabbing just prior to slaughter may nevertheless carry external contamination.

Number of Salmonella in infected birds

In addition to prevalence of Salmonella-positive broiler chickens, the number of organisms per positive bird is also a consideration, so that contamination in the processing environment can be modelled. Methods for determining the numbers of salmonellae in or on a bird can differ markedly, and a large degree of variability arises from different procedures. Results are reported in different units depending on the methodology, e.g. colony-forming units (CFU) or most probable number (MPN). In general, for risk assessments, CFU is the preferred unit of data, but MPN data can also be formulated such that they can be of use for estimation. In addition, the true number of organisms per bird is likely to vary from bird to bird. Consequently, there will be a large amount of variability in this estimate, and such variability may arise from a number of different sources.

Sampling information

For both prevalence and concentration, other information related to the collection of the data is also important. In particular, the test method used and its associated sensitivity and specificity must be considered. At the farm level, many different sample collection methods are used to determine the Salmonella status of individual broiler chickens or of the flock. For example, samples may be faeces, the caeca, cloacal swabs, and various environmental specimens. Other factors that influence results include the basis for the sampling strategy, the statistical validity of the sampling plan, information on farm management, the time of year of data collection, and the age of birds. Consequently, interpretation and combination of data can become difficult.

Summary of available data

Salmonella-positive flocks and within-flock prevalence

Tables 6.1 to 6.4 provide a summary of the flock and within-flock prevalence collected for this project. Initial observation of the data indicates that at the time of writing this report, information on Salmonella prevalence is missing for countries in a number of regions of the world. In particular, there is no or limited data for African, Asian and South American countries. Many countries within these regions provided some information for the 1995 Animal Health Yearbook (FAO-OIE-WHO, 1995), but information is restricted to details such as when the last case was reported and the level of occurrence. For other countries, no information appears to be have been reported.

For flock prevalence, in Table 6.1, much of the reported prevalence data include details of the numbers of flocks tested and the numbers of positive flocks. In cases where number of flocks tested and numbers positive are not provided, point estimates or ranges for flock prevalence are reported (e.g. studies by Mulder and Schlundt, in press; Hartung, 1999; White, Baker and James, 1997). In some cases (Tables 6.1 and 6.2), different sample materials are used to derive the flock prevalences, which introduces uncertainty. In addition, specificity and sensitivity of the various test protocols are rarely described. Few of the reports include information on how the results relate to the overall population of broiler chicken flocks, hence any variability due to, for example, differences between poultry companies (vertically integrated operations) is difficult to estimate. At the time of preparing this report, only one study (Soerardi-Liem and Cumming, 1984) had considered seasonality by sampling at different times of the year (Table 6.3).

Overall, it appears that flock prevalence is very variable between countries. However, it must be recognized that different sampling methods have been used in the different studies. In particular, in some reports environmental samples such as the litter, water and feed have been tested to determine positive flocks (for example, Lahellec et al., 1986; Jones et al., 1991a; Poppe et al., 1991). In contrast, other studies (such as Jacobs-Reitsma, Bolder and Mulder, 1994; Angen et al. 1996) involve direct testing of the broilers by examining the cloaca or caeca. Given these differences, comparison of country data must be undertaken with caution.

For within-flock prevalence, the data presented in Tables 6.3 and 6.4 indicate that there is very limited information relating to within-flock prevalence. In contrast to the flock prevalence data, several of the reported studies have considered variability among flocks, using the same sampling and testing protocols. For example, the data reported by Jacobs-Reitsma, Bolder and Mulder (1991) for the Netherlands show a large amount of variation in within-flock prevalence (a range of 0 to 80% for the caeca samples, and a range of 0 to 100% for the liver samples). Similarly, a wide range in values is reported from the Australian study by Soerjadi-Liem and Cumming (1984) (Table 6.3). As noted for some of the flock prevalence studies, the sample sizes reported in these surveys are small and thus there will be a large amount of uncertainty associated with any derived distributions for within-flock prevalence.

Number of organisms

At present, there are few data for numbers of Salmonella within infected broiler chickens (e.g. number per gram of faeces), or the numbers that may be present on feathers, skin, etc., of either birds that are infected, or birds that do not have intestinal carriage of the organisms but are surface contaminated. Most studies simply determine the presence or absence of salmonellae in the material tested. However, one study reported 100-1000 CFU of Salmonella per gram of gut content (Huis in ’t Veld, Mulder and Snijders, 1994). Humbert (1992) reported that samples of Salmonella-positive faeces in the environment contain between 10² and 10⁴ CFU salmonellae per gram. This small amount of information could be used to derive a distribution for the number of organisms, but there would be large associated uncertainty.

Data gaps

Overall, the following main data gaps have been identified for the primary production module.

• Salmonella prevalence information is available for some countries worldwide, but many of these studies give limited details of study design.

• Regions for which there is no or limited prevalence data include Africa, Asia and South America.

• No information relating to sensitivity or specificity of tests used is presented in the studies.

• There are very limited data relating to numbers of organisms per Salmonella-positive or contaminated bird.

Table 6.1. Salmonella flock prevalence data (see also Table 6.2).

Note: NS = not stated

Table 6.2. Flock prevalence and comparison of different sampling methods

NOTES: (1) 63 of 292 flocks. (2) 39 of 290 flocks

Table 6.3. Seasonal flock and within-flock prevalence of Salmonella in Australian flocks based on caecal samples (Source: Soeradi-Liem and Cumming, 1984).

Table 6.4. Within-flock prevalence and bird prevalence.

NOTES: (1) Caecal samples at 3-4 weeks, on farm. (2) Not stated if from one or more flocks, therefore considered as individual bird prevalence. (3) Cloacal swabs.

SOURCES: [1] Goren et al., 1988. [2] Jacobs-Reitsma, Bolder and Mulder, 1991. [3] Corrier et al., 1995. [4] Hadad and Mohammed, 1986.