Você está na página 1de 17

Risk Analysis, Vol. 27, No.

4, 2007

DOI: 10.1111/j.1539-6924.2006.00834.x

Special Issue on Campylobacter Risk Management and Assessment (CARMA)

A Risk Assessment Model for Campylobacter in Broiler Meat


Maarten J. Nauta,1 Wilma F. Jacobs-Reitsma,2 and Arie H. Havelaar1

A quantitative microbiological risk assessment model describes the transmission of Campylobacter through the broiler meat production chain and at home, from entering the processing plant until consumption of a chicken breast llet meal. The exposure model is linked to a dose-response model to allow estimation of the incidence of human campylobacteriosis. The ultimate objective of the model is to serve as a tool to assess the effects of interventions to reduce campylobacteriosis in the Netherlands. The model describes some basic mechanistics of processing, including the nonlinear effects of cross-contamination between carcasses and their leaking feces. Model input is based on the output of an accompanying farm model and Dutch count data of Campylobacters on the birds exterior and in the feces. When processing data are lacking, expert judgment is used for model parameter estimation. The model shows that to accurately assess of the effects of interventions, numbers of Campylobacter have to be explicitly incorporated in the model in addition to the prevalence of contamination. Also, as count data usually vary by several orders of magnitude, variability in numbers within and especially between ocks has to be accounted for. Flocks with high concentrations of Campylobacter in the feces that leak from the carcasses during industrial processing seem to have a dominant impact on the human incidence. The uncertainty in the nal risk estimate is large, due to a large uncertainty at several stages of the chain. Among others, more quantitative count data at several stages of the production chain are needed to decrease this uncertainty. However, this uncertainty is smaller when relative risks of interventions are calculated with the model. Hence, the model can be effectively used by risk management in deciding on strategies to reduce human campylobacteriosis.
KEY WORDS: Broiler meat; Campylobacter; farm to fork risk model; poultry processing; quantitative microbiological risk assessment.

1. INTRODUCTION Campylobacter is one of the main causative agents of gastroenteritis in Western countries. In the Netherlands the incidence of campylobacteriosis is estimated to be 80,000 cases per year in a population of 16 million.(1,2) Broiler chickens are generally reNational Institute for Public Health and the Environment (RIVM), MGB, Bilthoven, The Netherlands. 2 Wageningen UR, Animal Sciences Group (AS6), Lelystad, The Netherlands. Address correspondence to Maarten J. Nauta, National Institute for Public Health and the Environment (RIVM), MGB, PO Box 1, 3720 BA Bilthoven, The Netherlands; tel: +31302742805; fax: +31302744434; Maarten.nauta@rivm.nl.
1

garded as one of the main sources of campylobacteriosis. Campylobacter is frequently found in broiler chickens and in broiler chicken meat. Humans can thus be exposed to Campylobacter by the consumption of improperly heated broiler meat or other foods that are cross-contaminated with Campylobacter during food preparation of an (accompanying) meal with broiler meat. This article describes a farm to fork risk assessment model, which was constructed as part of the CARMA (Campylobacter Risk Management and Assessment) project,(2) which aimed to advise the Dutch government on possibilities to reduce the incidence of human campylobacteriosis in the Netherlands. The 845
0272-4332/07/0100-0845$22.00/1
C

2007 Society for Risk Analysis

846 scope of the risk model is to provide a tool to compare the effects of interventions all across the broiler chicken meat production chain. The model has been restricted to one route of exposure only, that is, exposure to salads cross-contaminated during food preparation of a meal with chicken breast llets. This route is chosen because chicken breast llet is a raw chicken product, which is frequently prepared in Dutch private homes. The presumed route of exposure from chicken breast llet is cross-contamination because the llets are usually cut and cooked. The Campylobacters on the llet are inactivated, but there is a potential for cross-contamination via hands and kitchen equipment. If the cross-contaminated food is consumed raw (like a salad) the risk of exposure is largest. In contrast to a more supercial treatment of many different routes, a detailed analysis of one route only allows the inclusion of details that may have an important impact on the (nonlinear) dynamics of the transmission of Campylobacter through the farm to fork chain. This article presents the baseline risk model, representing the present situation in the Netherlands, dened as the situation in the year 2000. (This year was chosen at the start of the project, as this was a recent year without any crisis in animal production in the Netherlands.) We discuss the model dynamics and the importance of the input data applied for the analysis. The effects of proposed risk management interventions at different points in the farm to fork chain on the baseline model results are analyzed and discussed in an accompanying article(2) and in a CARMA project report.(3) This report includes details that could not be added here. The model starts at the point where broilers enter the processing plant. A preceding model on the dynamics of Campylobacter on the farm,(4) an analysis of the costs,(5) the proposed interventions, and a cost-utility analysis with conclusions for risk management(2) are given elsewhere in this Risk Analysis issue. Our risk model builds further on studies performed in Canada,(6) the United Kingdom,(7) Denmark,(8) and for WHO/FAO.(9,10) In general, the modeling approach that we apply is more mechanistic than before, and more precisely deals with variability within ocks and between ocks. When there is a potential for cross-contamination, potential nonlinear dynamics are included in the model because this may affect the evaluation of effects of interventions. Also, the processing model(11) differentiates between Campylobacters on the birds exterior and Campy-

Nauta, Jacobs-Reitsma, and Havelaar lobacters in the birds intestines at the entrance of the processing plant, and thus allows a comparison of interventions aiming at either of these.

2. MODEL 2.1. Methodology The risk model follows the changes in the prevalence (frequency of units contaminated with Campylobacter) and numbers of Campylobacter in contaminated units from stage to stage over the food chain. Here, the unit is the potentially contaminated animal or food item, i.e., the bird exterior, the carcass, the llet, or the salad. The model describes exposure as the variability distribution of ingested numbers of Campylobacters (doses) in salads cross-contaminated by Campylobacters from chicken breast llets. A dose-response model then predicts the frequency of infection and illness in the Dutch population in a year. For the farm to fork exposure modeling we apply the modular process risk modeling methodology (MPRM).(1215) In this methodology, the food chain is split in modules, each of which can be described in terms of a basic process (bacterial processes growth and inactivation; food handling processes partitioning, mixing, cross-contamination, and removal). The level of detail of the basic process models depends on the relevance of the process step for the dynamics, the availability of data, and the specic risk management questions regarding this process step as considered in the analysis. In MPRM the rst step is to consider the basic processes to model, and then to obtain the appropriate data. Within the basic process models nonlinear effects that are considered important in the transmission dynamics are incorporated. The farm to fork risk model is a stochastic quantitative microbiological risk assessment model. Monte Carlo simulation is used to incorporate the effects variability in the analysis. First, there is variability between birds within a ock. Usually not the whole ock of birds is colonized, so the animal prevalence (the percentage of colonized birds in a ock) is lower than 100%.(4) Also, the level of Campylobacter contamination on the birds exteriors and the concentration of Campylobacter in the gut will vary between birds in the ock. Next, there is a variability between ocks reared and slaughtered in the Netherlands over a year. Not only will some ocks contain colonized birds and some

Risk Assessment Model for Campylobacter ocks none (i.e., the ock prevalence is lower than 100%), also the animal prevalence and the distribution of levels of Campylobacter between birds in a ock will vary between ocks. Hence, to describe this variability, we need probability distributions of, for example, the animal prevalence and the means of the distributions describing the variability between birds. It is not possible to combine these two sources of variability into one because the cross-contamination dynamics during processing depend on within-ock variability only. Also, there is uncertainty in the parameter values. This uncertainty represents our lack of knowledge due to scarcity of data, imprecise methods of measurement, inadequate representativity of data, etc. Like variability, this uncertainty is often represented by probability distributions. However, in our analysis uncertainty is not incorporated in the Monte Carlo simulations because at several places in the food chain it was not possible to quantify the uncertainty. Also, with the present two separate sources of variability in the analysis, incorporation of uncertainty would require a third-order Monte Carlo simulation. Our computer facilities did not allow us to do this. Uncertainty is therefore alternatively evaluated in a scenario analysis(3) (Section 3.3). The risk model is constructed as two linked models, a processing model and a consumer model, using @Risk 4.5 (Palisade, Neweld), an add on to Microsoft Excel (see Fig. 1.). In the processing model, one iteration simulates the fate of 500 birds from one ock processed in line. Some pilots have shown that this group of 500 is sufciently large to represent a broiler ock, and sufciently small to allow spreadsheet model calculations. Only infected ocks were analyzed, end results have been corrected for the ock prevalence. Model input informs the animal prevalence in that ock, and the distribution of numbers of Campylobacters on the exterior of the birds and distribution of the concentration of Campylobacter in the feces of the birds. The result gives information on within-ock variability. At the endpoint it yields the number of Campylobacters on 500 carcasses and/or chicken breast llets from that ock. To get a representative sample to describe the large between-ock variability, a set of 5,000 iterations, simulating 5,000 ocks, showed to produce reproducible results. The consumer model starts with 5,000 distributions of numbers of Campylobacter per llet, from 5,000 ocks. It samples llets from these distributions, which are assumed to be independently stored, handled, and prepared together with salad. Data

847

processing model 500 birds (one flock) processing in line scalding defeathering evisceration washing chilling cutting 5000 (infected flocks only)

500 fillets

consumer model 1 random fillet

storage preparation
*

20 fillets per flock 5000 flocks = 100,000 fillets

exposure

prob. of illness

Fig. 1. Schematic diagram of the risk model. The processing model simulates the processing of a ock, represented by 500 birds. It ends with 500 llets produced by the birds. The consumer model simulates the exposure and probability of illness after preparation and consumption of a salad in a meal with one llet. For a representative sample 20 llets are sampled per ock, for 5,000 ocks. Only meals where the salad is cut after the llet are simulated.

on consumption frequencies and meal sizes in the Netherlands are taken into account. A dose-response relation predicts the number of cases from the resulting doses. As a standard, the consumer model is run for 100,000 iterations, sampling 20 llets from the distribution of each ock, for 5,000 ocks (see Fig. 1). Finally, the risk model calculates the expected number of cases of campylobacteriosis resulting from the consumption of salad prepared after handling raw chicken breast llet. The exact numeric result is not a very important result itself, as there is quite some uncertainty in model parameter estimates, and many model simplications were necessary for the risk assessment. However, this result is used as a baseline incidence that can be used to assess and compare the

848
Table I. Overview of Units Modeled in the Stages of the Risk Model Input Unit
Farm Processing Cutting Storage Consumer preparation Ingestion/ dose-response Broiler ock Broiler ock carcass Broiler ock carcass Fillet Fillet Salad

Nauta, Jacobs-Reitsma, and Havelaar ered to be a crucial factor in predicting the effects of interventions. If published data are inadequate to fulll the data needs of the mechanistic model, (formal) expert judgment may be preferred above processing data. The latter can later be used for model validation. 2.2.1. Model Input 2.2.1.1. Prevalences. The ock prevalence and animal prevalences at slaughter are provided by the CARMA farm model.(4,17) It is found that the ock prevalence is 44.4% in the year 2000. The animal prevalence is always <100%. Of the 44.4% infected ocks, 21% has an animal prevalence <5% and 49% has a prevalence >98%. This bimodal distribution of within-ock prevalence is the consequence of the dynamics of the infection within the ocks after the introduction of Campylobacter: it takes some time before an infected bird is infectious itself and infects a group of others. Once this group gets infectious, the infection spreads rapidly over the whole ock.(4) 2.2.1.2. Concentration in the feces. In the Netherlands, between 1991 and 1994, ve fecal samples prior to hanging at the processing plant were taken per ock of 38 ocks and Campylobacter concentrations were determined, yielding a data set describing both within-ock and between-ock variability.(18) From 190 samples of 38 positive ocks, 176 were positive for Campylobacter by enrichment. Twenty-nine ocks were positive for all ve samples taken per ock. The results of the analysis of these data are used as input for the baseline model, which requires a description of within- and between-ock variability. It is assumed that within each ock the log of the concentration of Campylobacter in the feces follows a normal distribution, with a xed standard deviation of 0.73 in each ock, as derived from the data. The within-ock means are variable between ocks and are assumed to follow a normal distribution with mean 6.00 and standard deviation 1.52, also derived from the data. (So the between-ock variability is described by the standard deviation (1.52) of the means of the within-ock variability distribution.) In the Monte Carlo simulation the upper tail of the distribution is truncated at 10 log cfu/g to prevent unrealistically high values that may have large impact on the nal risk estimate. 2.2.1.3. Contamination of the carcass exteriors. For the CARMA project three ocks were sampled at the entrance of a Dutch processing plant, which previously tested positive at the farm.(19) For each ock 20 carcasses were taken from the processing line, just after bleeding and just before entering the scalding

Output Unit
Broiler ock Broiler ock carcass Broiler ock llet Fillet Salad Probability of illness

Note: Broiler ock (with quotation marks) refers to groups of products derived from one broiler ock.

relative effects of interventions on the public health risk(2,3) (see also Section 3.3).

2.2. Model Description The food chain is split up in ve major consecutive stages: (1) processing (slaughter and industrial processing of the carcass); (2) cutting (of the carcass to a skinned chicken breast llet); (3) storage (before and at retail, and by the consumer); (4) consumer handling (of the llet, and the salad prepared along with it); and (5) ingestion (potentially of a dose of Campylobacter). A dose-response relation is used to predict the consequential probability of illness. Table I summarizes how the units modeled change over the food chain. Parameter values are given in Table III. In the selection of data used in the risk model, we gave priority to data published in the scientic peer-reviewed international literature, describing the year-round situation in the Netherlands in the base year 2000. As these data were hardly ever available, we compromised on this by either taking data from other (recent) years, taking data from other (European, Western, industrialized) countries, or taking unpublished Dutch data. In case of absence of applicable research data, we based our parameter estimates on expert opinion. For industrial processing a formal expert judgment study was performed;(16) for other parts, parameter estimates were based on less formal expert estimates of parameter values. Note that the choice for using expert judgment as a data source was not only governed by availability of poultry processing data, but also by the suitability of these data to be used in the mechanistic nonlinear model applied here. In our modeling approach, applying the proper mechanistics in the risk model is consid-

Risk Assessment Model for Campylobacter tank. Campylobacter numbers on the exteriors were determined in the rinse water after rinsing the carcasses with 100 mL of peptone saline for 30 seconds. Normal distributions were tted through the logs of the counts per carcass, using maximum likelihood estimation methods. These distributions thus represent the variability within the ocks. The overall mean is 7.27 log cfu/carcass, close to other reported data.(20,21) The mean standard deviation within a ock is 0.56, the standard deviation of the means between ocks is 0.83. In the risk model these values are applied in normal distributions describing the log of the numbers of cfu per carcass. Samples are truncated at 4.5 at the lower tail and 10 at the upper tail, as the standard deviation between ocks is based on a sample of three ocks only, and to prevent too extreme values. We assume that all birds from an infected ock are contaminated at their exterior when they enter the processing plant due to cross-contamination from infected birds during rearing and transport, irrespective of intestinal colonization. Samples of the distribution of N ext are adjusted for animal prevalence <100%, as smaller amounts of Campylobacter from chicken feces will be spread through the ock if animal prevalence is lower. The (within-ock) mean of the normal distribution for log N ext,input , sampled from the between-ock normal distribution (7.27, 0.83), is lowered by the log of the animal prevalence, log (Panim ), to obtain the mean applied as within-ock mean. For simplicity it is assumed that values of N ext,input and Cfec are not correlated. This may not seem a realistic assumption, but we have shown that the effect of this simplifying assumption on the model results is negligible.(3) With this approach it is also assumed that birds from uninfected ocks are not contaminated on their exteriors. This may not be the case, as crosscontamination may occur during catching and transport. However, such ocks are not incorporated in the analysis because quantitative data on this phenomenon are lacking, and because it is expected that the numbers of Campylobacter on these ocks will in general be orders of magnitude lower than in infected ocks. As discussed elsewhere,(2,11) lower quantities of Campylobacter per ock are associated with much lower risks and, therefore, cross-contaminated negative ocks may be neglected. 2.2.2. Processing In fully automated poultry processing crosscontamination and removal may occur at all stages

849 of the process.(22) Scalding, defeathering, and evisceration, in which feces contaminated with Campylobacter may leak from the carcasses, are supposed to be the most critical points for cross-contamination during processing.(10,22,23) The nal number of organisms that are on any carcass is a result of the initial contamination and the effect of all stages of processing.(10) In our processing model, which is extensively described elsewhere,(11) the consecutive processing stages of scalding, defeathering, evisceration, washing, and chilling are each described by applying the same basic processing model. In the Netherlands, low scalding (with water temperatures 5052 C) and air chilling are applied for the production of fresh meat like chicken breast llets. The processing model(11) requires parameter estimates that could not be based on published data. Although some data on changes in levels of Campylobacter in chicken processing lines are available,(2426) these data could not be transformed in a format applicable to this study. Therefore, a formal expert judgment study was performed(16,27) in which 12 Dutch and one British expert(s) were consulted and weighted on the basis of their scores on a set of query variables. The study resulted in a set of interdependent uncertainty distributions of the model parameters for each stage of processing.(3) In the baseline risk model we apply the median values of these distributions as default estimates of all model parameters (see Table III).

2.2.3. Cutting At the cutting stage, chilled carcasses are deboned and portioned. First, breast caps are cut from the carcass, then (double) llets are cut and skins are removed. Most of this is done by automated equipment, but peoples handwork may be involved as well. Here, we dene a chicken breast llet as a single llet, i.e., half the llet originating from one carcass. Count data on the effects of cutting on the numbers of Campylobacter on chicken products are scarce(28,29) and could not be applied to our model. Being outside the scope of the formal expert study,(16) only one expert3 was interviewed to obtain estimates for this stage of the food chain.

J. Obdam, Head of Research and Development, Plukon Poultry BV, Wezep, The Netherlands. Estimates are to be considered as indicative, kindly provided at our request. The authors of this article are responsible for applying these estimates here.

850 As the available information on the dynamics of transmission of Campylobacter during this process is limited, a simple model is applied, based on the insights gained from modeling the previous stages. The relevant basic MPRM processes are partitioning (a breast cap is cut from the carcass) and crosscontamination/removal (from breast cap skin to llet, from equipment to llet, etc.). For these, we apply two separate models. 2.2.3.1. Partitioning. The number of Campylobacters from breast cap i, N ext,bc (i), will depend on the number on the carcass after chilling N ext,chill (i) and the probability for each Campylobacter on the carcass to reside on the breast cap, pcut . If this (variable) probability is Beta( 1 , 2 ) distributed, Next,bc (i) BetaBin[Next,chill (i), 1 , 2 ] (1) (where indicates is a sample from). Based on expert opinion we apply a Beta(1, 3.15) distribution for pcut , so the mean value is 0.24. 2.2.3.2 Cross-contamination/removal. The llet is cut from the breast cap and the skin is removed. As Campylobacters are supposed to be absent or rare on the skinless meat prior to slaughter, they are transferred to the meat by cross-contamination. This can either be by passing through tears in the skin, along with water leaking under the skin during processing, or during the cutting process. We assume the latter to be the dominant route. The number on the double breast llet cut from carcass i, N ext,df (i), is a function of the number on the breast cap and the expected number in the environment (equipment, hands, etc., as in the processing model). The latter can be estimated, and is linearly related to the expected number on the breast cap E(N ext,bc ).(3,11) This results in: Next,df (i) = a Next,bc (i) + E(Next,bc ). (2)

Nauta, Jacobs-Reitsma, and Havelaar At this step stochasticity is omitted because its effect is negligible.

2.2.4. Storage After cutting, the llets are stored and transported. In modeling terms that means that for some time they experience a variable temperature prole. Typically, the potential inactivation of Campylobacter during this process should be modeled by a secondary survival model. However, information and data are lacking to apply such a model here. Therefore, a simple log-linear reduction is assumed: log(Next,fs (i)) = log(Next,f (i)) rstorage . (4)

The decrease in survival after storage, rstorage , has to be estimated. For this, CARMA project count data(19) are applied. Fillets from six different ocks processed in the summer of 2004 in the Netherlands were sampled and Campylobacters were counted directly after cutting. Fillets of the same ock were individually stored in plastic bags for a week at 4 C, and then counted. A variable reduction is derived from these data setting rstorage BetaPert(0.1, 0.9, 2.1).

2.2.5. Consumer Preparation For consumer preparation we apply a model for cross-contamination during food preparation.(30) It describes cross-contamination from raw broiler meat to hands, cutting board, tap water, and salad, based on several studies on cross-contamination of bacteria in a kitchen environment. It is assumed that the broiler meat is well cooked, inactivating all Campylobacters on the meat. Cross-contamination to the salad is considered to be the only relevant route. The model holds transfer coefcients tX,Y , expressing the probability for each cfu Campylobacter to be transferred from X to Y during the handling X and Y. Transfer from and to chicken meat (C), hands (H), cutting board (B), and salad (S) are considered, the tap water is ignored because its effect is negligible.(30) Transfer coefcients tH,H , tB,B , and tS,S express the probability of survival per cfu Campylobacter during hand washing, board washing, and salad washing, respectively. The transfer coefcients are implemented as variability distributions derived from the literature, as given in Table II. The number of Campylobacters ending up in the salad for meal i, N salad (i), is:

Here, a is the contribution of the breast cap i that the double llet is cut from, and is the contribution of the environment in the total number of Campylobacters on the llet after cutting. The expert point estimate for the expected fraction transmitted from breast cap to double llet, a + , is 101.8 = 0.0158. Also, the expert estimates that about 17% of the contamination originated from the carcass itself, so a/(a + ) = 0.17. Therefore, estimated parameter values are a = 0.0027 and = 0.0132. Finally, the number on the single llet is calculated as: Next,f (i) = 0.5Next,df (i). (3)

Risk Assessment Model for Campylobacter

851

Transfer/Wash-Off Chicken to hand tC,H Chicken to cutting board tC,B Hand to salad tH,S Cutting board to salad tB,S Hand washing tH,H Cutting board washing tB,B Salad washing tS,S

Mean Value 0.0415 0.0125 0.207 0.343 0.0347 0.0000464 0.367

Variability Distribution Beta(1.78, 41.1) 10 Normal (0.098, 0.606) % Beta(0.6, 2.3) 10 Normal (1, 535, 0.32) % Beta(0.24, 6.67) BetaPert(1, 4.5, 7) log reduction BetaPert(0, 0.4, 1) log reduction

Reference 43 44 43 44 45 46 47 Table II. The Mean Transfer Probabilities per cfu Campylobacter and Wash-Off Effects Associated with Different Food Handling Activities (in the Risk Models, Variability of These Probabilities is Sampled from the Indicated Distributions).

N salad (i) = (tC,H tH,H tH,S + tC,B tB,B tB,S )tS,S Next,fs (i). (5)

with the (adjusted) number of table companions (see Reference 3 for details) 2.2.7. Dose-Response As a nal outcome, we aim to predict the probability of illness as a consequence of the consumption of a salad prepared together with a chicken breast llet. Therefore, the exposure assessment is linked to a dose-response relationship. The most frequently used dose-response relationship for Campylobacter is the beta Poisson model for probability of infection(33,9,34) based on the data of a volunteer study.(35) This model states that p1 , the probability of infection by one ingested Campylobacter, is variable between ingested doses, as described by a beta distribution: p1 Beta(, ). (7)

Next to transfer probabilities of Campylobacters, it is important to take frequencies of the type of handling into account. First, it is assumed that chicken breast llet is cut before the salad in 50% of the prepared meals. The effect of hand washing, board washing, or salad washing is only relevant if these are actually washed. If not, the transfer coefcient will have value 1. Also, chicken breast llets should be cut on the same (side of the) cutting board for cross-contamination via the cutting board to occur, else tB,B = 0. Data on the frequencies of these handlings are collected from the literature. Also, some of the data are derived from statements of the controls in the case control study on Campylobacter and Salmonella conducted in the Netherlands during 20022003.(31) 2.2.6. Exposure So far, the risk model assesses the number of Campylobacters in a salad prepared together with chicken breast llet from a chicken from a contaminated ock in a domestic kitchen. This salad may be eaten by several table companions joining the meal. Therefore, to estimate the distribution of exposures (i.e., the numbers of Campylobacters expressed as cfu ingested per person), this information is combined with Dutch consumption data.(32) This allows estimation of the number of person-meals with both chicken breast llets and salad consumed per year in the Netherlands at about 85 million in a population of about 16 million people. As explained elsewhere,(3) exposure per person can be assessed by accounting for the presumed number of table companions per chicken breast llet meal, the number of llets used per meal, and meal sizes.(32) The ingested dose per person is a sample from the number of Campylobacters in the salad: Ning Binomial(Nsalad , 1/), (6)

Here, we apply the model for discrete doses in a betabinomial model. The probability of infection due to the ingestion of n Campylobacters is then: Pinf (n) = 1 (1 p1 )n . (8)

Mean values of Pinf (n) can be used at a population level in the risk assessment model, that is,(36) Pinf (n) = 1 ( + ) ( + n) , () ( + + n) (9)

where () represents the gamma function. For Campylobacter(37) = 0.145 and = 7.59. Hence, the expected value of p1 is 0.019, implying a 2% probability of being infected after ingestion of a single living Campylobacter cell. Data(35) also include information on the probability of illness given infection. As a simple model for the probability of illness given infection, it is assumed that Pill|inf = 0.33, based on the fact that 29 people got sick out of 89 individuals that were infected.(9,35,38) The dose-response model for illness now becomes: Pill (n) = Pinf (n) Pill|inf . (10)

852 2.2.8. Overview of the Risk Model For an overview of the risk model, Table III shows the model parameters and their values applied in the baseline model, and Table IV summarizes the chain of models applied that link the number of Campylobacters on the exterior of the carcass and the meat product at each stage. 3. RESULTS 3.1. Baseline In the baseline risk model the model parameter values are the best estimates, as indicated in Section 2.2 (see Reference 3 for details). Variability between birds in a ock and variability between ocks is included; uncertainty in parameter values is not taken into account. The results of the baseline processing model until cutting the chicken breast llets are shown in Fig. 2. On average, the model predicts a steady decrease of the numbers of Campylobacter on the carcasses during processing. The mean value of log (N ext ) for contaminated ocks decreases from 6.71 (SD 0.56) per bird at the entrance of the processing plant to 4.36 (SD 0.39) per carcass after chilling in colonized ocks and 1.97 (SD 0.09) per llet after cutting. The variability between ocks and carcasses is large (see Fig. 2) and decreases within ocks due to the effects of crosscontamination. Occasionally, individual carcasses (in 19% of the model runs) and mean values per ock (17%) show an increase in the level during defeathering due to leaking feces with a relatively high concentration of Campylobacter. Moreover, the mean effect of processing (i.e., the log (N ext,chill ) log (N ext,input )) is positive in 3% of the iterations, indicating that N ext increases due to processing in 3% of the ocks. For contaminated ocks the prevalence of contaminated llets (N ext,f > 0) is 94%. After cutting and deboning, within-ock dynamics are no longer relevant and llets are randomly sampled from the simulated ocks. Hence, within-ock variability and between-ock variability can be combined to between-llet variability by joined sampling. As discussed in Section 2.2.4, storage affects survival, through which the distribution of N ext,f , the number of Campylobacters on the llets, shifts to lower values: after storage the prevalence of contaminated llets is 81%, with a mean value of log (N ext,fs ) of 1.4 (SD 0.08) for contaminated llets. Finally, as an effect of consumer food handling and potential cross-contamination to a salad, the distribu-

Nauta, Jacobs-Reitsma, and Havelaar tion of N consumption is once more shifted to lower numbers. The predicted frequency of exposure to Campylobacter is 0.80% of the individual meals with chicken breast llet and salad. Fig. 3 presents the exposure in dose classes of ingested numbers of Campylobacter along with the percentage of human cases attributable to exposure to those dose classes. It shows that the majority of exposures are to doses of 1 and 2 cfu. Less than 2% of the exposures are to doses > 100 cfu Campylobacter. Although 11.5% of the exposures are to doses >10 cfu, 45.4% of the human cases is attributable to these exposures. In the end, the baseline model predicts an incidence of 12,300 cases per year due to consumption of salad cross-contaminated by chicken breast llets. 3.2. The Impact of Input Variability The large variability between outputs of model runs is a reection of the variability in the input of the model, the animal prevalence (Panim ), the concentration in the leaking feces (Cfec ), and the numbers of Campylobacter on the exterior (N ext ). All vary per ock, and the last two also vary per bird within ocks. This variability is important in the risk assessment model, so its impact is explored. The model is run (1) with xed animal prevalence (Panim = 100%), (2) by setting the N ext,input = 0 or Cfec = 0 for all birds, and (3) by setting within- and between-ock standard deviation at zero, all other values remaining as in the baseline. The effects of (1) and (2) on the average processing dynamics are illustrated in Fig. 4. With Panim = 100% for all ocks, more birds are contaminated, more Campylobacters are introduced into the system, and higher levels are found at all stages of processing. With N ext,input = 0 there is no initial contamination on the birds and all Campylobacters found on the carcass exteriors result from fecal leakage. (After chilling the mean value for log (N ext,chill ) is 3.4.) With Cfec = 0 there is no carcass contamination from the feces. As this is mainly effective during defeathering, canceling this effect leads to a stronger decrease in the mean level after defeathering. (After chilling, the mean value for log (N ext,chill ) in 3.8.) These mean results suggest that human exposure and illness are predominantly the consequence of Campylobacters present on the exterior at the entrance of the processing plant, not of Campylobacters in the chicken feces. However, this is not the case: 78% of the total human incidence is due to Campylobacters

Risk Assessment Model for Campylobacter


Table III. Parameters and Default Values Applied in the Baseline Model Stage Entrance processing plant Parameter Pock Panim Cfec Cfec N ext,input N ext input (Low) scalding aext 1-afec benv cext cenv pfec ms ms aext 1-afec benv cext cenv pfec ms ms aext 1-afec benv cext cenv pfec ms ms aext benv cext cenv aext benv cext cenv 1 2 a/(a + ) log (a + ) rstorage f hw f bw f ob f sw Exposure Dose-response Pill|inf Meals Description Flock prevalence Animal prevalences Concentration in feces Standard deviation (SD) between birds within a ock SD of between ocks Number campy. on exterior SD between birds within a ock SD of between ocks Transmission of exterior to environment Transmission of feces to exterior Transmission of environment to exterior Removal from exterior Removal from environment Probability of leaking feces Mean mass of leaking feces SD of ms (log-normal) Default Value 44.4% Distribution(17) 6 0.73 1.52 7.27 0.56 0.83 0.86 1.40106 6.80106 0.71 0.036 0.48 1.84 0.001 0.91 1.03105 9.85104 0.049 0.11 0.70 1.72 0.001 0.46 2.75103 1.07105 0.042 0.085 0.68 1.94 0.001 0.31 2.75103 0.245 0.062 0.061 3.75103 0.035 0.015 1 3.15 0.17 1.8 BetaPert(0.1, 0.9, 2.1) 80% 62% 33% 60% See Ref. 3 0.145 7.59 33% 81,000,000

853

Reference/Source 17 17 18 18 18 19 19 19 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 See Section 2.2.3 See Section 2.2.3 See Section 2.2.3 See Section 2.2.3 19 (See Section 2.2.4) See Section 2.2.5 31 31 31 48 32 37 37 38 3

Defeathering

Evisceration

Washing

(Air) chilling

Cutting

(Cool) storage Consumer preparation

Parameter partitioning Parameter partitioning Relative contribution of carcass i to llet i Log-fraction of carcass contamination on llet Reduction by storage Transfer rates: see Table 2 Hand washing frequency Cutting board washing frequency Frequency of using other cutting board Frequency of salad washing Various consumption data Dose-infection parameter Dose-infection parameter Probability of illness given infection Number of chicken llet with salad meals

Note: Parameters for industrial processing stages (scaldingchilling) are described elsewhere;(3,11) values are medians of an expert judgment study.(16) Other stages are discussed in the main text.

854

Nauta, Jacobs-Reitsma, and Havelaar

Stage Scalding Defeathering Evisceration Washing Chilling Cutting (1) Cutting (2) Storage Consumer prep Exposure Response

Output N ext,scald N ext,def N ext,evis N ext,wash N ext,chill N ext,bc N ext,f N ext,fs N salad N ing Pill

Equation

Unit

Table IV. Overview of the Models Applied for Each of the Stages of the Risk Model

See Ref. 11 Carcass See Ref. 11 Carcass See Ref. 11 Carcass See Ref. 11 Carcass See Ref. 11 Carcass BetaBin(N ext,chill (i), 1 , 2 ) Breast cap = 0.5(a N ext,bc (i) + e(next,bc )) (Single) llet =10r storage N ext,f (i). (Single) llet = (tC,H tH,H tH,S + tC,B tB,B Salad (table companions) tB,S ) tS,S N ext,fs (i) Binomial(N salad , 1/ ) Meal (p.p.) (1
( + ) ( + Ning ) )Pill| inf () ( + + Ning )

Meal (p.p.)

Note: For explanations, see the main text.

9 8 7 6 5 4 3 2 1 0
ng ng t n pu rin tio ng illi di in he al ra hi ce as ch cu tti ng g

60.0% 50.0%

exposures cases

log cfu / unit

frequency

40.0% 30.0% 20.0% 10.0% 0.0% 1 2 3-10 11-100 >100 cfu ingested

sc

at

is

Fig. 2. Numbers of Campylobacter per carcass [log (N ext )] are given for the consecutive stages of processing. Straight lines represent mean 5, 50, and 95 percentiles of the within-ock variability of 5,000 ocks, dashed lines show the 5, 50, and 95 percentiles of the means of 5,000 ocks, representing the between-ock variability. On average there is a decrease in the level of contamination during processing. Between-ock variability is much larger than within-ock variability.

Fig. 3. The relative frequencies of dose classes and the percentages of human cases of campylobacteriosis attributable to those classes. Higher doses have more impact than the distribution of exposures suggests.

fe

de

ev

originating from the leaking feces (see Section 3.3). This is a consequence of the large variability in mean values of Cfec between ocks, as illustrated in Fig. 5. It appears that, occasionally, very high numbers of fecal bacteria are transmitted to the carcasses during processing and that these peak numbers drive consumer risk to a large extent. Hence, the model predicts that fecal leakage as a source of Campylobacter is more important than the contamination of the birds exterior at the entrance of the plant. In (3) the standard deviations of log (N ext,input ) and log (Cfec ) are set to zero, either within or between ocks, achieving similar results: there is little

effect on the mean values found during processing (slightly lower than the baseline in Fig. 4, not shown), but the variability decreases, so heavily contaminated carcasses are very rare. The total number of Campylobacters in the system (and thus the arithmetic means of N ext,input and/or Cfec ) and the predicted incidence are much lower if the variability is neglected. If both within- and between-ock variability is neglected, the predicted incidence is only 3% of that in the baseline. 3.3. Sensitivity Analysis The nal output of the model is the predicted human incidence of campylobacteriosis as indicated above. However, the main objective of the model is not to precisely predict this incidence, but to evaluate the effects of potential interventions in the poultry meat chain.(2,3) To study the effects of these interventions and other alternative scenarios with different

Risk Assessment Model for Campylobacter

855 rameter values, is analyzed to explore the sensitivity and uncertainty of the risk model. This small set is a selection of relevant scenarios from a larger set.(3) Results are shown in Fig. 6. The effects of potential interventions to control Campylobacter are described and discussed elsewhere.(2) 1. Input: The effect of modied input is explained in Section 3.2. As discussed there, Fig. 6 shows that the relative risk with Cfec = 0 is about 22%, and with N ext,input = 0 it is about 78%, indicating that fecal leakage is the dominant source of campylobacteriosis. In particular, the effect of the exclusion of between-ock variability in the input on the relative risk is very large. 2. Expert uncertainty processing model: The expert study for industrial processing(16) quantied the uncertainty in the expert estimates. Analysis of baseline model runs where the quantied expert uncertainty is included and between-ock variability is neglected show that the 5% and 95% percentiles of the uncertainty distribution show a decrease of about 1 log-unit and an increase of about 1.5 logunit in the mean value of N ext,f . With these effects added as changes in N ext,f in the baseline model, risks are 19% and 740% of the most likely estimate, respectively, illustrating that the uncertainty in the risk estimate due to expert uncertainty is large. Next, in the baseline model analysis we use the median values of the uncertainty distributions of the processing

baseline 8 7 6 log cfu / unit 5 4 3 2 1 0 p_anim=1 N_ext=0 C_fec=0

ra tio n

in pu t

ch illi ng

w as hi ng

ng

rin

sc al di

he

at

fe

Fig. 4. Dynamics of the mean values of log (N ext ) during processing for the baseline model and some alternatives with modied input. If the animal prevalence is set at 100% for all ocks, all mean levels increase. If N ext,input is set at zero for all birds, contamination of the exterior is introduced by fecal leakage, mainly during scalding and defeathering. Without a cross-contamination effect of fecal leakage (all Cfec = 0) levels on the carcasses after defeathering and evisceration decrease.

input and/or model parameter values than the baseline model, changes in the assessed incidence are compared, expressed as relative risks. Here, the relative risk of an alternative scenario is dened as the assessed number of cases in this scenario divided by the assessed number in the baseline scenario. A small set of alternative scenarios, in which the model is run with one or a few different model pa-

de

ev

is

ce

between flock variability after chilling


7 6 4
log cfu per carcass log cfu per fillet

cu

tti

ng

between flock variability after cutting


5

5 4 3 2 1 0 baseline C_fec=0 scenario N_ext=0

3 2 1 0 baseline C_fec=0 scenario N_ext=0

Fig. 5. Comparison of the birds exterior and their intestines as source of Campylobacter on the meat. The variability in mean results per ock are compared for the baseline and the alternatives N ext,input = 0 and Cfec = 0 for all birds in the ock. The boxes indicate the mean and the median, bars give the range between 5 and 95 percentiles. If the mean is smaller than the median, the box is white, else it is black. The left graph shows the result for log (N ext,chill ), right for log N ext,f . Central values for the Cfec = 0 scenario (only exterior contamination at the start of processing) are larger than for the N ext,input = 0 scenario (only contamination by leaking feces), but the 95% is lower. The latter has most impact on human exposure and the associated health risk, so the leaking feces is the dominant source.

856
N_ext = 0 C_fec =0 P_anim =100% no flock var no bird var no var at all 95% level 5% level mean unc vals experts unweighted cutting 2.3 cutting 1.3 storage = 0 f_hw = 50% f_hw = 90% f_ob = 85% f_ob = 100% Pill|inf = 25% Pill|inf = 40% other DR model 1% 10%

Nauta, Jacobs-Reitsma, and Havelaar

Fig. 6. Sensitivity analysis results. The predicted incidences of some scenarios relative to that of the baseline are shown by the bars, and are discussed in the main text. Top down scenarios relate to (1) model input, (2) uncertainty characterized in the expert judgment study,(16) (3) cutting, (4) llet storage, (5) consumer food handling, (6) dose response.

100%

1000%

10000%

Relative risk

model parameters because these distributions may be quite skewed.(3) If the baseline model is run with mean values instead, this yields a risk of 400% of the most likely estimate. The effect of weighing the experts(16) is limited. The risk estimate is slightly smaller if expert opinions are incorporated unweighed. 3. Cutting: The effect of cutting is based on some rough estimates. The baseline value for log (a + ) = 1.8 (with a + the expected fraction of Campylobacters transmitted from breast cap to double llet) is modied to 1.3 and 2.3. This approximately halves and doubles the estimated risk. 4. Storage: Neglecting an effect of storage between llet production and consumption (rstorage = 0) has a large effect on the risk estimate. As the baseline estimate for rstorage is derived from a small data set, better insight into the effect of storage (under different conditions) seems important. 5. Consumer preparation: The potentials of improved hygienic behavior by consumers is explored by some scenarios affecting the frequency of hand washing and the use of separate cutting boards. Results indicate that con-

sumer behavior has relatively little impact on the risk estimate. 6. Dose response: The effect of the doseresponse model is explored by applying modied probability of illness given infection (which is based on a rather rough estimate of 33% in the baseline), and by applying an updated dose-response model for Campylobacter.(3,39) Apparently the choice of the dose-response model has large impact on the assessed incidence.

4. DISCUSSION 4.1. Baseline The baseline risk model is a representation of the most likely situation for broiler chicken meat produced in the Netherlands in 2000. It uses most likely values for the model parameters, including variability in levels of Campylobacter between ocks and between birds (and carcasses) within a ock, stochasticity in fecal leakage, variability in effects of storage between industrial processing and consumer food preparation, and variability in food preparation and consumption.

Risk Assessment Model for Campylobacter The results show that, on average, there is a continuous decrease in the level of Campylobacter on the carcasses and the meat, N ext (see Fig. 2). This is the consequence of the absence of growth and the repeated inactivation and removal of Campylobacter during processing, as a consequence of submersion, washing, defeathering, skinning, storage, etc. Contamination of the carcass exterior by feces leaking from the carcass (predominantly during defeathering) may give a temporal increase in the level of Campylobacter when the concentration in the feces, Cfec , is relatively high and N ext is relatively low. On average this effect is dominated by a continuous decrease because (1) animal prevalence smaller than 100% implies that for many birds Cfec = 0, whereas it is assumed that N ext,input is always >0, and (2) values of N ext are relatively large compared to values of Cfec . The reason for both (1) and (2) is that apparently the birds exterior is already heavily contaminated with feces at the entrance of the processing plant. Furthermore, results show that cross-contamination during processing and food handling leads to a dispersal of Campylobacters over the units (birds, carcasses, llets) involved, and as a consequence the prevalence of contaminated units may increase, whereas the variability between units decreases.(11) The ingested doses at exposure are generally very low, the majority only 1 or 2 cfu (see Fig. 3). This is due to the inactivation, removal, and cross-contamination to salad during cutting, storage, and food handling. Due to these processes the distribution describing the variability of numbers of pathogens on food units is shifted to lower numbers. When the distribution shifts to the left, the major part of it shifts to zero cells per unit, and only a small part of the right-hand tail remains. This result complicates risk estimation because dose-response data usually concern much higher doses,(35) and the effects of low doses are to be inferred from these high dose data. In the model applied here, this inference is based on the single-hit model assumptions.(37) For the most frequent dose of 1 cfu the single-hit dose-response model applied here predicts 1.9% probability of infection and 0.6% probability of illness when 1 cfu Campylobacter is ingested. The baseline estimate of the incidence is 12,300 cases of campylobacteriosis per year as a consequence of the consumption of salad cross-contaminated after handling chicken breast llet. Epidemiologic studies in the Netherlands lead to a mean estimate of about 80,000 human cases (90% condence interval between 20,000 and 160,000 cases) of campylobacterio-

857 sis per year.(1,2) As the attributive fraction of broiler meat lies between 20% and 40%,(3) between 16,000 32,000 cases are estimated to be caused by the consumption of broiler meat. A fraction of these will be the consequence of the consumption of salad crosscontaminated from chicken breast llet. Other routes, like cross-contamination by other chicken meat products and insufcient cooking of the meat product and eating in restaurants or at barbecues, will also contribute to the total human incidence due to broiler meat consumption. The model estimate of more than 12,000 cases therefore seems rather high. However, assumptions on consumer behavior like the percentage of salads cut after handling the chicken breast llets (50%) and the percentage of consumers that put the fresh meat in the freezer (0%) proportionally affect this estimate. Given the overall uncertainty of the assessed exposure distribution, and the uncertainty of the dose-response relationship in addition, an imprecise estimate is expected. The predicted incidence seems to t very well into the range of acceptable predictions.

4.2. Validation Although the variability between runs of the model indicates that the possibilities for model validation based on small data sets is limited, some data are available that can be used to check the model results.(3) A set of ocks was recently sampled at different stages of processing.(19) Comparison of these data with the risk model results shows that the high variability between ocks as predicted by the processing model is conrmed.(3) International processing data(2426) show a frequent increase in the numbers of Campylobacters on the carcasses with defeathering and evisceration and a relatively strong decrease during washing, which does not t with the general results of the risk model. As other risk assessments are primarily based on these data, processing models developed by others reect this dissimilarity in results.(69) It may be due to different processing practices in other countries, higher levels of fecal contamination (Cfec ) than those assumed in our model, and difculties in tting sample data with modeling results.(11) Next, monitoring data of broiler meat at retail(40) and Campylobacter count data on chicken breast llets are available.(3,19) Taking into account a realistic detection limit, both studies indicate that the model may underestimate the numbers of Campylobacter on chicken breast llets.

858 In general, there is a striking similarity in the results of this risk assessment and a previous one on Shigatoxin producing E. coli O157 in steak tartare.(41) In both studies the majority of ingested doses is very low (110 cfu) and the estimated incidence on the basis of the available dose-response relation is too high compared to epidemiological data. Also, compared with (monitoring) data at retail, both risk models seem to underestimate the prevalence and concentration of pathogens in the food products. This suggests that the dose-response relationships used overestimate the probability of illness in the general population.

Nauta, Jacobs-Reitsma, and Havelaar 4.4. Modeling Method The risk model is constructed applying MPRM (modular process risk model) methodology.(12) The industrial processing model integrates crosscontamination and removal, incorporates some simple mechanistics, and is nonlinear.(11) A simplied version is applied for the cutting stage. Along with the inclusion of both Campylobacters on the exterior and in the feces as distinct sources of contamination, the structural application of this nonlinear model is a novelty in process risk modeling. It yields results that cannot be obtained by models based on log linear input-output relationships only. The latter has been the approach of previous risk assessments of Campylobacter, which are mainly based on some count data of chicken processing and linear relationships between the means of log concentrations of Campylobacter over the processing stages.(610) The advantage of such an approach is that it is based on actual microbiological data, and may thus seem to be a more realistic representation of the actual (present) situation. However, although in general microbiological data are to be preferred above expert opinions, the correct use of these data for risk assessment models is not without complications.(3,11) The disadvantage of data-based linear models is that they ignore nonlinear effects of cross-contamination. This may have a serious impact on the assessment of the effects of interventions where other levels of contamination occur than those in the situation where the data were collected, and therefore these models may be less suitable for risk management.(11)

4.3. The Impact of Input Variability As a novelty in food chain risk assessment, the risk model explicitly includes variability in prevalence and numbers of Campylobacter within and between ocks. The analysis described in Section 3.2 shows that this has a large impact on the model results. A crucial factor here is the assumption that the distributions of N ext,input and Cfec are log-normal. As microbial counts, like those referred to in this article, are usually expressed in logs, this is an obvious assumption. With a log-normal distribution, the arithmetic mean is larger than the geometric mean, which is based on the mean of logs. The reason for this is that the arithmetic mean also depends on the standard deviation of the logs. The total load of Campylobacters introduced in a processing plant is the product of the number of birds in a ock and its arithmetic mean, and therefore depends on both the mean and the standard deviation of the log counts. The separation of within-ock and between-ock variability is important because the within-ock variability affects the processing dynamics and decreases due to cross-contamination between carcasses from a ock.(11) Between-ock variability has the largest impact on the nal risk estimate, the assessed human incidence, and increases due to the variable effects of processing. These results indicate that data on ock prevalence and animal prevalence and even data on means of log counts alone are not sufcient as input for a broiler chicken risk assessment model. A good characterization of the variability in N ext,input and Cfec within and between ocks is crucial for risk assessment because this variability has a large impact on the processing dynamics and the estimated risks.

4.5. Uncertainty The effect of the uncertainty in parameter estimates as quantied in the expert study(16) is large (see Section 3.3), but a full analysis of this uncertainty falls outside the scope of this article. The total uncertainty about the baseline risk estimate is even larger than that of the processing model alone, and could only be analyzed quantitatively to a limited extent. Reasons for this are that at some stages it is very hard to quantify uncertainty, that it was difcult to implement the uncertainty in the computer program developed for the risk model, and that the nal interpretation of this uncertainty is complex.(3) Simplifying assumptions, like the suitability of the chain model for Dutch broiler meat production, the appropriateness of data, and the adequacy of the knowledge of the

Risk Assessment Model for Campylobacter characteristics of Campylobacter, have to be made in any model. They may affect the accuracy of the model output, especially if this is a risk estimate that may be sensitive to rare events not incorporated in the model. In the application of the risk model as a tool to evaluate the effect of interventions, the uncertainty in the estimated relative risks due to modied exposure is more important than the uncertainty of the human incidence as risk estimate of the baseline model. The same measure, the ratio of the predicted incidence of the model with modied parameter values and the baseline prediction, has been calculated in Section 3.3 for some alternative scenarios to analyze the sensitivity of some model assumptions. Apart from the impact of variability as discussed in Section 4.3, this analysis stresses the need for improved knowledge and more data on parts of the chain that usually get little attention in food safety management of the processing chain, like the cutting and storage stages. The uncertainty of the relative risk is expected to be substantially smaller than the risk estimate itself because the uncertainty in the numerator and the denominator may cancel out. This is particularly true for the uncertainty in parameters that are linearly related to the nal risk estimate, like the percentage of salads cut after handling the chicken breast llets and the percentage of consumers that put the fresh meat in the freezer. For model parameters higher up in the food chain uncertainties do not cancel out completely, but the decrease in uncertainty is still substantially large.(3) 4.6. Data Gaps To lower the uncertainty in the model results, future data collection should be focused more on quantitative data. Currently, much effort is put in collecting data on ock prevalences at farms, but these data could only be applied to a very limited extent in the farm phase of the risk model.(4) In comparison, highly relevant quantitative data on animal prevalences and levels of Campylobacter in chicken feces and on birds entering the processing plant (let alone the variability therein) are very scarce. In the production chain more quantitative data on the effect of processing steps could be helpful to obtain more precise estimates of the model parameter values now estimated by expert judgment. Here, better knowledge about particularly, the stages of defeathering, cutting, and storage at refrigerator temperatures is important. For the consumer phase of the model, data on actual daily life practices of consumers are preferred

859 above data on statements about behavior, but such data are scarce. At least the relationship between what people say and what people do should be made more transparent. In combination with the need to more precisely assess the potentials of risk intervention at the consumer phase, transdisciplinary research in this area involving social scientists, microbiologists, and risk modelers is required.(42) ACKNOWLEDGMENTS The study described in this report would not have been realized without the help of many. The following people and teams are gratefully acknowledged for their contributions by discussions, providing data, or otherwise: Hanne Rosenquist (DFVF, Denmark), Janet Corry (University of Bristol, UK), FAO/WHO drafting group Risk assessment of Campylobacter in broiler chickens,(10) Decisionalysis Risk Consultants Inc. (Ottawa, Canada), Peter Enthoven (CCL-Research), Jaap Obdam (Plukon Poultry, Wezep), the expert panel consulted for the expert judgement processing model,(16) Roger Cooke, Louis Goossens (Technical University Delft), Ine van der Fels-Klerx (RIVM, now at RIKILT), Eric Evers, Sido Mylius, Peter Teunis, Yvonne Doorduyn, Gert Steentjes, Martine Bakker, Annemarie Pielaat, Anne Mensink (RIVM), CARMA industry forum, CARMA project group, and CARMA steering committee. REFERENCES
1. Mangen, M.-J., Havelaar, A. H., Bernsen, R. A. J. A. M., Van Koningsveld, R., & De Wit, G. A. (2005). The costs of human Campylobacter infections and sequelae in the Netherlands: A DALY and cost-of-illness approach. Food Economics, 2, 3551. 2. Havelaar, A. H., Mangen, M. J. J., Nauta, M. J., De Koeijer, A. A., Bogaardt, M.-J., Evers, E. G., Jacobs-Reitsma, W. F., Van Pelt, W., Wagenaar, J. A., De Wit, G. A., & Van der Zee, H. (2006). Effectiveness and efciency of controlling Campylobacter on broiler chicken meat. Risk Analysis, 26(6). 3. Nauta, M. J., Jacobs-Reitsma, W., Evers, E. G., Van Pelt, W., & Havelaar, A. H. (2005). Risk Assessment of Campylobacter in the Netherlands via Broiler Meat and Other Routes. RIVM Report 250911 006, Bilthoven, The Netherlands. 4. Katsma, E., De Koeijer, A., Mangen, M.-J., Jacobs-Reitsma, W., & Wagenaar, J. (2006). Assessing interventions to reduce Campylobacter prevalence in broilers using transmission models. Risk Analysis, 26(6). 5. Mangen, M.-J. J., Havelaar, A. H., Poppe, K. P., & De Wit, G. A. (2006). Data gaps and their impact on costs and efciency when controlling Campylobacter on Dutch chicken meat. Risk Analysis, 26 (6). 6. Fazil, A. M., Lowman, R., Stern, N., & Lammerding, A. (1999). Quantitative risk assessment model for Campylobacter jejuni in chicken. Abstract CF10, p 65.

860
7. Hartnett, E., Kelly, L. A., Gettinby, G., & Wooldridge, M. (2002). A quantitative risk assessment for Campylobacter in broilers: Work in progress. International Biodeterioration & Biodegradation, 50, 161165. 8. Rosenquist, H., Nielsen, N. L., Sommer, H. M., Norrung, B., & Christensen, B. B. (2003). Quantitative risk assessment of human campylobacteriosis associated with thermophilic Campylobacter species in chickens. International Journal of Food Microbiology, 83(1), 87103. 9. Anonymous. (2001). Hazard Identication, Hazard Characterization and Exposure Assessment of Campylobacter spp. in Broilers. Report No. MRA 01/05. Rome, Geneva: FAO / WHO. 10. Anderson, S., Christensen, B. B., Fazil, A., Hartnett, E., Lammerding, A., Nauta, M., Paoli, G., & Rosenquist, H. (2003). A draft risk assessment of Campylobacter ssp. in broiler chickens. In Joint FAO/WHO Activities on Risk Assessment of Microbiological Hazards in FoodsInterpretative Summary (pp. 124) WHO/FAO. 11. Nauta, M., Van der Fels-Klerx, I., & Havelaar, A. (2005). A poultry-processing model for quantitative microbiological risk assessment. Risk Analysis, 25, 8598. 12. Nauta, M. J. (2001). A Modular Process Risk Model Structure for Quantitative Microbiological Risk Assessment and its Application in an Exposure Assessment of Bacillus Cereus in a REPFED Report No. 149106 007. Bilthoven: RIVM. 13. Nauta, M. J. (2002). Modelling bacterial growth in quantitative microbiological risk assessment: Is it possible? International Journal of Food Microbiology, 73(23), 297304. 14. Lindqvist, R., Sylven, S., & Vagsholm, I. (2002). Quantitative microbial risk assessment exemplied by Staphylococcus aureus in unripened cheese made from raw milk. International Journal of Food Microbiology, 78(12), 155170. 15. Nauta, M. J. (2005). Microbiological risk assessment models for partitioning and mixing during food handling. International Journal of Food Microbiology, 100, 311322. 16. Van der Fels-Klerx, H. J., Cooke, R. M., Nauta, M. N., Goossens, L. H., & Havelaar, A. H. (2005). A structured expert judgment study for a model of Campylobacter transmission during broiler-chicken processing. Risk Analysis, 25(1), 109124. 17. Katsma, W. E. A., de Koeijer, A. A., Fischer, E. A. J., Jacobs Reitsma, W. F., & Wagenaar, J. A. (2005). Campylobacter Prevalence in Broiler Flocks in the Netherlands. Modelling Transmission Within and Between Flocks and Efcacy of Interventions. Report ASG05/I00113, Animal Sciences Group (ASG) Wageningen-UR. 18. Enthoven, P. Unpublished data. CCL Research, The Netherlands. 19. Jacobs-Reitsma, W., Nauta, M., & Wagenoar, J. (2005). Counting Campylobacter along the broiler processing line. In Abstracts of the 13th International Workshop on Campylobacter, Helicobactor, and Related Organisms (p. 118). Gold Coast, Australia: CHRO. 20. Berrang, M. E., Buhr, R. J., & Cason, J. A. (2000). Campylobacter recovery from external and internal organs of commercial broiler carcass prior to scalding. Poultry Science, 79(2), 286290. 21. Stern, N. J., Clavero, M. R. S., Bailey, J. S., Cox, N. A., & Robach, M. C. (1995). Campylobacter spp. in broilers on the farm and after transport. Poultry Science, 74, 937941. 22. Hafez, H. M. (1999). Poultry meat and food safety: Preand post-harvest approaches to reduce foodborne pathogens. Worlds Poultry Science Journal, 55(3), 269280. 23. Kist, M. (2002). Food-borne Campylobacter infections. Bundesgesundheitsblatt Gesundheitsforschung Gesundheitsschutz, 45(6), 497506. 24. Oosterom, J., de Wilde, G. J. A., de Boer, E., de Blaauw, L. H., & Karman, H. (1983). Survival of Campylobacter jejuni during

Nauta, Jacobs-Reitsma, and Havelaar


poultry processing and pig slaughtering. Journal of Food Protection, 46, 702706. Izat, A. L., Gardner, F. A., Denton, J. H., & Golan, F. A. (1988). Incidence and level of Campylobacter jejuni in broiler processing. Poultry Science, 67, 15681572. Berrang, M. E., & Dickens, J. A. (2000). Presence and level of Campylobacter spp. on broiler carcasses throughout the processing plant. Journal of Applied Poultry Research, 9, 43 47. Cooke, R. M., Nauta, M. J., Havelaar, A. H., & Van der Fels, H. J. (2004). Probabilistic inversion for chicken processing lines. In Proceedings of the Conference on Sensitivity Analysis of Model Output. Los Almos, NM. Berrang, M. E., Ladely, S. R., & Buhr, R. J. (2001). Presence and level of Campylobacter, coliforms, Escherichia coli, and total aerobic bacteria recovered from broiler parts with and without skin. Journal of Food Protection, 64(2), 184188. Uyttendaele, M., De Troy, P., & Debevere, J. (1999). Incidence of Salmonella, Campylobacter jejuni, Campylobacter coli, and Listeria monocytogenes in poultry carcasses and different types of poultry products for sale on the Belgian retail market. Journal of Food Protection, 62(7), 735740. Mylius, S. D., Nauta, M. J., & Havelaar, A. H. (2006). Cross-contamination during food preparation: A mechanistic model applied to chicken-borne Campylobacter. Risk Analysis, 26(6). Doorduyn, Y., Brandhof van den, W. E., Duynhoven van, Y. T. H. P., Wagenaar, J. A., & Pelt van, W. (In prep.). Risk factors for endemic Campylobacter jejuni infections in the Netherlands: a case-control study. Kistemaker, C., Bouman, M., & Hulshof, K. (1998). De consumptie van afzonderlijke producten door Nederlandse bevolkingsgroepenVoedselconsumptiepeiling 1997 1998. TNO-Rapport V98.812, TNO Voeding, The Netherlands. Medema, G. J., Teunis, P. F., Havelaar, A. H., & Haas, C. N. (1996). Assessment of the dose-response relationship of Campylobacter jejuni. International Journal of Food Microbiology, 30(1-2), 101111. Teunis, P. F., Nagelkerke, N. J., & Haas, C. N. (1999). Dose response models for infectious gastroenteritis. Risk Analysis, 19(6), 12511260. Black, R., Levine, M. M., Clements, M. L., Hughes, T. P., & Blaser, M.J. (1988). Experimental Campylobacter jejuni infection in humans. . Journal of Infectious Diseases, 157, 472479. Haas, C. N. (2002). Conditional dose-response relationships for microorganisms: Development and application. Risk Analysis, 22(3), 455463. Teunis, P. F., & Havelaar, A. H. (2000). The beta Poisson doseresponse model is not a single-hit model. Risk Analysis, 20(4), 513520. Havelaar, A., de Wit, M., & van Koningsveld, R. (2000). Health Burden in the Netherlands (19901995) Due to Infections with Thermophilic Campylobacter Species. Rijksinstituut Voor Volksgezondheid En Milieu RIVM Rapport Nr. 284550004. Teunis, P., Van den Brandhof, W., Nauta, M., Wagenaar, J., Van Pelt, W., & den Van Kerkhof, H. (2005). A reconsideration of the Campylobacter dose response relation. Epidemiology and Infection, in press. Van der Zee, H., Wit, B., & De Boer, E. (2002). Monitoring pathogenen in kip en kipproducten. In Keuringsdienst Van Waren Oost, Afd. Signalering. Zutphen, The Netherlands. Nauta, M. J., Evers, E. G., Takumi, K., & Havelaar, A. H. (2001). Risk Assessment of Shiga Toxin Producing Escherichia Coli O157 in Steak Tartare in the Netherlands. Report No. 257851 003. Bilthoven: RIVM. Fischer, A. R. H., De Jong, A. E. I., De Jonge, R., Frewer, L. J., & Nauta, M. J. (2005). Improving food safety in the domestic

25. 26.

27.

28.

29.

30.

31.

32.

33.

34. 35. 36. 37. 38.

39.

40. 41.

42.

Risk Assessment Model for Campylobacter


environment: The need for a transdisciplinary approach. Risk Analysis, 25(3), 503518. 43. Montville, R., Chen, Y., & Schaffner, D. W. (2001). Glove barriers to bacterial cross-contamination between hands to food. Journal of Food Protection, 64(6), 845 849. 44. Kusumaningrum, H. D., van Asselt, E. D., Beumer, R. R., & Zwietering, M. H. (2004). A quantitative analysis of crosscontamination of Salmonella and Campylobacter spp. via domestic kitchen surfaces. Journal of Food Protection, 67(9), 18921903. 45. Chen, Y., Jackson, K. M., Chea, F. P., & Schaffner, D. W. (2001). Quantication and variability analysis of bacterial

861
cross-contamination rates in common food service tasks. Journal of Food Protection, 64(1), 7280. 46. Cogan, T. A., Slader, J., Bloomeld, S. F., & Humphrey, T. J. (2002). Achieving hygiene in the domestic kitchen: The effectiveness of commonly used cleaning procedures. Journal of Applied Microbiology, 92(5), 885892. 47. Smith, S., Dunbar, M., Tucker, D., & Schaffner, D. W. (2003). Efcacy of a commercial produce wash on bacterial contamination of lettuce in a food service setting. Journal of Food Protection, 66(12), 23592361. 48. Worsfold, D., & Grifth, C. J. (1997). Assessment of the standard of consumer food safety behavior. Journal of Food Protection, 60, 399406.

Você também pode gostar