Figure VI-1. Schematic diagram of the V. parahaemolyticus (Vp) risk assessment model showing integration of all the modules.
Risk characterization is the integration of the Exposure and Dose-Response assessments (Figure VI-1). This phase describes the probability of illness caused by consumption of oysters harboring pathogenic V. parahaemolyticus, and discusses the impact of the risk assessment. In this section, the predicted number of illnesses associated with each region and season are presented based on the model assumptions discussed in the previous sections. The distributions of the probable number of illnesses that may occur associated with regional/seasonal harvests, presented here, were derived based on the projected number of occasions when raw oysters were consumed.
|
Figure VI-1. Schematic diagram of the V. parahaemolyticus (Vp) risk assessment model showing integration of all the modules. |
As stated previously, the overall structure of the risk simulation was divided into three modules: Harvest, Post Harvest, and Public Health. The Harvest Module simulated the variation in total and pathogenic V. parahaemolyticus densities as a function of underlying environmental conditions. Analysis of the effects of salinity on the levels of pathogenic V. parahaemolyticus in the Harvest Module, suggested that salinity was not an important variable. The salinity component was removed and a model developed for V. parahaemolyticus growth based solely on water temperature. The output of the Harvest Module was the distribution of total and pathogenic V. parahaemolyticus densities in oysters at the time of harvest. The Post Harvest Module simulated the effect of current oyster handling practices and the possible effects of mitigations to derive predictions of the distribution of total and pathogenic V. parahaemolyticus densities at time of consumption. The Public Health Module estimated the distribution of the probable number of illness, which may be expected to occur within any given region and season on the basis of the predicted distribution of pathogenic V. parahaemolyticus densities at time of consumption. Throughout the simulation we utilized derived distributions of influential parameters and relationships between parameters to obtain estimates of the distribution of pathogenic V. parahaemolyticus densities at various stages along the pathway from harvest to consumption. Figure VI-1 shows a schematic representation of all the parameters used in the simulations for each module and how the output of each module becomes a parameter for the following module.
The distribution of ingested dose of pathogenic V. parahaemolyticus per serving was calculated by combining the distributions of meat weight per serving and distributions of density from the Post Harvest model. This was accomplished by following the Monte Carlo method of resampling from these input distributions and multiplying the sampled values to generate the distribution of consumed doses. Simulated samples from the distribution of ingested doses were then converted to a corresponding distribution of risk per serving. For each region and season, the number of illnesses occurring in 100,000 servings was then simulated as a sequence of independent, but not identically distributed Bernoulli random variables. Although larger simulation sizes are possible, the need to run multiple simulations (e.g. different regions and seasons) necessitated limiting the size of each simulation to a practical level. Predicted numbers of illnesses were determined by projecting the number of illnesses per 100,000 to the number of servings estimated based on the NMFS landings statistics. The simulations were repeated 50 times to simulate the year-to-year variation in the number of illnesses which may be expected due to the year-to-year variation in such parameters as the distributions of water temperatures and percentage of total V. parahaemolyticus which are pathogenic. The website where the worksheet showing the formulae and parameters used for the model can be found is in Appendix II.
Figure VI-2. Effect of structural uncertainty of dose-response on projected number of illnesses associated with V. parahaemolyticus (Vp) consumption in the Louisiana Gulf Coast, summer harvest). |
The distribution of the probable number of illnesses predicted by the Monte Carlo simulation was somewhat sensitive to the choice of the dose-response model. The difference in model predictions is shown in Figure VI-2 for the Gulf Coast (Louisiana) harvest during the summer. These histograms of the distributions are approximations based on 50 repeated simulations with randomly varying factors such as mean seasonal water temperature and percentage of total V. parahaemolyticus which are pathogenic. As shown in Figure VI-2, the predictions of illness under the Gompertz dose-response model are considerably more variable than under either the Beta-Poisson or Probit dose-response models. Preliminary simulations had suggested that a 10-fold shift of the ID50 (e.g. in regard to possible food matrix and immunological effects) would be sufficient to make model predictions generally consistent with CDC estimates of annual illness. However, the extent to which the estimate of the ID50 for the human feeding trials underestimates the "true" population ID50 is not known. If, for example, the difference in ID50 in the feeding trials versus conditions of general exposure was 10,000-fold, then it would not be the case that a similar shift of the ID50 would make model predictions generally consistent with CDC estimates regardless as to choice of dose-response model. For the present, the uncertainty in the magnitude of the ID50 under conditions of general population exposure was not fully evaluated.
For example, as shown in Figure VI-2, the expected number of illnesses (i.e., mean of the distribution) associated with summer harvest in the Louisiana Gulf Coast, predicted by the Gompertz, is approximately 3,300 which is slightly higher than the expected 2,400 cases predicted by the Beta-Poisson. The expected number of illnesses predicted by the Probit model is 270, which is considerably less than that of the other two models. The results indicate that a 10-fold increase of the ID50 was sufficient to make model predictions consistent with CDC estimates for either the Gompertz or Beta-Poisson dose-response models. A larger shift of the ID50 would be necessary to predict the same number of illnesses based on the Probit dose-response. Also, as evident in the figure, there is considerable variation about the expected values; particularly for the Gompertz model, which is more heavily skewed towards higher values. This variation represents the "statistical" uncertainty of the dose-response parameters under each model (i.e., distribution of bootstrap parameter estimates) as well as the effect of other influential variability parameters, particularly the distribution of water temperature. Figure VI-2 suggests that, congruent with Figure V-4, there is considerably more statistical uncertainty associated with the Gompertz dose response than the other two dose-response models.
The distribution of probable number of illnesses predicted by the Beta-Poisson model for the whole Gulf Coast, Pacific Northwest and Mid-Atlantic regions are shown in Figures VI-3 through VI-6. Clearly, the largest numbers of projected illnesses were attributed to the Gulf Coast harvest. For the Gulf, the average number of illnesses projected to occur associated with current levels of consumption is 25 during the winter; 1,200 during the spring; 3,000 during the summer, and 400 during the fall. The spread of the distribution is approximately 2-fold. For the Pacific, the average number of illnesses projected to occur is approximately 15 during the spring and 50 during the summer. The variance of the projected number of illnesses for the Pacific during the summer is comparable to that projected for the Gulf. This similarity is primarily a consequence of the assumptions concerning the year-to-year variation of percentage of total V. parahaemolyticus which are pathogenic. As discussed in the Harvest Module section, it has been assumed that the extent of year-to-year variation of this parameter relative to the mean is the same in the Pacific as in the Gulf.
Figure VI-3. Probable number of V. parahaemolyticus (Vp) illnesses associated with spring and summer Gulf Coast harvests.
Figure VI-4. Probable number of V. parahaemolyticus (Vp) illnesses associated with fall and winter Gulf Coast harvests.
Figure VI-5. Probable number of V. parahaemolyticus (Vp) illnesses associated with spring and summer Pacific Coast harvest.
Figure VI-6. Probable number of V. parahaemolyticus (Vp) illnesses associated with spring and summer Mid-Atlantic harvest. |
Model-predicted illness associated with the Pacific Coast harvest during the spring and summer, predictions are low relative to epidemiologically based estimates. From 1990 to 1996 an average of 8 culture-confirmed cases were reported in Washington State during the summer months (72). This statistic includes recent El Nino years that are not typical but it does not include confirmed cases in other Pacific Coast states, or the outbreaks in 1997 and 1998. Consequently, in a typical summer season at least 8 culture-confirmed cases are expected in the Pacific Northwest. This corresponds to about 160 cases assuming that illnesses are culture-confirmed at a rate of 5%. Consequently the average model-based prediction (50 illnesses) is 3-fold lower than the estimate based on the epidemiology. With regard to this discrepancy, it is possible that intertidal exposure of oysters to ambient air temperatures may have an appreciable effect on total and pathogenic V. parahaemolyticus densities in the Pacific growing areas (56). This phenomenon is not reflected in the present risk assessment where V. parahaemolyticus densities at harvest are predicted based on water temperature alone. It is also possible that pathogenic strains in the Pacific (urease positive) are somewhat more virulent than those strains indigenous to the other regions of the country.
In comparison to the relatively large number of illness projected for the Gulf harvests, expected number of illnesses projected for the Mid-Atlantic harvest are 10 during the spring and 12 during the summer. As indicated in Figure VI-6, the occurrence of 30 or more illnesses associated with Mid-Atlantic summer harvest is predicted to be a relatively rare event. Expected numbers of illness likely to occur due to current levels of consumption of Northeast Atlantic oysters are estimated to be 12 during the spring, 30 during the summer, and 7 during the fall. For the winter, the distribution of probable number of illnesses for Northeast Atlantic oysters is below the resolution of the simulation based on simulation of 100,000 servings. Taking into consideration the NMFS landings data (104), and then assuming 50% consumed raw plus average serving size of 12, in the Northeast Atlantic region, many more oysters are consumed in the fall (2.2 million oyster servings) vs. the spring (400,000 oyster servings). Consequently, the predicted mean number of illnesses is more in the fall than in the spring for this region in consideration of the number of servings estimated.
The low numbers of projected illnesses due to Northeast Atlantic and Mid-Atlantic oysters is attributable to both the colder water temperatures and the relatively modest harvest from these regions during the warm summer months. Considering the extent of underreporting of illness that is likely to occur for self-limiting gastroenteritis due to V. parahaemolyticus infection, these low numbers of predicted illness appear to be generally consistent with the infrequency of culture-confirmed illness associated with oysters harvested from these regions. Underreporting and infrequency of culture-confirmed illness may be partly caused by the unfamiliarity of consumers and healthcare practitioners with Vibrio- associated illnesses, and lack of expertise in diagnosis and in laboratory detection (34).
Based on the assumptions outlined previously with regard to description of the Public Health Module, the projected distribution of probable number of cases with septicemia which may occur in any given year is shown in Figure VI-7. Most of these cases are predicted to be associated with Gulf Coast oyster harvest with few cases due to Pacific Northwest harvest. Although the most probable number of cases of septicemia per year associated with ingestion of V. parahaemolyticus, is 4, the overall mean of the distribution is 6 cases per year, based on a sample of 50 replicated simulations. As evident in the figure, the occurrence of 15 or more cases of septicemia in a single year for the entire country is projected to be an infrequent event (i.e., an event with probability less than 0.1).
Figure VI-7. Distribution of probable number of cases of V. parahaemolyticus-associated cases of septicemia occurring per year (all seasons and regions). |
The effect of three Post Harvest mitigations was evaluated in the simulation: (a) mild heat treatment (5 min at 50°C), (b) freezing (-30°C), and (c) rapid cooling immediately following harvest (e.g., aboard ship). As discussed in the Post Harvest Module, the effect of mild heat treatment has been shown to reduce the density of V. parahaemolyticus to nondetectable levels (at least a 4.5 log10 reduction) and freezing at -30°C has been shown to reduce the density by approximately 2 logs.
Figure VI-8. Effect of potential mitigations on the distribution of probable number of illnesses associated with V. parahaemolyticus (Vp) in oysters harvested from the Gulf Coast in the summer. |
All three potential mitigation strategies have a substantial effect on the distribution of probable number of illnesses. The effect of these mitigations was evaluated under the assumption of the Beta-Poisson dose-response model. For the Gulf Coast summer harvest (Figure VI-8), a shift in the distribution of probable number of illnesses down from a mean of 3,000 illnesses to approximately 240 illnesses is predicted under the mitigation of rapid cooling. The mean number of illnesses projected to occur under the freezing mitigation is approximately 15. As evident in Figure VI-8, the variance of the predicted distribution is also reduced under these mitigations.
The effect of mild heat treatment was found to reduce the mean risk of illness per serving to substantially less than 1 in 100,000. As discussed previously, practicalities of performing the Monte Carlo simulation necessitated limiting the simulations to 100,000 servings per season/region combination and then projecting to an estimated number of servings. Consequently, the distribution of probable number of illnesses under mild heat treatment mitigation could not be accurately determined. For the Gulf Coast summer harvest the mean of the distribution is certainly less than 10 illnesses.
With the exception of rapid cooling, the effect of these potential mitigations on the number of illnesses is similar for other regions and seasons. The relative effectiveness of rapid cooling in the Pacific Northwest is predicted to be much less than in the Gulf Coast or Mid-Atlantic regions due to cooler air temperatures and a shorter duration of harvest. The effects of the mitigations on the mean risk per serving are shown in Figures VI-9 through VI-11 for the Pacific, Mid-Atlantic and Gulf Coast harvest for all seasons. As evident in these figures, the effectiveness of mitigation is more pronounced during the summer since the potential for growth is much greater during this time of the year.
Figure VI-9. Effect of potential mitigations on the distribution mean risk of V.
Figure VI-10. Effect of potential mitigations on the distribution mean risk of V.
Figure VI-11. Effect of potential mitigations on the distribution mean risk of V. |
FDA had previously indicated that V. parahaemolyticus in shellfish should not exceed a level of 10,000 viable cells per gram (64). The 1999 V. parahaemolyticus Interim Control Plan adopted by the ISSC includes the 10,000 viable cells per gram guidance. In areas where levels of greater than 10,000 cells/g oyster tissue are found, the area would need to be resampled. While the critical cause of illness in humans is the level of pathogenic V. parahaemolyticus, the total V. parahaemolyticus is used as a convenient surrogate indicator of higher risk of illness. The risk assessment cannot critically evaluate the control plan because, as the model is constructed, there is no mechanism included to account for the possibility of persistence of V. parahaemolyticus in specific oyster harvesting areas. Moreover, the rapidity or sensitivity of tests performed by individual laboratories for detection of V. parahaemolyticus is not well determined. Both of these factors are critical in evaluating the control plan.
The risk assessment does, however, allow us to ask what would be the predicted impact on the incidence of disease if we were able to exclude oysters at the time of harvest that had a specified level of V. parahaemolyticus. This also includes estimating the impact of what excluding oysters that had a specified level of V. parahaemolyticus would have on the percentage of oysters that would no longer be available. The impact of such a criterion was evaluated using parameters from the Gulf summer (Louisiana) region/season parameters. A total of fifty simulations of 33,333 iterations (Monte Carlo samples) were run individually.
The results were sorted by whether or not they caused illness and then the initial V. parahaemolyticus levels in the environment (i.e., at harvest) were sorted into "bins" of half log intervals. The proportion of illness associated with each half-log interval of initial V. parahaemolyticus level was calculated and the results were then used to estimate the potential effect of various guidance levels for "at harvest" densities on reduction of illness and associated cost in terms of percentage of total harvest lost (e.g., diverted from raw consumption market).
The results are shown in Figure VI-12 for the 10,000 per g standard, with levels of 100, 1,000 and 100,000/g included for comparison. An "at harvest" perspective was adopted here due to the fact that the current guidance level of 10,000 viable cells per g included in the 1999 V. parahaemolyticus Interim Control Plan pertains to oyster samples obtained at time of harvest. Although the guidance level of 10,000 viable cells per g may apply to shellstock at any time post harvest, current monitoring efforts are generally directed towards "at harvest" samples. Consequently, the potential effects of different guidance levels for monitoring of shellstock at the wholesale or retail level was not evaluated.
Figure VI-12. Potential effect of control of total V. parahaemolyticus per gram |
On average, the simulation results suggest that 15% of the illnesses are associated with the consumption of oysters for this region/season combination that contain greater than 4 log10 (104) V. parahaemolyticus per g at time of harvest. The corresponding fraction of the harvest containing greater than 4 log10 (104) per g was 5%. Therefore, if all shellstock could be evaluated for total V. parahaemolyticus at time of harvest, the simulation results suggest that excluding all oysters that had levels of 10,000 viable cells per g would reduce (sporadic) illness by 15% at a loss of 5% of the total harvest from the raw consumption market. This relatively low (potential) reduction of illness is attributable to the large proportion of the harvest that would remain with a lower, but still significant, associated level of risk. In comparison, the simulation results suggest that in the absence of subsequent post harvest mitigations, "at harvest" guidance levels of 5 log (105), 3 log (103) and 2 log (102) total V. parahaemolyticus per g could (potentially) reduce the illness rate by 2%, 50% and 90% with corresponding losses of 0.3%, 25% and 70% of the harvest, respectively.
A tornado plot is a convenient way of describing the factors in our model that most affect the results. The plot is called a "tornado plot" because of the similarity of the image of a tornado with the graphical arrangement of the factors from most influential at the top to least influential at the bottom. In our model, we see that the most influential factor driving the results for all the harvesting areas during the summer (Figures VI-13, VI-14, VI-15, VI-16) is the levels of V. parahaemolyticus. Because our model assumes that the rate of increase of pathogenic V. parahaemolyticus, is the same as that for total V. parahaemolyticus, then V. parahaemolyticus-associated illness results when the level of V. parahaemolyticus increases, thus increasing the levels of pathogenic V. parahaemolyticus.
In the Gulf Coast, both Louisiana and the remaining Gulf Coast regions, where the temperatures are the warmest compared to the other regions, time to refrigeration was determined by sensitivity analysis to be the second most important effect on occurrence of illness (Figures VI-13, VI-14).
Figure VI-13. Tornado plot of influential parameters on log10 risk of V. parahaemolyticus (Vp) illness per serving of raw oysters (Gulf Coast excluding Louisiana, summer oyster harvest). |
Figure VI-14. Tornado plot of influential parameters on log10 risk of V. parahaemolyticus (Vp) illness per serving of raw oysters (Louisiana Gulf Coast summer harvest). |
Figure VI-15. Tornado plot of influential parameters on log10 risk of V. parahaemolyticus (Vp) illness per serving of raw oysters (Pacific Northwest Coast summer harvest). |
Figure VI-16. Tornado plot of influential parameters on log10 risk of V. parahaemolyticus (Vp) illness per serving (Mid-Atlantic summer harvest). |
The other factors analyzed have significant effects, but to a lesser extent. As one would expect, the more oysters one eats, the more likely it is that one will become ill. Also, unsurprisingly, conditions that allow for the V. parahaemolyticus to grow within the oyster (length of time oysters are unrefrigerated, time it takes to cooldown the oysters, water and air temperature) increase the risk of illness. Since the levels of V. parahaemolyticus decrease during cold storage, the length of time the oysters are refrigerated is negatively correlated with the risk and the factor points on the tornado plot in the opposite direction.
Reasonableness of model predictions and the appropriateness of the modeling assumptions that have been used in the risk assessment, can be evaluated by comparing model output to relevant data that were not used to develop the relationships and distributions of parameters in the model per se. With regard to the prediction of illness there is no independent data available for this purpose. As indicated previously, with due consideration of estimated levels of pathogenic V. parahaemolyticus at time of consumption and oyster landing statistics, the dose-response under conditions of feeding trials is not consistent with CDC estimates of annual illness. Consequently, the epidemiological data were used to adjust the dose-response from the conditions of human feeding trials to conditions of population exposure and do not constitute a point of validation of the model. Independent data are available on the levels of total V. parahaemolyticus at retail and these data have been compared to model predictions in order to assess the appropriateness of the model with respect to the Harvest and Post Harvest Modules.
A collaborative nationwide survey of V. parahaemolyticus densities in oysters at the retail level i.e., restaurants, oyster bars, wholesalers, etc., was conducted by the ISSC and FDA in 1998 and 1999 (44). A total of 370 oyster samples were collected during the study and the harvest state was identified for all samples. This study provides the most comprehensive information available on seasonal and regional differences in density of total V. parahaemolyticus at time of consumption. In particular, this information provides a point of empirical validation of the assumptions used in the Post Harvest Module to predict the extent of growth that occurs post harvest. To facilitate this comparison, simulations of the distribution of total V. parahaemolyticus densities were carried forward through the simulation in addition to the distribution of pathogenic V. parahaemolyticus.
Figure VI-17. Observed retail level distribution of density of total V. parahaemolyticus (Vp) compared to model predictions for all seasons (Gulf harvest) (28).
Figure VI-18. Observed retail level distribution of density of total V. parahaemolyticus (Vp) compared to model predictions for all seasons (Mid-Atlantic harvest) (28).
Figure VI-19. Observed retail level distribution of density of total V. parahaemolyticus (Vp) compared to model predictions for all seasons (Pacific harvest) (44). |
In the ISSC/FDA study, V. parahaemolyticus densities were enumerated by an MPN method. A relatively high proportion of the non-Gulf Coast samples had nondetectable levels. To appropriately adjust for the varying proportion of nondetectable V. parahaemolyticus across the different regions and seasons, estimated means were obtained by fitting a Tobit regression to the complete data set (n=370) with different harvest region and season combinations as a predictor variable. The variance about the group means was assumed to be homogeneous i.e., the same across different regions and seasons. The limit of detection varied somewhat from sample to sample but was generally 0.18 MPN/g. Clearly, estimates of regional/seasonal means near or below this threshold are an indication that a high proportion of samples from that particular grouping were not detectable and the estimate of the mean is strongly influenced by the assumptions underlying the Tobit model. In particular, estimates for the Pacific Coast were poor due to lower levels of V. parahaemolyticus in that region year round and the low number of samples obtained from the West Coast during the study.
Comparison of estimates of mean and population standard deviation of log10 total V. parahaemolyticus densities based on the ISSC/FDA study versus model predictions are shown in Figures VI-17 through VI-19 for the Gulf Coast, Pacific Northwest, and Mid-Atlantic harvest regions. In so far as the mean and standard deviation of the log10 densities predicted by simulation varies from year to year due to environmental conditions, model predictions are presented after averaging out year-to-year variation. A certain degree of deviation from ISSC/FDA estimates is to be expected in this comparison since the empirical data were obtained over the period of a single calendar year.
Generally, the estimates of the means based on ISSC/FDA data compared well with those predicted by the simulations. In particular, model predictions of mean log10 densities are in good agreement with ISSC/FDA data for both the Gulf and Mid-Atlantic regions during the summer when the risk of illness is highest. For the Gulf Coast, model predictions of mean log10 densities in the fall are somewhat lower than those obtained by the ISSC/FDA study. With regard to this discrepancy, water temperature measurements indicate that the fall season of 1998, corresponding to the time of ISSC/FDA sampling, was somewhat warmer than usual. The NBDC buoy at Dauphin Island has been disabled since September 1998; however, the average daily noontime water temperatures in nearby Weeks Bay AL (an NERR site) during the fall of 1998 was 22.5°C, compared with a typical average of 18°C at Dauphin Island. A difference of 4.5°C corresponds to an average of 0.50 log10 higher density of total V. parahaemolyticus at time of harvest. Furthermore, warmer air temperatures would entail more post harvest growth. From January 1999 through September 1999, corresponding to the remaining period of ISSC/FDA sampling and the other seasonal comparisons, water temperatures in Weeks Bay did not differ greatly from the overall averages measured at the Dauphin Island buoy (i.e., ~1°C difference).
For the Pacific Northwest harvest, average model predictions were higher than the estimates based on ISSC/FDA data. The difference in the summer appeared to be in the range of what can be expected due to year-to-year fluctuations of environmental conditions. However, model predictions during the spring are much higher than those based on the ISSC/FDA data. A possible explanation for this discrepancy is the lack of precision associated with the estimate based on ISSC/FDA data. The number of samples on which the estimate is based was very small and consequently, due to the generally low levels of total V. parahaemolyticus in the spring season, the estimate of the mean is poor. Estimates could not be obtained for the Pacific winter or fall season. Similar results were obtained when considering Northeast Atlantic harvest (data not shown). Model predictions of mean density were consistent with the ISSC/FDA estimate for the summer but 1 to 1.5 log10 higher during other seasons of the year.
The common population standard deviation about regional and seasonal mean densities was estimated to be 1.5 log10 based on the ISSC/FDA data. This compares well with model predictions of the spread of the distribution when allowing for the fact that variance of measurements obtained in the survey are also a reflection of method error which has been adjusted for in the simulation. The error bars in Figures VI-15 through VI-17 denote one standard deviation above and below the mean. The interval is generally larger for the ISSC/FDA retail study data than for the simulation output.
Additional simulations were performed to examine the effect of uncertainty and variability parameters on the variance of the distribution of probable number of illnesses obtained by simulation. These simulations were directed towards determining the influence of three parameters: (a) relative growth rate of V. parahaemolyticus in oysters versus broth model (axenic rate); (b) combination of variability and uncertainty in the overall percentage of total V. parahaemolyticus that are pathogenic (% pathogenic); and (c) variation of water temperature.
The average risk per serving for Gulf Coast summer harvest under a series of five alternative assumptions is shown in Table VI-1. Fifty repeated simulations were performed under each set of assumptions. In two of the series, the parameters were either varied according to their distributions as specified in the description of the risk assessment or they were held constant at their mean values ("vary all" and "hold all"). In the remaining three series of simulations one parameter was held fixed while the other parameters varied according to their specified distributions.
Table VI-1. Effect of selected uncertainty and variability parameters with respect to Gulf Coast summer harvest: average and relative variation of predicted risk per serving for the V. parahaemolyticus risk assessment model
| Vary allb | Hold axenic (4.0) | Hold % Vppath(0.2%) | Hold temp ( =28.9),  =1.5 |
Vary None (hold all) | |
|---|---|---|---|---|---|
| Average risk per serving | 0.00134 | 0.00150 | 0.00160 | 0.00163 | 0.00162 |
| Coefficient of variationa | 98.0% | 72.2% | 86.0% | 75.9% | 53.7% |
| a standard deviation divided by the mean
b baseline (vary all parameters), no growth rate uncertainty (hold axenic), average percentage pathogenic known (hold % path), no temperature variation (hold temp), no uncertainty/variation in all three parameters (vary none) | |||||
Due to differences in means for the five series of assumptions, the variation in the model output (risk per serving) is summarized in Table VI-1 by the coefficient of variation (the standard deviation of a distribution divided by its mean). A clear difference is evident between the distribution of risk per serving obtained when all parameters were varied versus that when all parameters are held fixed. The coefficient of variation is 54% when all three parameters are held fixed and 98% when all parameters are varied according to their specified distributions. Consequently, the three parameters considered here account for approximately 45% of the total variation in risk per serving associated with the Gulf summer harvest. Individually, the uncertainty in growth rate proportionality and percentage pathogenic account for 26% and 12% of the total variation, respectively. In comparison, water temperature, which is a variability parameter, accounts for 22% of the total variation. These results suggest that, of the uncertainty factors considered here, the variation in the output would be reduced the most by determining the appropriate proportionality constant between growth rate in oysters versus axenic culture ("hold axenic").
An additional and important source of uncertainty underlying the predicted distribution of illness is that associated with the dose-response extrapolation from frequency of illness in feeding trial studies to conditions of human exposure. The effect of this uncertainty has not been fully evaluated in the present risk assessment. Based on consideration of CDC illness estimates, a plausible shift in the ID50 was determined for which model predictions of illness were consistent with the CDC estimates. A more complete evaluation of this uncertainty could be examined as a refinement of the present risk assessment.
This model incorporates similar components of other risk assessments, but has several unique aspects. This model has analyzed risk in terms of region and season. Other microbial risk assessments have only looked at aggregate yearly risk. This model is scalable in that it may be applied to finer levels of analysis as data become available. Other microbial risk assessment models must be restructured to incorporate finer levels of analyses. This model has separated variability from uncertainty by identifying four key variables as uncertain, selecting values for these variables according to specific distributions, and inserting these values into a model simulation loop. In this manner, parameters that represent variability of the model are not mixed with parameters that are uncertain. This separation has allowed us to analyze the reduction in the overall uncertainty of the analysis that we will gain if the uncertainty of an individual variable is reduced. Other microbial risk assessments have separated variability from uncertainty; however, this risk assessment has investigated the gain in information that results from reduction in uncertainty of individual variables. We discuss each of these points in turn.
This model analyzes risk within the four seasons for five primary harvesting regions (Northeast Atlantic, Mid-Atlantic, Gulf of Mexico [divided into 2 regions], and Pacific Northwest), due to differing harvest practices and climates. We could have subdivided the analysis further; however, the limitations of acquiring the data for the next level down are such that we did our analysis at the regional level. Analyzing the regions separately allows the assessors to look at mitigations that may be tailored to specific regions and seasons. Those results may then be used in a subsequent cost benefit analysis. Other microbial risk assessments have not performed region/season analysis. This limits the range of possible mitigations that can be examined.
The model is amenable to further subdivision of locality and season because it is scalable. What we mean by scalable is that the model simulates from harvest to consumption of oysters with a specific level of V. parahaemolyticus. Given the existence of appropriate data, the model can simulate this process from any specific harvest location at any specified time. Our analysis (regional and seasonal) was done at the level at which we had the required data set of water and air temperatures, harvest practices, V. parahaemolyticus prevalence, and shellfish landing information and would give us a complete US risk assessment. We could have further refined our assessment if we had complete data for the individual states, but time and resources were limiting factors. If the data were available, the model could be applied to risk assessments at the state level, shellfish harvesting area level, or, still finer, division of shellfish location and season. New uncertainties will arise if the model is applied to harvesting areas for which the required data set is incomplete. We were not able to explore the effect of incomplete (or inaccurate) data on the results of the model at this time, but we may do analyses in the future.
We have separated variability from uncertainty in our analysis because it provides the best characterization of the available information. We distinguish between model inputs that are less well characterized because of our lack of knowledge (uncertainty) and model inputs that are heterogeneous (variable). By describing some model inputs as heterogeneous we mean that we characterize some of the model inputs as being naturally variable. For example, the water temperatures of the different regions will vary in the model according to a normal distribution with a given mean and standard deviation. At the same time, we have characterized the conversion factor between V. parahaemolyticus growth in axenic culture and V. parahaemolyticus growth in oysters as an uncertainty model input. We did this because of our lack of knowledge of the true conversion factor. With more study we could reduce our uncertainty. The result of our making the distinction between model inputs that are uncertain and model inputs that are variable is that we may analyze our model and examine the effect on output of reducing the uncertainty of each of the uncertainty variables separately. In this way we provide our risk managers with insight on which uncertainty has the greatest effect on the final results. Risk managers may then prioritize which uncertainty to reduce first in an effort to reduce the uncertainty inherent in the risk characterization of the risk assessment.
The model can be improved. At present, the model simulates risk for a set of uncertainty factors with defined variability largely based on the relationship of V. parahaemolyticus levels to temperature and a random selection (within defined limits) of percent pathogenic V. parahaemolyticus. The model does not allow a quantitative prediction of the reduction in risk resulting from implementation of the FDA/ISSC V. parahaemolyticus interim control plan (adopted by the ISSC in July 1999), because we cannot model the effect on the risk from embargoing subsequent harvesting after measuring an unsafe level of V. parahaemolyticus in a shellfish harvesting area. For us to be able to model the plan, we have several modeling needs: first, we need to scale the model to the shellfish harvesting level. To do this we need complete data sets for the individual harvesting areas. Second, we will need sensitivity and specificity data for the virulent V. parahaemolyticus gene probe methodology by the individual laboratories doing the tests. Third, we will need to extend the model to account for rapidity by which pathogenic V. parahaemolyticus levels change in specific areas. At present the model predictions are primarily based on temperature plus a random factor for population variation. The model might be improved by tailoring the rapidity of turnover of water in the shellfish harvesting area based on levels of freshwater flows, tide changes, wind direction, and depth of harvesting area. These are all areas of planned future study.
Hypertext updated by bap/kwg/cjm/dms 2001-JAN-30
); freezing (
); heat treatment (
); rapid cooling (
).
