Animal & Veterinary
Presentation of CVM Risk Assessment
Dr. David Vose
DR. VOSE: Thank you. Good morning.
The CVM risk assessment, what I am going to try and cover in the 40 minutes that I have got is, first of all, what we modeled and why, the logic associated with that model. And I am sure that that appears to something of a black box to at least a few of you.
I am going to talk the results that we have gleaned so far, uncertainty analysis which is a large part of what we have been doing, recognizing the degree to which we do and don't know. And as Wes pointed out in his presentation, that a great deal of the value of risk analysis is to work out what it is you know and don't know. And I am also going to describe how one might use the model in brief form to help make your regulatory decisions.
Well, first of all, of course, I have to recognize the team that I have been working with. First of all, Sharon Thompson who is my boss so she comes at the top of the list. Sharon took over with this project halfway through from Peggy Miller. And I take my hat off to her because that is a tough job to do. It is halfway through and she suddenly has got to understand what we have been doing. And it was a very complicated problem that we had to deal with.
I also have to thank Peggy Miller who was the initiator of this project. And I have to recognize to Peggy that she was the person who originally thought of this approach to assessing this risk, as much as I would have liked it to have been me. I simply executed what was a very clever idea from her.
There is me, the consultant, of course ---. Kathy Hollinger -- just in case you don't know because you will end up in the wrong place if you don't know that, somewhere in Germany.
Okay. Kathy Hollinger, as Dr. Sundlof has said in his opening remarks, Kathy put an awful lot of effort in. And she sort of reminds me of a bulldog. I am British. And so she has the tenacity of the bulldog who will go out and just keep collecting information and not be satisfied. She would often come up with a comment to me, "But it is not that simple, David", which gets very irritating because I would like it to be. It's a model. But all power to her. She kept me on line.
As did Mary Bartholomew who spent a lot of time helping collect the data and analyzed the data that was given to us in forms that weren't necessarily exactly what we needed. In quantitative risk analysis, you need numerators and denominators very often because you want to work out uncertainty.
People will tell you, "Oh, well, we found 30 percent resistance." They don't like to tell you that they only checked five chickens. So we need numerators and denominators if we are to say what that means. And Mary has done a great deal of helping obtain that information.
Okay. This is the only slide with this much information on it. So I apologize. Why do we model fluoroquinolone-resistant Campylobacter in chickens? Well, this was originally set up as a pilot study to determine the feasibility of doing the risk assessment on antimicrobial, bacterial, blah, blah, blah.
We wanted to look at the data needs that would incur and we wanted to look at the source of information where we may be able to find that data. As others have pointed out, Campylobacter is the most commonly known cause of bacterial food-borne illness in the U.S.
Ninety-nine percent of Campylobacteriosis are sporadic illnesses which makes life a lot easier from a mathematical point of view. If they were these outbreaks, then we would have a more difficult problem.
Chicken is, as others pointed out, the most commonly identified risk factors for Campylobacteriosis in the U.S. It has been -- Campylobacter has been reported to develop resistance quickly to fluoroquinolone which, again, makes our problem much more simple. Fluoroquinolones are important antimicrobials, of course. It is a valuable drug to us and we want to make sure that we guard the value of that drug.
And most importantly, we felt that certainly as we started to move along this part, we felt that there was enough data in order to produce a meaningful quantitative risk analysis. I am a quantitative risk analyst. I am involved in the mathematics of things.
Another option is to go down the qualitative route where you just simply identify the factors and talk descriptively about the problem. And other organizations have done that.
Okay. Well, this risk assessment modeled direct transfer of resistance because fluoroquinolone resistance is on the chromosome. It is not transferred to other bacteria. This is something I know absolutely nothing about. But because it is not a two-step process, it makes, again, our mathematics a little simpler.
You can see that we have picked out this particular problem for two reasons then, Campylobacter fluoroquinolone resistance in chicken. A) Because it is a big issue. But B) because there is data there. And C) the math makes it feasible.
Now, we are going to try some further analyses on the risk initiatives underway to look at other microbial resistance issues such as indirect transfer. That may or may not be something that we can feasibly do quantitatively. But we are certainly not going to start out saying yet we are going to be able to do everything else quantitatively because we could do this one so.
But -- so the point to take away I suppose here is that if we couldn't have done it quantitatively on this risk issue, we certainly wouldn't be able to do it on the others. But we can, so we have got some feeling of security that we can proceed on.
Okay. The problem we modeled, imagine you have poultry in a shed. They get some disease, e.g. collibacillosis. I probably said that wrong. They are all treated with a fluoroquinolone. Then that fluoroquinolone-resistant Campylobacter, it proliferates in the drug because -- sorry, proliferates in the poultry gut because all of the other bacteria have been erased.
Then us humans go and eat that chicken and they get contaminated with that Campylobacter. And then they go to the doctor and the doctor says, "Oh, you are ill. Take some fluoroquinolone." And nothing happens. So to how many people would that occur is what we are trying to work out.
Now, there will be a lot of people I think who would criticize this model because it is not a predictive microbiological model. A predictive microbiological model would say, for example, look at the number of pathogenic organisms in the chicken and then flow through, see how many were gotten rid of in chillers that the gentleman in the back was talking about through evisceration, etcetera, etcetera.
How many would be lost through natural attenuation of the numbers from chilling or freezing, and then the cooking. And, oh, it just goes on and on. I mean, you can think of so many things. Even if we just dealt with the one last issue. Here is a quantity of chicken that has fluoroquinolone-resistant Campylobacter on it and you go feed it to someone. Well, who do you feed it to. You know, if I gave all of you out here the same dose with the same pathogenicity, you would have varying reactions.
There would be any number of you who would have light illness. Some would have no effect at all. And it would depend on, for example, what you had -- when you had your cup of tea, did you have some yogurt if there was any out there? Did you -- have you had a full meal? Have you had nothing to eat yet this morning like me, etcetera, etcetera.
So it is an extremely complicated problem if you want to look along the microbial part. And certainly from the point of view of the regulator, the Food and Drug Administration here, it really wasn't relevant to look at all of those parts.
Now, from the point of view of industry, I can quite imagine that they would want to work out ways that they can try to reduce the number of bacteria that actually were loaded in their chickens. Absolutely right. It is fair to say is it fluoroquinolone that should be used or should we try and work out ways of reducing its use; is there any effect on the chicken population. Right.
So we chose this rather simple model as being the most appropriate. Now, although it is simple, we can make corrections to the original assumptions for changes in the system. For example, if changes in human feeding patterns -- if we eat more chicken or less chicken or if we tend to eat it more cooked or less cooked, things like that. We can probably start making some kind of fudge factors, but reasonable guess fudge factors that will allow us to update our model as the system changes, if it does.
But the essential real benefit of this is it provides a responsive means of continually assessing the risk. By responsive, I mean if we keep monitoring the problem, we can assess month by month or quarterly by quarterly, we can assess the size of the risk.
Now, if we had gone down to a predictive microbiological model with so many changes to the system like they change the number of chillers that they use or the frequency with which they clean them out, well, we would have to go all the way back and do a much more complicated analysis.
So the point of this is it is easy to use. And we can get a quick idea of the size of the risk that we are exposing the U.S. population to.
Okay. Now, to my mind, this risk analysis -- this microbial risk analysis is unique in that we found data to quantify all the model parameters. I say unique because I have been involved in a number of microbial risk assessments including the United States of America. And almost always -- well, always we have somewhere along the line to make a guess. We have to assume something that we really wouldn't like to have to assume. We have to use a surrogate bug for the dose response, etcetera, etcetera.
Well, in this particular risk assessment, we have thanks largely due to Kathy and Mary found data to quantify every single parameter. And that data has come from a number of sources, from FoodNet surveys, physicians' reports, CDC's attempt at a case control study, NARMS, from poultry industry, data on consumption and production, U.S. population records, etcetera.
Data didn't just have to be collected, but it had to be collected in a form that allowed us to perform uncertainty analyses. So we had to dig out not just the information like prevalences and percentages. But we had to, as I said before, talk about numerators and denominators.
Now, given all of that, 1998 was the first year that we were able to produce a complete set of data. So we had both sides of the equation that I am going to talk about in a minute, we had data for everything. What I had originally imagined doing and I had hoped that we would be able to achieve is to compare several years of data from the past. And we would get a much more firm understanding of what was going on.
So I suppose at this point we are in the first year of what I hope will be several years of data collection that will make us more and more able to understand the connection between Campylobacter-resistant fluoroquinolone in chickens and the effects on the chickens.
All right. If you read through the risk assessment report that we have done, probably a lot of you will be confused about this quantifying uncertainty. Uncertainty is about the state of our knowledge. There is in theory some parameter value that is out there that could be known. But we will never have perfect data. We will never have perfect information about that parameter.
And if we just take at face value some of the data that we have when we have a very small amount of data, we can be very wrong. We can be overly conservative. We could be overly pessimistic. We don't know. But we would be very wrong if we just take the data at face value.
If I toss a coin three times and I get two heads, you are not going to tell me that the probability of the heads is 66 and two-thirds percent. It wouldn't make sense. Well, that is the same principle.
In this particular problem -- analysis, we used a Bayesian approach. And there were good reasons for that. First of all, it allows us to combine dissimilar data. So we were able in a couple of instances to take a set of information over here with a particular certainty with information involved in both of those two different studies.
A potential criticism of the Bayesian analysis is that we have to introduce something called a prior distribution. And that would introduce a very small bias. And Dr. Cox, who is following me here this morning, will probably mention that being a Bayesian mathematician.
But having said that, the data set sizes mean the results pretty much equate to the classical statistics estimates which is perhaps the things that you remember from university and certainly less controversial, although Bayesian inference is certainly growing in use.
And so quantifying uncertain analysis not only tells us how much we really know and how good our predictions can be. But it also tells us where we should be able to collect more data and how it would be useful.
So here is an example. This is a distribution of uncertainty about a particular probability. And you have -- I'll get my laser pen here. You have three distributions here. The first one, which is this broad curve here, is talking about your estimate of a probability.
If you were, say, to take -- go to a population and say -- oh, let's talk about Republicans and Democrats -- ask four people, "Are you going to vote Republican or Democrat?" -- and I can do that because it is 50/50, so I am all right. Two say Republican and two say Democrat.
Well, if I am trying to extrapolate to the true population, I know that I don't really know very much about the proportion of people that are going to vote Democrat or Republican. And so this description here is describing the amount of uncertainty.
Well, it is pretty much somewhere between zero and one, not very sure. But as I accumulate more data, I go through this -- the beta (3,3) is talking about four people, two of each side; a beta (11,11) is 20 people, ten on each side. And you can see my distribution is becoming a bit narrower.
And then here we have got a beta (21,21) which is 20 people of each side. So 40 people are asked and 20 people said Republican, 20 Democrat. And there you have a much narrower level of uncertainty. So the point of it is that if we accumulate more data, so we become progressively more certain about what the truth is out there.
For those of you who are more technically inclined, here is a little graph to show that although Bayesian inference has a slight bias to it, the classical statistics of an estimate for this particular type of problem, when you had four people, two Republicans and two Democrats, well, the classical statistical estimate will be this thing here, this red line.
It is a binomial distribution. And there is an approximation there in blue which is the normative approximations of the binomial versus this green line which is the Bayesian estimate.
Well, what I am trying to show here is that with this red step line, that is the perfect classical estimate as they call it. And yet they frequently represent that with this blue line. It is a little more helpful for them for a majority of the analyses they do. So if a classical statistician is willing to take this step line here and make it into a blue, then going from blue to the green, that is not a big deal.
More importantly, as your data sets become bigger, so the difference between this three of them, and you can't see the blue and the green, the classical versus the Bayesian. They just completely overlay on each other. And that is not even for a very large number of data points, just 20 data points.
So there isn't really any controversy between Bayesian and classical inference in this particular model.
Okay. Now, I do -- the difficulty that people will have I think in understanding what I have tried to d here is looking at this idea of a nominal expected number of people who will come out with Campylobacteriosis. I say nominal because I wasn't really very interested in the actual number of people.
CDC put a lot of analysis into trying to determine the true number of people out there in the population of America who got ill. Well, I was more interested in something called the intensity of that system because I want to know whether if we were to take that same number -- that same system and one year we note that 30 people became ill. Well, the next year we are not going to note the same 30 even though there was the same risk out there. Maybe it is going to be 35. Maybe it is going to be 25. I want to know that if you were able to repeat that year many, many times, what would the average be which is my much better estimate of the true risk to the human population.
So here is an example of a Poisson distribution which is the appropriate distribution in this circumstance. And you can see, I have got -- this is the probability. And for a given intensity -- this is for a given size of risk if you like. On average risk, two people per year would die, whatever, ill.
Then you would see we could quite easily have zero people one year or we could have one person or two persons or three or four all with the same amount of risk. And yet we can observe different things from one year to another. And that gives you some idea that we should be a little bit cautious about interpreting changes, reasonably small changes from one year to the other in what we observe in the illness out there because it could simply just be a sampling error. It is just that we -- it's just there is so much randomness out there, it is quite possible you will have a small sample one year and a larger one for the other, and yet have the same level of risk.
So I am very keen that when we do this risk assessment, we use it to quantify the risk. But we should be completely cognizant of the randomness that is out there that could if we are not careful sway us from making overly cautious decisions or underlie cautious decisions. And the purpose of doing the uncertainty analysis was to stop us from doing that.
Okay. Model overview, how I set this model up was, first of all, to look at the number of Campylobacter culture confirmed cases observed in the U.S. population. And this comes entirely from CDC data except that I am interested in the nominal expected number.
So I am interested in that two value if you like from that Poisson distribution versus the actual observed numbers. So I am trying to get a sense of how many people out there are getting those Campylobacter cases.
And from there, this is in Section 2, I am looking at the total number of Campylobacter infections in the U.S. population. So it is the total number of Campylobacter infections in the U.S. rather than those that were culture confirmed cases because culture confirmed cases are the only ones that you actually observe in your health system because they have to be identified. You have to get them, thus, in scooping the poop and doing the microbial analysis.
So we extrapolate from there to work out the total number of people that are ill in the population. In Section 3, I am looking at those -- the number of those people who would have been ill from the fluoroquinolone-resistant Campylobacter because, clearly, those are the people who would be at risk.
And I want to see how many of those who were infected with the fluoroquinolone-resistant Campylobacter then went to the doctor and were prescribed an antibiotic and that antibiotic happened to be fluoroquinolone because, clearly, those are the only people out of everyone that had Campylobacteriosis, those are the only people who are going to have any observable difference in their final outcome.
Over here in Section 4, I am looking at the number of -- the quantity of meat consumed, of chicken meat consumed that is contaminated with fluoroquinolone-resistant Campylobacter. And the idea is to say if we take the Poisson intensities if you like of those two things, we can correlate them together in a sort of generic dose response model.
And with a constant of proportionality, we can estimate or we can relate the human health cases to the chicken. So Section 5 deals with how we go about making that connection.
Okay. So let's deal with Section 1 quickly. In a fairly simple analysis in Section 1, I simply took the U.S. population data down here. I worked out the -- we had data for the number of observed and invasive cases from FoodNet, etcetera. I put that through.
This is uncertainty for about a Poisson intensity. And we simply extrapolated that out to a population. And then we split it between those that would have enteric and non-bloody, and enteric-bloody infections.
In Section 2, we were looking at -- all right, the only people that you observe are those -- who were culture confirmed cases. So we missed a lot. We missed lots of people. If I go from the bottom, we missed those people, for example -- let me see, which way should I go -- well, we've got the number of people who sought care. We have the number of people who submitted a specimen. We have the number of people for whom the specimen then tested positive.
And so only those people who went through all of those chains actually ended up being observed in your FoodNet data. So we need to extrapolate back and divide by all of those proportions, all of those probabilities if you like, to work out the total number of people who truly were -- who had Campylobacter.
And if we do that, I have -- this is a distribution where on the vertical axis I have a description of relative uncertainty, so -- confidence if you like. And you see the value range. In this case, we've got values that range from, say, about 0.9 million up to about, say, 4.8 million.
If you look at this on the cumulative frequency curve where this vertical axis here means the probability or my confidence that the true value is less than or equal to whatever the X axis value is. So, for example, I can read off here that I am five percent sure that the value is at least 1.3 million or something like that. And over here I am 95 percent sure that the value is less than, what would that be, about 3.8 million or something like that.
And over here on the bold line, I have the CDC estimate of the actual number that were observed in 1998 and -- which it rather fortuitously I suppose turns out to be at around about the 50 percent mark. So CDC and our data agree which isn't surprisingly because we used their data.
Now, I would like you to bear in mind that you shouldn't see this as the actual number, the distribution as the actual number of people. I know this is a difficult concept to get. But it is not distribution of the actual number of people uncertain about that.
It is the distribution of the intensity which has more uncertainty because we are taking into account the fact that we have a small sample from what really might have been out there. If we repeated that year, we could have seen different values occur from one year to another.
Okay. So in Section 3, we are interested in those people who had those Campylobacteriosis cases who would not have benefitted from -- would have sought care and who would have received through it fluoroquinolone, but then obviously it didn't work. So we have to go -- we have to back through here. We take the number of people and then we work out those -- the proportion that relates to domestically consumed chicken because, of course, the fluoroquinolone we are interested. The administration is to domestic chicken.
And then we look down here at those who went off and sought medical care, those who were treated with some medication, the proportion of those who sought care and were treated with medication for which that medication was actually fluoroquinolone.
And then by calculating by taking the total number of Campylobacteriosis cases and dividing by all of these, we multiplied by all these probabilities or proportions. We ended up with estimates of the total number of people who would have had invasive infections and were treated, but unfortunately treatment didn't help them because fluoroquinolone was of no benefit and those who had enteric bloody and enteric non-bloody infections.
And I have distributions here describing our uncertainty about what those values are. Again, these are Poisson means, intensity and uncertainties. And you see here we have got the confidence that the true number of invasive cases. Well, in 1998, it would be somewhere between, say, ten and 30. And there is the distribution.
It shows -- the back square there shows your mean. So the mean of that distributions means if you are going to pick one value that you are going to tell the press, that would be your best sort of guess if you like. And you can see the uncertainties.
Here we say somewhere between, say, 11 and 29 people with --- percent confident. It was within that range. And then over here I have got bloody diarrhea. And we have got distribution of uncertainty, somewhere again between, say, 700 and a bit less than 2,500.
And finally, I have got non-bloody diarrhea -- bloody diarrhea in the first one and non-bloody diarrhea enteric illness. And we have got somewhere between, say, 2,000 and 6,500 people.
And if you add those all together, the total number of people with invasive, bloody and non-bloody enteric infections, then we get a total somewhere between, say, 2,000 and 8,000 people a year in 1998 who would have been to the doctor, prescribed fluoroquinolone, but to whom it was of no benefit.
And I suppose you should compare that with, say, the two and bit million of people who have Campylobacteriosis. And so we've got 4,000 out of two million. That is a cumulative distribution, again, saying that it is somewhere between, say, two and a bit thousand and a bit more than 8,000.
So in Section 4, I was interested in looking at the contaminated chicken because I want to compare humans and the contaminated chicken populations. And this is a very simple analysis. I simply looked at the prevalence of Campylobacter in chicken carcasses at the end of the sorting process. And that is a point estimate -- sorry, that is a point in the process in which we are measuring.
If we had measured at the beginning of the process, we would have a different estimate of prevalence. So if you measured them at the slaughterhouses that they came in the slaughterhouse, you would have a different measure.
And for the purposes of this risk assessment, it is not really so relevant where we measure except it would be nice to measure as close as we can toward the consumer. So the first, so long as we can go towards the consumer. And this happens to be a good place because at the end of the chiller, they are then going to go off into a whole bunch of different paths that we can't monitor so easily.
So I took the prevalence of Campylobacter in chicken carcasses which is based on -- well, we have data on that and, again, the prevalence of fluoroquinolone-resistant Campylobacter among Campylobacter isolates. And so if you multiplied those two together, you get a good estimate of the prevalence of fluoroquinolone-resistant Campylobacter carcasses.
And from data, we have data on the consumption of the boneless, domestically-reared chickens in the U.S. in pounds. And so the volume of chicken consumed is the average per person multiplied by the population. And then we look at the total quantity of boneless, domestically-reared chicken contaminated with fluoroquinolone-resistant Campylobacter in the U.S. And that is just simply the total volume consumed multiplied by that Campylobacter-resistant prevalence.
And this is the estimate we came up with. It says that there is somewhere between 1 X 109. That would be 1,000 million pounds and, say, 2 X 109, 2,000 million pounds worth of Campylobacter-resistant fluoroquinolone -- fluoroquinolone-resistant Campylobacter contaminated chicken pounds.
Okay. Section 5 is trying to make a connection between the contaminated chicken that is consumed and the human health impact. We take this expected incidence which I have called in my model N3T. It is the total number of people would have had some human health impact out of the resistance from Campylobacter.
And we say that it is proportion to the poultry product -- poultry production Vi. And so there I have this constant of proportionality, K. And because N3T and Vi are very uncertain, we will have a lot of uncertainty about K. It turns out that this works quite nicely under certain fairly minimal conditions because of something called a conditional probability identity.
Okay. Now, how can we use this value of K, if you like, to make predictions about the future? Well, what we do is we say imagine Vn is a future annual volume of fluoroquinolone-resistant Campylobacter contaminated chicken that has been consumed. And we can work out what count that would be by monitoring the amount of chicken that is consumed and monitoring the prevalence of Campylobacter amongst chicken isolates and by monitoring the prevalence of fluoroquinolone resistance amongst those Campylobacter isolates.
So if we can keep monitoring this and have a good idea of maybe those trends, we don't even need to know very well what those trends will be. If we monitored them fairly consistently, we don't have to model the trends. We can just simply see where we are at any one point.
And we can use this very simple equation here which would tell you the number of new human infections. And that is going to be a Poisson distribution where the N here is this new amount of contaminated chicken, and divide it by K.
So at any stage, we can start to talk about the risk that actually is out there by having this prevalence of fluoroquinolone-resistant Campylobacter.
Now, this model does assume that the value of K remains constant. In other words, that human behavior remains constant. But I would say particularly with respect to things like behavior in the kitchen.
Now, we had a previous speaker talking about they didn't think that most of the contamination, most of the illness came from directly consuming poorly cooked chicken, but from poor handling practices. Now, we also had Doug Powell stand up and say you can't educate people. And I suspect it is going to take quite some time before you really will start people to handling the food a bit more properly.
I had fun yesterday coming back on the plane. We were -- Louise and I -- she is from England, as well. We were sitting on the bus. And the bus is taking us out from the airport to our car. And it is say, "Don't forget to put your seatbelt on." So at least you, too, try and teach your people. We don't do that at all. We think it is funny.
You remember how English have quirky sense of humors. That is us. So it tells us human behavior remains constant. It also assumes that the resistant pathogen retains the same level of pathogenicity. And it also perhaps more so -- a more difficult assessment is that it assumes that the microbial load in a contaminated portion remains constant.
Now, if, for example, you were to introduce irradiation as a process, then that would -- this assumption would fall down. Mind you, at the same time, you probably wouldn't have the risk anymore. So that wouldn't be such a bad thing.
Okay. Now, if we quantify -- how do you quantify the human health risk per year? And this is really a large part of why we are all here. It is a policies decision. But in order to present the results of my risk assessment, I have presented four different things here.
I have talked about -- if you remember those -- the total number of people who were actually affected because they had -- they consumed that domestically-reared chicken, they went to the doctor because they got Campylobacteriosis. The doctor said, "Here, have some fluoroquinolone. You will be fine." And they weren't.
Well -- and there is a bit of argument about what that would represent, perhaps an extra two days of illness, who knows. Anyway, what risk does that represent? It depends who you are. If you are just your average person in the U.S. population, then we can say the risk is if you like the actual number, the average number of people who would be -- in a year who would be affected in that way divided by the total population.
So the denominator is the U.S. population here. And for those people, for the likes of you and I who hopefully are not sitting here with Campylobacteriosis thinking about going to the doctor this afternoon, well, then the probability is maybe one in 61,000 or so. That is an expected value. There is uncertainty around that.
Or if you want to look in terms of probabilities, it is 0.0019 percent. And for most of you, you are not going to say, oh, 0.0019 percent. It doesn't mean a lot. But maybe one in 60,000 means something more to you.
Now, if you were sitting here with Campylobacteriosis, then the risk to you is something more like one in 521. On the other hand, if you had definitely decided that you were going to see the doctor this afternoon and you had Campylobacteriosis from the domestically consumed chicken, then it is going to be something like one in 63 versus if you actually went there and the doctor said, "Yes, you are ill", and he decided to administer -- or prescribe an antibiotic. That risk increases to one in 32.
Okay. So I have got a number of uncertainty distributions about that. Here we have the one in X kind of format where we have the U.S. citizen. I just want to show you what I mean by there is still some uncertainty about it. So we have a considerate amount of uncertainty around those values I am giving you.
Okay. Now, we need to analyze the uncertainty. We can use spider plots which are a nice little technique to determine where those key uncertainties are. If we know where they are, we know where we can take some more information.
And if we look at this analysis I will show you in a second, it shows that we -- in my view, we still have comparatively little knowledge of human health cases which is a very strong argument for increasing your FoodNet survey. I think -- well, if it were not for this FoodNet survey data, we would never have gotten started. And if it had wider coverage, we would certainly have a much better estimate of the human health impact.
Okay. So, well, what on earth is this? This is a spider plot. And here I have got all of the key uncertain parameters associated with estimating the total number of Campylobacteriosis cases in the United States in 1998.
And the vertical axis here represents if we were to know that each one of those parameters was at its actual five percentile or its 20 percentile or its 50 percentile, these are places along the distribution of the uncertainty. We have about what that true value is.
If we would be able to say, now, we know what that value is, if it turns out that it was at its fifth percentile, then our estimate, the mean estimate of N3T here would be at that value. So if I take this little black dot one and it was at ninety-fifth percentile, well, then it would be this value.
In other words, this vertical range here represents in sensible terms, terms we can understand, the effect of actually really knowing what that value is. For some of those where they -- the flatter, well, really knowing the value doesn't make any value to our analysis. In other words, our analysis is relatively insensitive to what that value might be or, in other words, what it really means is that we have sufficient data about those particular components and we should be concentrating our efforts in understanding other parts.
Well, the three bits from the point of view of estimating this total number of Campylobacteriosis cases, the three parts are the expected observed enteric infections in FoodNet data. More FoodNet data would be marvelous. Also, in here, the second most important was the probability that a specimen, a stool specimen tests positive. And there may be some amount of controversy about that.
We certainly had to use -- the one point where we used data that didn't direct apply to the U.S. population. It came from New Zealand data. But our choice was either to assume it was 100 percent or to use some data. And it was the only data that I know of that was available to us. So as people have said before, if any of you out there have information for us, it would certainly help us improve our estimates.
And here we have the probability that the stool requested and submitted for non-bloody. So what is the probability that if a person goes to the doctor that the doctor will say, "Oh, you better give me a stool specimen." And we have a lot of uncertainty about that.
In terms of the volume of contaminated chicken, well, we -- essentially it is the fluoroquinolone-resistant prevalence in poultry which is no surprise that we really have our most uncertainty about. But the really interesting thing is then to compare the ratio of N3T to -- divided by Vi. And that gives us -- from the point of view of the whole assessment, it tells us where we really need to concentrate on uncertainties.
And you see here, the PRC, which is that variable or parameter that is marked, is the only poultry-related parameter. So, essentially, it is the human health side that is really contributing the greatest amount to our inability to predict what the future holds.
So if I look at the total amount of uncertainty, you can see where this is -- the quotient of N3T/Vi is what I am interested in. And you can see that if I wrap up all the uncertainties of one versus the other, this is the human health. And human health has a great deal more uncertainty, in other words, has a much larger vertical axis range than the chickens.
So in conclusion, because I have got my stop light here, in conclusion, the modeling approach is simple. And simple will annoy some people. But it will also make an awful lot of other people very relieved.
It is simple -- I would like to think it is very transparent. And it makes few assumptions. I hope that we have done a good job of being quite explicit about what those assumptions are.
It is fairly easily updatable. And, therefore, if you choose to use it, it can remain very current. And a very key part of this is it recognizes uncertainty. And, of course, as I have said a couple of times already, our uncertainty would improve a great deal if we were to be able to collect more data.
And I believe that the model can be used as an aid to regulatory decision-making. And you will notice that I have written down "aid to." And as we have had our speakers say before, it isn't -- numbers are not the only thing. You have to look at a lot of other input parameters into a decision. So I in no way believe that this is the conclusion to your decision-making. Thank you very much.
DR. BEAULIEU: Questions?
DR. KASOFF: Mark Kasoff from London. I found it a very interesting talk. I have trouble with one step which is where the patient was acquiring the organism, has symptoms and has required a resistant organism is given the drug because we know that the great majority of these patients don't need any antibiotics.
How did you estimate the extra morbidity because he is taking the drug against the resistant organism? What estimate did you put in for that? Because in the end, maybe --- for the overall damage to the society of this resistant organism.
DR. VOSE: You have a good question. And certainly, we did look at the effects of the extra morbidity. We have very varying data, very widely varying data and not a great deal of consensus about I think what that true value is. But roughly speaking, it turns out to be I think about two extra days of illness for the vast majority.
I didn't include it because essentially it becomes a constant parameter. You multiply the number of people by the number of extra days of illness. And so I have left it at just the number of people.
But if you wish to convert that for yourselves into the human health impact in terms of morbidity, multiply it by two and call that days -- personal days of illness. And I think you have got a reasonable estimate. I wouldn't hang my hat on it, but it would be reasonable.
DR. BEAULIEU: Yes?
(Away from microphone.)
MR. : The model --- statistical uncertainty ---.
DR. VOSE: I agree with you. And the mathematics, of course, can only describe the statistical uncertainty. But Kathy Hollinger and Mary Bartholomew -- I hope I don't put you on the hot spot for saying this -- but they are going to describe the biological assumptions and uncertainties in their presentation this afternoon. So perhaps better to address that question to them.
(Away from microphone.)
MR. : Mr. Vose, my --- 61,000 ---.
DR. VOSE: Well, the bottom line, let me just drive up here. You can say -- I think the bottom line depends on what you want to say, you know. I mean, I am sure that if you are -- well, you can be on different sides of a particular fence here.
So I am not going to give you a bottom line figure. It really -- I think what I tried to do is by very explicitly talking about uncertainty, I let you decide what you mean by a level of risk.
And I think it is quite -- I think that is very appropriate because if you are on the side of human health, then obviously you would like to try -- you would see any human health impact as being awful for you and you would take it -- one would say -- some would say an alarmist's view.
But you would take a conservative view about that assessment versus if you were some other person, you might take a completely different view. So you choose what value you want to make out of those distributions. I am not going to give you that. That is not a cop-out, I promise.
MR. : I think one thing that is useful to consider is that although mention of the model, but wasn't shown in your presentation, is to try to understand the population of people from whom these people with potential harm are arising.
And if the model predicts two million people with Campylobacter infections and the data demonstrates that seven percent of those people have a fluoroquinolone-resistant infection and the resistance is a consequence of fluoroquinolone use in poultry, there are about 140,000 people a year who have a fluoroquinolone-resistant infection. And the resistance is a consequence of using fluoroquinolone in poultry.
That is the population from whom we tried to decide what the harm might be. And the model shows that there are about 5,000 people from that 140,000 that are affected that you modeled.
Those are the people who are sick enough to seek care and the physicians concerned enough to get a culture and also concerned enough to prescribe fluoroquinolone. And so we would say that those are -- that it isn't appropriate to look at those numbers.
But also you should consider that amongst those 5,000 people, there are about 20 the model predicts that have a bloodstream infection, an invasive infection. And it would be a simplification to judge that those people with a bloodstream infection would just suffer two additional days of illness that might be more severe consequences for them. And we would judge them to be severely harmed by this.
We do agree that the moderately harmed people do equal about whatever the 5,000 minus 20 is. Those are the moderately harmed people, people with two additional days of diarrhea.
There are also people that are mildly harmed that are not in the model. And those people that are mildly harmed are people that are ill enough to seek care, are -- but the physician does not prescribe antibiotics.
And there is increasing data or at least we have preliminary data that demonstrates that people with a fluoroquinolone-resistant infection, even if they don't -- aren't given antibiotics, have a longer duration of illness. So that needs to be more fully explored. But there is potential harm to people, even to those who aren't prescribed antibiotics. Then we have -- I mentioned there were 140,000 people in the model.
DR. VOSE: At least, yes.
MR. : And we just described the 5,000 people that were severely and moderately affected and the 10,000 people that are mildly affected, if you follow my logic. That leaves 125,000 people who are ill, but do not seek care and do not get cultured. And because they don't get cultured, we can't break them into groups of who is resistant and who is susceptible.
So it is very difficult to study those people to see if there is a difference in illness between those two groups. So there is this very large uncertainty, a group of people that the harm is uncertain but is theoretically possible.
And so I just wanted to point out the misstatement about the two days of diarrhea is not the only harm that your model portrays.
DR. VOSE: Okay. If I can reply to that, I entirely agree with you, Fred. And it was an approximation to say two days. But the reason I did it is because if we look at, say, the bloodstream infections. Perhaps there is an extra eight days. We had no data at all about the extra illness there would be -- that that would equate to.
But I took the approach that we are talking about 5,000 people versus 30 and the 5,000 times two days versus 30 times eight. It was a second order effect to include the extra days of human health effect.
But I agree with you. I thought it was better really to present the three sub-population as they were rather than aggregate them from the point of view of human health days because it felt to me that we were making more of an assumption that we haven't really needed to do.
Now, I also -- I think you have a point about the denominator that you want to talk about. But although that is one point of view, we spent a great deal of view discussing what that denominator should be in terms of estimating the risk to human population.
You say it is the number of people who actually have a Campylobacter infection that is fluoroquinolone-resistant. Now, perhaps that is the right one. But also, if the person never seeks care, does it really make any difference whether it was one strain of Campylobacter or another? If there was no difference between the human health impact on them, I would maybe argue that it wasn't relevant. But I am afraid -- yes.
MR. : But the point is just because they don't seek care and we don't have data where there is a difference in severity of illness does not equate to no difference in severity of illness. That is an assumption that should be explicit that, in fact, there is biological reason to believe that there would be a difference in severity of illness.
DR. VOSE: Yes, okay. And there was some data that -- we certainly explored that, whether there was any difference between the level of illness that somebody had if they had fluoroquinolone-resistant Campylobacter or non-resistant. I don't know whether Kathy is going to talk about that. But we took a judgement call. And I hope we are explicit about it in our report.
We assumed that there was not. But if there is data that says that there is, then obviously we would have to address it. I totally agree with you.
MR. : Yes, I would like to come back to a comment that Tom Shryock made a while ago regarding the numerator and not the denominator. The FoodNet data shows a three-time increase case load of Campylobacter and even larger than that with Salmonella in infants less than a year of age.
Now, we recognize that infants don't eat chicken, raw chicken particularly. And it is probably unlikely many of those cases even arose from contaminated chicken juices, although some could have. That clearly is a possibility.
Not knowing where those cases came from, it seems to me that when you get on down in your calculations in estimating the number of cases that come from chicken consumption, we must adjust for that because clearly those numbers cannot be related to chicken consumers. If you look at the data, the case load is something like 55 per 100,000 in infants less than a year of age and it drops dramatically to 20 after a year of age.
So there seems to be adjustment that would be needed in the data, in the modeling when you go from the estimated cases to those that are related to chicken which seems to me would reduce the numerator quite a bit when you get down to those cases of chicken. So I would suggest that that data needs to be re-examined.
DR. VOSE: Thank you. Well, I have to say that a very big difficulty we had in doing this risk assessment is to determine the proportion of people for whom the Campylobacteriosis really originated from chicken. And that is an incredibly difficult assessment to make.
And I think that some of that falls into what you are talking about because we don't have -- we can't -- if somebody comes to the doctor and they say, well -- you work out they've got Campylobacteriosis, how to work out where they got it from.
By the time that they've got it and it has been three or four days, and goodness knows whether you eat chicken and beef and play with cat and dog and any number of things. So if there is information in there that would let us be more specific about who really are getting it from chicken, that would be great.
And just to remind you that one of the previous speakers was talking about that they didn't think it directly comes from chicken, but from handling of chicken was in a large number of cases. Now, isn't that difficult to work out, how many people really got it from chicken when it comes from handling. Certainly not by looking at the amount of foods that they cooked.
MR. : I spent a number of years in pediatric practice and I actually don't have any problem understanding how infants under 12 months of age would acquire food-borne infections. They spend a lot of time in the kitchen along with mom while she is preparing the food. They receive solid food much earlier than us pediatricians recommend.
We, you know, tend to recommend formula only until six months. And that is the exception in my experience in pediatric practice; that they very often at their six-week check-up, you find that mothers are giving them solid food because they think it helps them sleep better through the night or something like that.
But the point I am making is that, you know, infants -- the definition of infant is less than 12 months. By the time they are 12 months old, they are pretty mobile. They spend a lot of time in the kitchen. I just don't have any problem understanding it.
(Away from microphone.)
MS. : --- breast milk --- the mother is constantly handling the baby ---.
DR. BEAULIEU: We will have to limit our questions to those folks who are at the mike now in the interest of time.
MR. : Two sort of technical comments about that problem. The first is that even though the incidence may be higher in people under a year of age, there are a lot fewer people under a year of age than there are in all the other age groups. So a high incidence doesn't mean a high fraction of cases.
The second is that there are three studies in the report from which the proportion due to chicken consumption were estimated. One of those uses students. So that is not a representative sample. The other two are population-based and, therefore, should take into account that age distribution.
DR. VOSE: Thank you.
MR. : I have a question. It goes to the front end of the cycle, not the back end of the cycle. Apparently, there is a 17 or 18 percent incidence of Campylobacter-resistant fluoroquinolone -- fluoroquinolone-resistant Campylobacter on poultry. And I have read through the document. And I wanted to listen here to see what you said.
And you touched very briefly at the start of your presentation about one shed of chickens receiving the fluoroquinolone and all the chickens in that shed receiving fluoroquinolone. And I don't question that, although in this country we call them houses. But that part doesn't make any difference.
But my question is the -- and there has been studies done on this that show that approximately one percent of the chickens grown in the United States are treated with fluoroquinolone. That is pretty low. Let's just say that it is even two, three, four, five percent. I am curious, in the development of this model, how have you accounted for that?
Because, you know, I don't question that the usage of the drug will lead to some resistance. I don't question that part. But I also question that there is other mechanisms for development of resistance. And I didn't hear anything in here that accounts for the very low usage of fluoroquinolones in chickens.
DR. VOSE: Okay. You have a very interesting point. And, okay, you have a different way of housing your chickens than -- rearing your chickens than I am used to. You have deep litter processes here. So there could be a connection from one --
MR. : What do you mean by deep litter?
DR. VOSE: Deep litter, isn't that what -- where you re-use the litter?
MR. : There are areas of the country that re-use litter. There are areas of the country that do not.
DR. VOSE: Do not, okay. Absolutely. All right. So there is a potential mechanism even though you may -- you treat a chicken from the past, another chicken can get it at a later time though it has never directly received fluoroquinolone itself. But -- and there are a number of -- there are all sorts of different things that can happen.
For example, at the plant, there could be cross-contamination galore at the plant, particularly in the chiller and in -- I know the thing that takes off the feathers, the machine that eviscerates the poor thing. And, you know, I have been there. I have a diary where I was taking notes. And I have this page -- this double page splattered with blood. I will never forget that.
But they have this thing that goes whoosh and removes the whole of the inside, you know, the poultry carcass in one hit. It is quite an impressive piece. But it goes round and eight times later, it is doing the next one. There are all sorts of different things, although I know that certainly the slaughterhouse that I went to was an incredibly clean place.
And certainly, the industry -- and I have to say that this is not in the United States, though I am sure that you have some of the practices -- but the industry took enormous pains to try to reduce the amount of cross-contamination.
I don't pretend to know exactly where that -- what levels of mechanisms of cross-contamination occur. And I think that that is one of the real difficulties of doing the farm-to-fork risk assessment, is being able to quantify all those levels.
So what I have tried to do is say, look, I don't know how they got all fluoroquinolone-resistant Campylobacter on that chicken. I admit that I presume that the fluoroquinolone resistance comes from administering at some point to some chicken fluoroquinolone. And we can certainly argue about that.
But having made those assumptions, at this point here, out of the chiller comes this chicken. It is contaminated or it is not. And I don't know how it really got there. But that is the thing that is going out to the consumer. Okay.
MR. : I really don't question that part of it. But what I am saying is I think somewhere in here, you should try to separate the resistance from the usage of the drug. I am not saying drug usage leads to resistance. But there are other things that lead to it, also. And let me just raise a point -- and I know people want to go to lunch.
But there is a point, we had checked broiler houses where we have moved chickens and just gone in on built-up litter and have done what we have done a wash-down which is cleaning and sanitation. And we could find Campylobacter in that house before we went through this process. But we could not find it after we went through that process.
And I am not saying we get them 100 percent all the time. All I wanted to get you to do was to think about the fact that there could be something else involved in this whole mechanism besides using fluoroquinolone. That's all.
DR. VOSE: Okay. Thank you. If I can just make a final comment to you. It would seem to me very worthwhile if the industry, the poultry industry was to -- and it sounds like you are doing it right now -- but was to try to do a risk assessment to identify where that contamination comes from. Now -- and that's -- I don't say that that is not a big job. I think it would be a big job.
MR. : No, no. It is.
DR. VOSE: But you would have some clue as to where it came from and the ways that we can change the use of fluoroquinolone -- or one can change the use of fluoroquinolone to minimize the resistance in poultry at the end of the process.
MR. : Those, in fact, are things that we have been doing. We don't have them all because there is a lot of these questions we don't know the answers to either. And one of the things is the chicken industry that we are working on is guideline manuals for use of products that we use to minimize the kinds of problems that you are talking about.
But you are right. I think you see it the way I do. This thing is more complex than what it appears.
DR. VOSE: Thank you very much.
MR. : Thank you.
MR. : I am from the Canadian Food Inspection Agency. I would like to weigh-in on both sides of the developing camps here. On the one, we have done a quantitative risk assessment for Campylobacter jejuni in fresh poultry in collaboration with Norm Stern at ARS, Russell Research Center in Athens, Georgia.
In our model, we actually did take a stab at modeling the cross-contamination impact in the kitchen as well as the preparation and consumption of cooked poultry. And in the manner that we modeled it, we came up with final estimation of risk approximately 200 times the risk of the cross-contamination in dripped fluids on counters, etcetera, being approximately 200 times that at consuming prepared and cooked chicken.
Now, there is a huge amount of uncertainty in that. There is a great need for further investigation and further work. So I don't -- but I don't -- from that, I don't have any difficulty in sort of buying into that theory or hypothesis about cross-contamination having a very important role.
On the other side, like the gentleman before me, you said initially, David, that one unique thing about getting into developing this model was that you had data for all the points along the way. And that is always an important concern in developing process risk models, quantitative risk assessments.
But I am not hearing that you really have data on that front-end association saying that the -- that very large assumption saying that the fluoroquinolone resistance in that ---
(Audio missing due to technical malfunction.)
DR. VOSE: --- it sounds logical if you take your chicken and it lives its life in the shed or house, whatever you call it. And then it goes from there to the poultry slaughter plant and it hasn't really been anywhere else. Then I guess to me it strikes me as a reasonable assumption. But certainly if there was any data that would say otherwise, then, of course, we would be delighted to look at it.
MR. : I don't disagree that it seems somewhat reasonable. It's just that the question is we don't really have the data.
DR. VOSE: No, there is no causal link.
MR. : Yes.
DR. VOSE: Yes. And I don't think this risk assessment has ever set out to prove causal links. If you have a criticism in that regard, I think it is reasonably unfounded. In any microbial risk assessment, there is never an attempt to make the causal link, just to look at the quantitative implications under a certain set of assumptions.
So we -- in microbial risk assessment, we make assumptions. And scientists find the causal links to either back us up or tell us we are wrong.
DR. BEAULIEU: Last comment.
MS. : I just wanted to say that a risk assessment question given to us by our risk managers really did not have that question or that issue within the scope of the question. We were to look at what is the impact of fluoroquinolone resistance from -- in humans from exposure to chicken.
So we really didn't address the drug use issue at all in this risk assessment. There was no attempt made. So I would like to say that, you know, that is part of the reason why this really isn't in this risk assessment. It would probably more be part of, you know, the risk management decision to use this risk assessment question that we needed to address.
(Away from microphone.)
MR. : It just seems that critical ---.
MS. : Yes, I think that we have seen that -- you know, evidence coming out of other countries or we have seen clinical trials. And we have seen resistance develop in relation to use in both humans and animals and in laboratories when you use it in bench top tests to create nalidixic acid resistance, for example, in microbes to mark them for further studies. So you see it, you know, developing fairly readily in response to exposure to the drug. And it is a characteristic of that class of drugs.
DR. BEAULIEU: Thanks. Thank you.
DR. VOSE: Thank you.
DR. BEAULIEU: We are running significantly behind schedule. You might have noticed. I am going to try to get at least -- let Tony Cox speak this morning. We may have to -- I will talk to Steve and see what he wants to do about his presentation.
Our next speaker at any rate is Tony Cox. He is President of Cox Associations, an independent Denver-based applied research training and consulting company that specializes in health safety and environmental risk analysis. He holds advance degrees in risk analysis and operations research from MIT. And he has lectured widely on topics in risk analysis, applied mathematics and computer science. Dr. Cox.