 My name is Hamilton Staples, and I have the great pleasure of introducing our first speaker today here in Bancroft Hall, Richard Feynman. Dr. Feynman is a professor of cell biology at SUNY Downstate Medical Center. The original title of his talk was fructose and fructophobia, the threat to paleo and the opportunity. I hear that title has changed, and he'll reveal the new title during the talk. Please join me in welcoming Dr. Richard Feynman. Well, the the title has changed. The opportunity in fructose was that it was a lot of the discussion was based on basic metabolism. Which emphasizes the scientific background of the problem. The threat, however, is that partly we might not get the metabolism right and that we're a little bit out of control in condemning fructose and the former mayor of New York thought we should even tax it, which would certainly have raised money, but not clear that it would do anything else. So exaggeration is not good, but in taking a paleolithic viewpoint, there are probably bigger threats. In particular, there's a certain amount of pressure to turn the country into what I call carrot nation. If you don't know the reference, I'll carry nation was one of the famous prohibitionists, and she went around smashing whiskey bottles with her axe. But there are actually many threats, a threat of red meat and even white rice. So as I looked into the threats, there was more and more material. I found myself like the apocryphal story of a graduate student, as I heard it in Berkeley, who was writing a thesis whose title was the character of Shylock and the Merchant of Venice. After a year or so, he had so much material that he changed the title to the Jew in Shakespeare, and ultimately the Jew in Elizabethan literature, and after some time the title became the Jew, and he's supposedly still working on it. So there really are a lot of threats here, and what I'd like to do is to describe some of the problems in the medical literature and the new title is Hill's Criteria, Bay's Theorem, and Murphy's Law, and Feynman's Matrix. So I'm going to discuss some of the original points on Fructose and give you a golden rule for statistics to apply to the medical literature, and I'm going to describe Bradford Hill as the person who first made the association between cigarette smoke and cancer, and his explanation on when an association does imply causality, and Bayes will describe the philosophy of statistics, and we all know Murphy's Law and Nutrition can go wrong, they can prove it right with statistics anyway, and in order not to be overly negative, I'll try to give a possible take on what we can do that might improve communication in this area. Now in Fructose, the key question as stated by Denise Minger's blog post was whether the fruits that we're eating now are like anything we ate before, and the idea is that we have not evolved to deal with high intake of sugar in the environment, and she showed that there really are a lot of sweet fruits that have not been through genetic engineering or Del Monte processing, and that they are sweet and have a lot of carbohydrate. But the real question is undoubtedly true that the availability of carbohydrates specifically Fructose was much lower for our ancestors, but the question is does that show up as an evolution of the metabolism for low intake, and I would suggest that in fact just the opposite that if there was a rare availability of Fructose, that finding a berry patch for our ancestors was equivalent to finding a coupon for Haagen-Dazs. Moderation was not what was going on. So looking at the metabolism, the basic idea that was a problem in terms of the science is the emphasis on, this is the basic glycolytic pathway by which Glucose is processed either for lactate or in the TCA cycle, and there's an emphasis on that being different than how Fructose is processed in the liver. And what's said is that Fructose by-passes PfK1. Fossil Fructocognase 1 is considered one of the major control points in glycolysis, so it responds to ATP and 10 other metabolites, and so regulates the metabolism of the flow of Glucose metabolites. The trouble with that is that that's not a really accurate way of looking at metabolism. The idea that this causes this and that causes that is not how metabolism is really set out. It's more what I call the pinball machine analogy and that's really not an accurate position description of metabolism, because in fact Fructose 1 phosphate activates Glucocinase. In other words, the appearance of Fructose calls for more Glucose, so what has evolved in the liver is a system for taking in both together. And Fructocognase requires ATP, and what that means is that if you lower ATP, you're going to turn on Fossil Fructocognase. So the basic idea is that whereas it's true that a high Fructose intake will lead to these intermediates, the so-called triosphosphates, which internally to fat and other downstream effects, the presence of Fructose also activates the conversion of Glucose to these intermediates. So the big conundrum if you want to ask questions about metabolism in Fructose is how do the triosphosphates know whether they came from Glucose or Fructose? Well they don't of course, and the further complication is because everything is connected, you also have substantial Gluconeogenesis. And in fact up to 60% of ingested Glucose will be converted to Glucose. I'm sorry, ingested Fructose will be converted to Glucose. So it's going to be very hard to sort this out, and you better be sure you have very accurate data. In addition, although under-emphasized, Fructose is a major source of glycogen. For a long time it was considered a glycogenic substrate. And finally, contrary to what you may see on the internet, ethanol is not processed in any way like Fructose. They're completely different in their metabolism. So where does this play out in terms of practical considerations? If you have a high Fructose input, you do see statistically higher triglycerides for example in the blood. But if you look at this, you can see that the differences are not great. And more important, there's a very large variability. It's not clear that anybody in here is really better than anybody in there because of the huge error bars. The real problem with it from a practical standpoint is that all these experiments are done at 55 percent total carbohydrate. And the question is really not, how can we maintain 55 percent carbohydrate? And if we do that, should we put Glucose in place of Fructose? I don't think that's the question. But more important, this kind of data is not really informative. And so I'm going to bring you the minus one principle. And that is that in nutrition, as in Lake Wobegon, nobody is average. And the new rule is habeas corpus d'aturum. We have to see all the data. We have to see what individual people... You don't want to know whether you're likely average. You want to know whether it's worth your while to bet on a particular diet. The zeroth principle, the major principle comes from a book called PDQ statistics. It's a very good statistics book. It's a sense of humor is somewhat beyond even PDQ Bach. So I'll just leave this up here if you can read it. But the statistical principle that it says is that the onus is on the author to convey an accurate impression of what the data look like using graphs or standard measures before beginning the statistical shenanigans. And that's going to be the main theme here, is that the medical literature is not doing well on that. I won't go through this. I have too much detail here. I misread the instructions. I thought I had 40 hours. The important point though is the idea of relative risk. And it's roughly speaking the same as odds ratio or hazard ratio. And the way it plays out practically is if the odds ratio is 50-50 or the hazard ratio is 50-50 and the hazard risk is one, then there's no difference between the two things that you're trying to compare. Now if the odds ratio is 2 to 1, then that's a suggestion that you may have something. That's probably what it would take to be considered seriously if you had a toxic tort case in a court of law. If you came in with less than 2 to 1, you'd probably have to have other evidence. I don't think it's cut in the right. So the main caveat is that relative or percent differences hides information. You have to know what the real risk is. As one example I give is Alice has 30% more money than Bob, but you don't know how much money she has. They may both be on welfare. Or I can give you a good way of doubling your odds of winning the lottery. You buy two tickets instead of one. Obviously that's not going to help a lot. Now the so-called father of modern epidemiology is Bradford Hill. And he, this is a remarkable document. It summarizes the question of winning the association could be considered causal. And it's beautifully written and very modest. I recommend it even though it's 50 years old. That's Bradford Hill. And as I said, what he asked is in what circumstances can we pass from the observation to a verdict of causation? And he produced nine rules. He was very modest and said that it was common sense. He didn't mean that they were absolute criteria. And the first of the rules, and this is what you should ask yourself when you look at any paper, are we talking about something big changes? Because his first study was to get medical records from physicians who because of their jobs were required to have medical records. And of the 789 deaths, Richard Dahl was his co-worker and 36 of the deaths were attributed to lung cancer. So when he went to look at the counted smokers versus non-smokers, the correlation virtually sprang out. All 36 of the deaths had occurred in smokers. So that's the kind of data that will tell you this is likely causation. This is described in the Emperor of Ormalities which is an excellent history of cancer. So let me tell you how you might apply Bradford Hill's first principle. This is a recent study on the Mediterranean diet. This is actually the most highly accessed paper last year in the New England Journal of Medicine. And their conclusion was that a Mediterranean diet supplemented with extra virgin olive oil or nuts reduced the incidence of major cardiovascular events. If you look at the data, however, the data is presented quite honestly here right up front. And you can see that there's a big difference between the controlled diet and the Mediterranean diet. But if you look in absolute terms, if you actually look at the scale here, this is pretty small. It's about 2%. And the authors rather honestly plotted this on a overall picture. The hazard ratio is .7. So the way to think of hazard ratio given this way is it's the inverse of the reciprocal of the risk. So .7 corresponds to or .5 corresponds to 2 to 1 odds. So they have not even hit 2 to 1 odds here. But in criticizing these things, one usually asks, maybe there's a reduction in cardiovascular risk, but what about overall mortality? And if you look at that, you can see that the Mediterranean diet, one of them was worse than just about the same as the controlled diet. In other words, there's virtually no difference here. And the analysis I just added my own graphic, you really have to look pretty hard to find anything here. So the point here is that Dolan Hill didn't have to do anything at all. The trial that was supposed to be the most rigorous statistical analysis, barely required elementary mathematics to prove its point. And that's I think a major point. If you want to get people to change their behavior, you have to get something real. Let me turn to another case, which is red meat, which is red meat and type 2 diabetes. And there's the claim that red meat consumption caused type 2 diabetes. And the most recent papers said that changes in red meat consumption. In other words, they had two different time points for red meat and the subsequent risk of type 2 diabetes. They came to the conclusion that increasing red meat over time is associated with an elevated subsequent risk of type 2 diabetes. And I'm going to try to analyze this particular study as a guide to what you have to deal with in the medical literature. When you look at this, what you're confronted with is rather than easy presentation of the data, you see absolute mind-numbing tables. So the question is, how do you get the information out of this? And what you want to know is, well, what is the risk? Because again, the first principle is, is this a big risk? And you can calculate that. The incidence, I mean that's the simplest thing to ask, what is the incidence of diabetes in this case if you're a high meat eater? And you just take the number of cases which you can find in table one, and I'm sorry, table two, and divide that by n which you can find in table one, and you come out with 7.4%, which is something. It's not great. But you have to compare this to those people who didn't change their red meat intake. And when you do that calculation, that's 6.8%. So the absolute difference is less than 1%. And when you consider that this is a change, what that means is that they measured red meat consumption at two different time points, and this kind of data is like weighing the captain by weighing the ship when he's on board and when he's not. You're really looking for trouble. Another way to describe the data is to take the reciprocal of the absolute difference which is called the number needed to treat. So what this says is that you would have to get 167 people to substantially reduce their intake of red meat in order to save one person from type 2 diabetes. Along these lines, though, is the Hill's next principle which was consistency. In other words, does the observation fit in with everything else that's been going on? Has it been observed by different persons in different places, circumstances, and times? Well, it's been certainly observed by the same group at Harvard over time. But overall, what most people observe is that during the period that we refer to as the epidemic of diabetes, diabetes went up substantially. The units are in millions of people, and this is in pounds of red meat. So there is a substantial lack of correlation here. And nonetheless, the authors say our results add further evidence that limiting red meat consumption over time confers benefits for type 2 diabetes prevention. No, it doesn't. If anything, it shows just the opposite. I tried to do this fairly quickly. I don't want to put a lot of text on the screen, but the reason this is uniquely offensive is that both the authors and the press suggested that the risk in red meat was comparable to the risk in cigarette smoke. My overall impression of epidemiology is the correlation between cigarette smoke and cancer is not only the classic example, but it may be the only example. So I wrote to the editor when this came out and I said that the association is very low, and the author reply was that no, it's not low, that the 1.8 hazard ratio that they found was the same as the ratio they found for smoking, which first makes you wonder why the paper wasn't called a risk from smoking. But 1.4 is not high. The definition of high is not what you think is going to be a cause. It's what the number says. And Bradford Hill found the death rate from cancer in cigarette smokers is 9 to 10 times the rate in non-smokers and the rate in heavy cigarette smokers is 20 to 30 times. In addition, the important point is that he said that if you look at the cancers that do associate with cigarette smoking, it's lung and throat. It's those things that you expect, it's not. And he gave the example of thrombosis where he said 2 to 1 was already suspicious. So I think this is very serious criticism and I'm not finished with this paper and that journal. And then there's temporality and this is just a question of which came first to the car of the horse and I won't discuss this much except you do see a lot of reports about how diet soda is associated with obesity. Who drinks diet soda? A really important question though is the biological gradient and what's meant here is that not only 20 times what non-smokers found, but the more there was a direct correlation between how much you smoked and the risk of cancer. So if we go back to the red meat paper, we calculated the incidence for high red meat, we calculate for small intake and you get 5.5%. But you should look at what happens if there's a decrease in red meat and what you find out is that it's the same as high red meat. So what I did is I went back and calculated the risk dose response curve and what you see is this. And what this is saying is that part of this curve is that when red meat goes down, diabetes goes up. I posted this on Facebook, somebody said what is the meaning of it going back up? And the meaning is that this data doesn't mean anything at all. The error and the poor methodology is such that it could come out right. Now people know that if you have a federal grant and you falsify your data, you can go to jail. But you can put out the data and misinterpret it any way you want. So how did they find that the actual risk corrected was 1.59? It's not that great, but it's something. And the answer is that they corrected for confounders. Now what this means is, for example, if you found that there was a direct correlation between, say, carbohydrate and obesity, you would have to recognize that people who took in a lot of carbohydrate might also be taking in a lot of calories. So you would have to correct that data for calories and if it held up, the association might be causal. If it didn't hold up, you'd have to throw it out. However, if you start from something that is wrong and you add confounders, well, they actually, the time-dependent, tax-proportional hazard regression model means they just looked at how the rate changed over time. And they had to correct for race and family history and smoking and red meat and initial changes and other lifestyle. They got it to come out right. Probably if it hadn't come out right, they'd have to put it in their shirt size and other variables. Well, overall I describe this with reference to Gulliver's Travels. He's a country where they refer to the thing which was not because their language doesn't have a word for lying. This is meaningless data. Another view of the biological gradient, this compared, again, changes in availability of sugar with the incidence of diabetes. Now what's good about this paper is that it actually put out all the data. It did not lump things together. It did not give you a mind-numbing table. And the thing about this is what I did is you have to recognize that here there was no change in diabetes and here there was no change in sugar availability. And what I did is just break this into quarters and colorize it. I haven't changed any of the data here. And what you can see is that it's as likely that, well, almost as likely that if you lower sugar you get an increase in diabetes, or if you increase sugar you get lower diabetes. So this data really, if anything, it says there's no association. But when you consider that they're measuring changes in sugar availability, the whole thing could be changed if a couple of guys in Hoboken hijacked a crate of sugar every couple of months. I won't go through these. I'll just post them. He'll have a total of nine criteria. And, well, as I say, I'll just post these. He was very careful about saying, about biological plausibility that that changes over time. Let me switch now instead to Bayes theorem. Now, have a, I'm going to skip most of the detail to give you the bottom line. Which is that Bayes theorem relates statistics to psychology. It tells you what you really mean when you say you have a 20% chance of rain. Now, most people interpret that as not meaning anything at all. It means you know what you knew just before you turned on the radio. But it recognizes that what you have in statistics is a belief state. And I'll skip through to the bottom line. And, well, Bayes put it in an equation. Bayes was actually an 18th century clergyman. This dates from 1743. And he, what the equation is, it's one of the order equations to understand. He says that you start with an a priori assumption about the likelihood that cardiovascular disease and saturated fat are associated. And then you correct that in some sense by the actual experiment that you do about the probability that there's a lot of cardiovascular disease and the probability that the people in your study are actually consuming saturated fat. And then you have an a posteriori conclusion. In other words, you have a new belief state based on the experiment. And in practical terms, you may not be able to put a number on the a priori belief. But that's the likelihood for the experiment. So, coming to the conclusion, whatever the belief that saturated fat was a causal agent, after the Framingham study, our belief had to go down. We were much more suspicious that that was true. Or we should have been. And the Oslo Hart study would further give us a belief about the hypothesis that was lower than when we started the experiment. And there's a whole bunch of these winding up with the Women's Health Initiative, after which we wouldn't believe it at all. So what's the take-home message here? If you apply this kind of thinking to a meta-analysis, this is where people try to average different experiments, you can see that the assumption is very bad. Several of these came out recently comparing what happens when you substitute saturated fat for other things. And the remarkable thing, the statistical rule is that if the error bar crosses the hazard ratio of one, then there's no difference. And so there are only two statistically significant studies here. So what should have been is rather than averaging these, each one of these should make the others less likely. And so what you're doing in a meta-analysis is you're trying to show that many wrongs add up to something real. And that doesn't make sense. Let me try to finish up by showing you what I think is a good way to present some of the data. And we published this a couple of years ago. It was based on the idea. We said, you know, if you're going to do statistics, the underlying assumption in statistics is that your independent variable, which is what people eat, for example, or, you know, this is relatively identical and that the error is due to some kind of random disturbance in the environment that one guy gets somewhat more mayonnaise than his tuna fish salad or something. But we know that that's really not true in a diet experiment and the twin study shows you that, in fact, if you look at energy efficiency, how many change in energy compared to the predicted, you can see the twins are very close together. But between pairs of twins, there's a lot of variability. So what we did is we tried to take account of this and give the, would be the most modest approach to the diet study. So what we did, this is, we took Volek's data from a low carb diet, very low carbohydrate ketogenic diet and a low fat diet. So the, we took the low carb diet, we took all of the weight loss and we put them in order and put them across the top. We took the low fat group and put those down the side. And then we took all the differences and the differences are the individual matrix elements. So what we're saying here is, we don't know whether two people are alike, but let's look at all the possibilities. What will be, what will happen if one person just happens to be like each of the others. And so this is the broadest picture of what could go wrong. And then we color code this and if it's evenly divided, then you know there's no real difference between the two diets. But in fact you can see that the low carb diet was drastically better. More important, I don't think you can see the color here, but the people who really did well was much greater in the low carb group. So we published it, we did it because it made sense and it was a good way to look at the data, but we kept saying to each other, we can't have invented this. Somebody must have done something like this before. We ran into the problem of communicating with statisticians, which turns out to be harder than I thought. And it took us a while before, they didn't know what we were talking about. We finally found what this really is like. And it's like the non-parametric method. A non-parametric test is what you do if you are not sure that your data is normally distributed, like a bell-shaped curve. And all you do is you, this compares the response time for alcoholics, for alcohol ingestion and placebo and they get all these different times. And all you do is you just rank them. You put them in a list and then you add up all the ranks. And the group that has the higher rank is, in this case, the slower one. So it's clear that alcohol has much lower reaction time. And so we realized that this is really equivalent to what we were doing. We made a matrix and we put one group on the top, the other group on the side, and we calculated all the differences. Then when you color code it, you can immediately see that one group is better than the other. So we're going to call this the graphical non-parametric matrix algorithm on the chance that it's so complicated that people start calling it the fine-man-fine method. And we'll see whether that happens. So these are the take-home messages. You've got to see all the data. Statistics always, group statistics always hides data. And it has to be a strong association. And the onus is on the author to convey the reader accurate impression of what the data looked like. Otherwise you've got to view it as an expression. Okay, that's it. We probably have time for one or two quick questions. Does anyone want to ask Dr. Feynman? Anybody in the back? I can walk to you. Hi, Dr. Feynman. So do you disagree with the fact then that fructose creates more fat in a person's body than glucose? Yeah, I disagree with that. As a general statement, and there are certainly conditions where it may, it does not actually contribute to fat. If you actually look at what happens, it can contribute the glycerol backbone. But it's just what I said. It can contribute more fat or it can contribute less fat. It depends on everything that's there. I mean the way I describe metabolism is it's like American football. You know, you can follow the quarterback, but the key to the play may be downfield blocking or what everybody's doing. You can't make a statement like that. Sometimes it increases fat, sometimes it decreases. But most of all, the main take-in point here is that the data supporting a unique effect of fructose comes from studies where you have a high background carbohydrate. And the question is not should we convert, should we change some fructose in our diet to glucose? The question is, if we do that, is that comparable to taking out carbohydrate across the board and replacing it with fat? Which is better? Replace carbohydrate with fat, keep carbohydrate high and replace fructose with glucose. Experiment hasn't been done. Why not? Well, I don't think they want to know the answer. Well, you know the answer. Well, he'll point it out. You never really know the answer till you do the experiment. But I think we know, yeah. That's all the time we have. Please join me in thanking Dr. Feynman. Our next talk will start at 11 o'clock.