The Postprandial Elephant

Hello readers. Before I get started I wanted to let you know that you can follow me on Twitter Twitter or Facebook. I'm not a heavy user of social networking sites, but I've been posting here and there with things that are too small to go on the blog. For example, I have recently been digging into the data from the National Health and Nutrition Examination Survey (NHANES), looking for associations related to salt and blood pressure (e.g. uric acid and kidney health, as I wrote about in the context of the salt/blood pressure experiment). Thanks to the folks at the Centers for Disease Control and Prevention for making all of this great data freely available to the public.

Today's topic is something that is seldom discussed when health and diet topics come up. I started working on this post about a year ago, but first wanted to work through some other ideas that I thought were important for understanding what is going on. It was hard to write and could be even harder to understand. This post covers topics from computational biology to lipopolysaccharide and the microbiome. Basically a little bit of everything (for my present definition of "everything"). The main topic is the usefulness of measuring one's triglycerides after meals. But first a little digression on measuring things, and on measuring blood sugar in particular.

Measuring Things

Once something can be measured it becomes a target for tracking and manipulation. Conversely, things that cannot be easily measured are often overlooked. As a result it is easy to place undue importance on measurable quantities at the expense of those that are more difficult to observe.

Gary Taubes mentions this idea in reference to cholesterol in Chapter 1 of Good Calories, Bad Calories. He writes: "what kept the cholesterol hypothesis particularly viable through the prewar years was that any physician could measure cholesterol levels in human subjects." Once discovered to be easily measurable, cholesterol became a target for management, and interventions that seemed to affect cholesterol levels were suddenly of great interest in the treatment of heart disease, regardless of any known or hypothesized relationship between those interventions and heart disease itself. Total cholesterol got managed because it was routinely measured. Later, other biomarkers were identified and studied in turn, such as HDL and fasting triglycerides. More recently, measurement of specific lipoproteins and particle density subfractions has come into vogue.

Blood Sugar

While fasting biomarkers are interesting, a lot of the action occurs in the postprandial ("after eating") state. Your doctor probably does not do any postprandial measurements unless you are pregnant or a suspected diabetic, in which case an oral glucose tolerance test may be indicated. In the low-carb and paleo worlds, the effect of a carbohydrate-containing meal on blood sugar is well known. Many folks on the curious/nerdy spectrum even own their own glucometers and test themselves from time to time after meals.

Knowing a bit about your postprandial blood sugar response seems like a good idea. There is plenty of evidence that high blood sugars are a problem (see review papers here and here). Jenny Ruhl over at Blood Sugar 101 has written extensively about this.

How does blood sugar go wrong? Tolstoy said that happy families are all alike, but that each unhappy family is unhappy in its own way. Perhaps the same may be said for carbohydrate metabolism. While young healthy people seem to be able to regulate blood sugar within a narrow range at all times, including after meals, people in various stages of metabolic derangement exhibit different patterns of abnormal blood sugars. Sometimes they spike too high but quickly return to normal. In others, they may rise gradually and hang out at excessive levels. Other people have so-called "reactive hypoglycemia", where a high-carbohydrate meal triggers an abnormally low blood sugar a few hours later.

As the patterns of metabolic disturbance differ, we should expect that there are many different underlying causes, including a variety of disease states, short term and chronic injuries, toxin exposures, and genetic polymorphisms. A metabolic maze of twisty passages, all different. Once started, a pattern of high blood sugar could be a problem in itself, with blood sugar reaching levels toxic to organs, nerves, and other tissues.

A model of disease mediated by glucose disregulation.

So I put postprandial blood sugar in the category of biomarkers that are easy to measure, and therefore more likely to be managed by many health conscious folks.

Many factors can affect postprandial blood sugar. These could include your glycogen status, fasted vs. fed state, proximity to exercise (e.g. post-exercise glucose uptake), rate of gastric emptying, the size and composition of the meal, carbohydrate content of the baseline diet, how fast you eat and how good your are at chewing, sleep deprivation, stress, and glycemic index.

Lipopolysaccharide exposure can also cause insulin resistance and raise postprandial blood sugar. The microbiome is involved in its own complex manner, as we saw with the TLR-2 knockout mice. Even ambient temperature may have an effect. Finally, the behavior of pancreatic beta cells is known to be quite complex and non-linear, so the blood sugar response to a particular meal may be difficult to predict.

The Elephant in the Room

Carbohydrate (primarily glucose) is not the only circulating energy substrate. It is one of four main categories, the others being fat (usually packaged in triglyceride form and carried by lipoproteins), amino acids (which can be glucogenic or ketogenic), and ketones. A few other molecules such as lactate and pyruvate can also serve as circulating energy substrates.

When you eat a lower carbohydrate diet, your fat intake necessarily goes up, and when you eat fat, that fat is absorbed and enters circulation in triglyceride form. So at some point I began to wonder whether eating a higher fat diet was spiking my triglycerides at the same time as it was normalizing my blood sugar. While it is well known that fasting triglycerides go down on a low carb/high fat diet (see e.g. Volek and Westman 2002), it was not clear to me what happens to triglycerides after a high-fat meal.

Some Fat Diabetic Rats

The figures below come from a study on diabetic rats (Motojima et al 2008). I like it because it shows the substitution of high postprandial blood sugar on a standard high carbohydrate rat diet for high postprandial triglycerides on a high fat diet. (Note also that NEFA (non-esterified fatty acids) in the lower right chart does basically the same thing that triglycerides do. We'll come back to those in a bit.) Also note that the peak in triglycerides comes a bit later than the peak in blood sugar. This is likely due to the fact that fatty acids absorb more slowly than carbohydrates. That said, the pattern is pretty much the same between fats and carbs: after a meal, circulating energy substrates peak and then go back down.

Postprandial values for standard (solid circles) and high fat (open squares) diets in Goto-Kakizaki diabetic obese rats. They had to train the rats to eat their chow during a one-hour window, twice a day, because rats would ordinarily not do that. From Motojima et al 2008.

One last item to note in these charts is what happens on the right hand side, long after the meal has been digested. Note that the high fat diet causes a long term elevation of not only triglycerides and NEFA, but also of insulin, compared to the standard chow. In fact, it even looks like the triglycerides creep higher from the 12:00 to the 15:00 readings. For now, let's arbitrarily label this extended post-meal phenomenon "inflammation."

We know by the way that the high fat diets used in these sorts of studies cause metabolic problems in rats, which is why they are used as a model of obesity. We should not therefore assume, as many wrongly do, that high fat diets are also unhealthy for humans. However, we can look at this sort of animal research as a guide to what happens when a diet does induce metabolic problems in humans (elevated fasting triglycerides would be an example of that kind of diet).

So while it is normal for circulating energy in the form of blood sugar and triglycerides to go up after meals, if it goes up too much or for too long you end up in a sort of metabolic gasoline fight. Like elevated blood sugar, elevated postprandial triglycerides is associated with heart disease and stroke and a host of other bad consequences. See, e.g. a recent review ("Triglycerides and Heart Disease, Still a Hypothesis?", Goldberg et. al. 2011). A scientific statement by the American Heart Association cites a pile of research related to this topic as well.

Diabetes is a metabolic disregulation that causes excess postprandial blood sugar. It also causes excess postprandial triglycerides. For details on how this works, you can consult this review article (Tushuizen et. al. 2004). On the other hand, a metabolically healthy person should be able to eat reasonable amounts of a wide range of macronutrient combinations without adverse postprandial effects.

Expanded model: a variety of bad things can disrupt one or more regulatory systems, resulting in postprandial abnormalities in glucose and/or fat metabolism.

How do Carbohydrates Affect Triglycerides?

You may have heard in low carb diet circles that it is actually carbohydrate, and not fat, that makes triglycerides go up. This is both true, misleading, and false. Let me explain.

What is almost certainly true is that a high carbohydrate diet can cause fasting triglycerides to go up. This is widely understood to be the case, and was noted by the German dieticians in the guidelines I linked to in my writeup on carbohydrates and HDL. In fact, HDL and triglyceride levels are tightly interconnected, and it seems to be the case that most things that raise HDL will also lower triglycerides. This probably has something to do with the action of cholesterylester transfer protein, but my head hurts every time I try to figure it out.

What about postprandial triglycerides? Well, ingestion of a sufficient amount of fat causes an acute rise in triglycerides, generally peaking between 2 and 6 hours. This fact is very well known in the medical community and has been known for a very long time. See e.g. this description of an oral fat tolerance test given to dogs by Arthur Knudson 1917. He fed fat to dogs and then watched the fat in their blood go up as they digested it. The protocol they used was also commonly used on humans and is remarkably similar to the methods used in modern studies investigating the same phenomena.

Absorption of fat is also associated with an acute phase inflammatory response. Lipopolysaccharide ("LPS") is involved, as discovered by Clett Erridge et. al. in 2007 ("A high fat meal induces low-grade endotoxemia: evidence of a novel mechanism of postprandial inflammation"). Postprandial increases in triglycerides have been found to be correlated with LPS in obese subjects (see Clemente-Postigo 2012). So the question is not whether fat consumption raises triglycerides. The question is how high does it raise them and how long does it raise them for?

A study from Jeff Volek's lab (Sharman et. al. 2004) showed that consumption of a very low carbohydrate diet can significantly reduce postprandial triglycerides in response to a high fat test meal. The study used two six-week dietary interventions (low carb and low fat) in a crossover design. The low carbohydrate diet contained approximately 10% carbohydrate. Analysis based on diet records confirmed that the participants were eating only 36 grams of carbohydrate per day on average. So in contrast to many other low carbohydrate diet studies, this one really used a low carbohydrate diet, for a sufficient duration to allow the initial phases of fat adaptation to take place.

Note that both diets were hypocaloric. As these were overweight individuals, we might expect metabolic health to improve on any reasonably dietary intervention of this duration that is low in calories.

At baseline and after the end of each six week intervention period, each participant was given an oral fat tolerance test consisting of a standardized high-fat test meal (the same idea as what Arthur Knudson did to his dogs in 1917). Serum triglycerides were measured hourly following the test and the results are shown below.

Triglycerides after high-fat test meal. Sharman et. al. 2004.

The first thing to note is that these were metabolically unhealthy individuals. The dotted curve shows the baseline result after the subjects had been consuming their normal diets. Note that the peak occurs at 4 hours and is very high (multiply mmol/L by 88.5 to get mg/dL, the peak looks to be approximately 290 mg/dL).

Second, note that both diets improved matters significantly compared to baseline in these overweight men. Both diets reduced fasting triglycerides (the "pre" point), and the low carbohydrate diet reduced them more as you would expect. The low carbohydrate group also has a much greater improvement in postprandial numbers, both in the height of the peak (it looks to be about 185 mg/dL), but also in its earlier time of occurrence. The total "area under the curve" is of course much lower in the low carbohydrate group compared to the low fat group or the baseline diet. The long tail of elevated triglycerides in the baseline and low fat groups is reminiscent of what I arbitrarily decided to call "inflammation" when we saw it in the obese diabetic rats. For the time being lets call it "inflammation" here too.

Now take a look at the points at 6-hour mark. This is where a person might typically be starting their next meal. If triglycerides are still high here from the last meal, the next meal will pile on top and drive them even higher. The triglyceride level for the low fat group at 6 hours is about where the peak triglycerides were in the low carb group at hour 3. The triglycerides for the low carb group on the other hand have dropped to about where they were when the low fat dieters were in the fasted state. A second high-fat meal at this stage would be much worse for the low fat dieters than for the low carb dieters.

Of course, typically the low fat dieters would not be having a fatty meal, they'd be having a crunchy carby meal. And that brings up a sensible criticism of this work. You might naturally expect the low carb group to better tolerate a high-fat meal because of their higher baseline fat consumption, in the same way that low carb-adapted folks may not do to well in a glucose tolerance test unless they are given the opportunity to adapt to carbohydrates over a couple of days beforehand. There is probably an element of truth to this. However I would point out that a metabolically healthy individual should be able to undergo a typical oral fat tolerance test while showing results similar to what the low carb group shows here. It looks to me as if the low carb intervention resulted in a substantial correction of a metabolic abnormality.

In addition, my experience testing my own triglycerides over the past year suggests that even adding a relatively small amount of carbohydrate to a high fat diet is enough to worsen postprandial triglycerides. At least this seems to be the case for me, and it would be very interesting to see if there is any research on this phenomenon. I would think the threshold would probably vary from one person to another, in the same way that people who successfully lose weight on low carbohydrate diets may have different threshold levels of carbohydrate consumption before things start to go pear shaped.

So What About That Pesky Lipopolysaccharide?

As mentioned, Clett Erridge et. al. showed that fat consumption permits lipopolysaccharide to enter the body through the gut, thereby inducing low-grade inflammation. It's a neat paper that looks at the problem from multiple angles -- worth a skim if you have time.

Lipopolysaccharide's infamous lipid A.

One might criticize Erridge's experiment by pointing out that the "high fat meal" consisted of a cup of tea and three slices of buttered toast. It was indeed high in fat, but you could argue, if you had PaleoTM inclinations, that it is actually the gluten in the bread and not the fat causing the gut barrier dysfunction. Gluten -> Leaky Gut -> LPS = Bad.

I would point out that Erridge's experiment was carried out in the United Kingdom, and it may have been considered unacceptably rude, or even against IRB requirements, to recruit subjects for a multi-hour ordeal without offering them a nice up of tea and a buttered toast sandwich. Anyway, subsequent work (see Deopurkar et. al. 2010) found the same effect after ingestion of dairy cream alone. I would like to see more replications of this result with different fats, but I won't hold my breath until someone tries it with a Paleo ApprovedTM wild caught grass fed organic fat source. Part of the trouble is that lipopolysaccharide is maddeningly difficult to measure, coming as it does in picogram quantities and having a potential circulating half-life measured in a few minutes.

The 2004 Sharman paper from Jeff Volek's lab (in addition to my own experiences) strongly suggests that LPS is not going to be a problem on a decent high fat diet. In that study, we saw the profound anti-inflammatory effects of a few weeks of fat adaptation. I suspect that, when and if the experiment is done, researchers will discover that fat adaptation either blunts the absorption of LPS, speeds its clearance from circulation, or attenuates the body's inflammatory response to it. We won't know for sure until they do the study, so lets hope at least that someone is working on it.

What About Those Pesky NEFAs?

Evelyn has written extensively at CarbSane about the potentially toxic effects of non-esterified fatty acids, which increase in circulation in the postprandial state. I don't know much about those, and I will be doing some more reading to get up to speed. As we saw with the fat diabetic rat study, it may be the case that NEFAs move in concert with triglycerides, so we can measure them by proxy. In other words, it may be the case that NEFAs come out of adipose tissue at the same time, and for the same reason, as the liver is producing excess VLDL particles. I will be on the look out for studies that suggest otherwise. For the time being, I'm not going to hold my breath until someone publishes (in an open-access journal) the exact NEFA study I'd like to see.

Some Personal Observations

So the CardioChek meter I have can measure triglycerides. Ordinarily this feature would be used in the fasted state as part of a standard three chemistry lipid panel (with total cholesterol, HDL and fasted triglycerides). However, nobody will stop you from using the triglyceride strips after meals (I won't tell). So I ordered a box last fall to play around with. The strips showed up a week into my sweet potato experiment, which I talked about previously in "Do Carbs Lower HDL?". I was still learning how best to use them (e.g. how long after a meal I should test), but I got some interesting results. I was surprised in a couple of instances to see numbers in excess of 200 mg/dl, for example after having a rib eye steak, a salad and a cup of nuts for dinner. These days, on a lower carbohydrate diet, I seldom see a peak reading over 150 despite some rather high fat meals. For example, readings 2.5 hours after a 1,000 calorie omelette usually range from 130 to 150 but can be much lower the day after heavy exercise. Readings at 3.5 hours are almost always lower than the 2.5 hour reading, indicating that the peak occurs before that time (the omelette contains 110 grams of fat, 75 grams of which are saturated, which is higher than the amounts used in the fat tolerance tests we've looked at).

My postprandial triglycerides were generally higher during my one month "safe starch" experiment, which is entirely consistent with the significant drop in HDL that I experienced during that time. I don't measure that frequently, but my maximum triglyceride level these days usually stays under 155.

A Digression on Modeling

I've written a little bit about the idea that we can better understand complex systems by building models, and then playing around with the models to see how they behave. Of course, models are no good if they don't reflect reality. On the other hand, a good model can be not only a useful clinical tool (see e.g. the minimal model of insulin), but can also help bring about important discoveries (e.g. the dominance of the kidney fluid mechanism of blood pressure control using the Guyton molel).

So I decided to construct a simple model to see if it could help me understand the triglyceride readings I've been seeing. The model is described by the diagram below. It is the simplest thing I could come up with.

A simple model of triglyceride absorption and clearance.

The model has two compartments: the gut and the systemic circulation. The rest of the tissues in the body take up lipids from circulation, but are modeled as an infinite sink and are not separately modeled. Ingested lipids are absorbed from the gut into circulation, and then cleared from circulation into the tissues. The purpose of the model is to understand what the dynamics of that process may look like.

The initial conditions are determined by the amount of lipids ingested. This quantity (L) decreases as lipids are absorbed, at a rate equal to a constant (p) representing the permeability of the gut, multiplied by the amount of lipids remaining. Mathematically, the gut compartment is described by a simple first order differential equation, dL/dt = -pL.

Circulating lipids, represented by the variable C, increase when lipids are absorbed and decrease when they are cleared. The absorption term is the same as that for the gut compartment (pL) with the sign reversed. Intuitively, this means that every bit of lipid that leaves the gut immediately enters the circulation. The clearance of lipids from circulation is modeled by a term similar to the term describing absorption in the gut compartment. Lipids are cleared from circulation in proportion to the amount remaining in circulation, multiplied by a clearance factor (c). There is no accommodation for the metabolic state or storage capacity of the adipose or other tissues of the body. Mathematically, this gives us another first order differential equation, this time with two terms: dC/dt = pL - cC. In this model, the amount of fasting lipids is normalized to zero. In real life there will be a constant offset to the measured value of C.

The goal is to understand how the amount of circulating lipids changes over time. This could give some guidance in interpreting triglyceride readings taken at different times after a test meal. The model is very simple so far and does not yet take into account the mysterious elevations in triglycerides that I have so far been arbitrarily calling "inflammation." So let's look at it for now as a picture of the postprandial triglyceride curve for an idealized perfect human that has no inflammation.

Triglycerides vs. time in response to a single ingested bolus, based on a simplified model.

I will not delve deeply here into the mathematics behind the model. Suffice it to say that after solving the two differential equations and making a biologically plausible simplification, you get a curve of the form C=axe^-bx+k. The model has three degrees of freedom, corresponding roughly to the height and width of the curve, plus a constant factor which can be calibrated to the fasting triglyceride measurement. This equation can now be taken into a statistical computing environment and tested against real data.

 

A Test of Three Foods

Here is a little bit of the data I have collected so far. The chart below shows the results of a series of test meals each containing a single food: avocados, macadamia nuts, and coconut oil (which was emulsified in warm water). In each case the meals were standardized to 75g of fat based on tables from the USDA nutrient database. Each food is predominantly fat, but the fatty acid profiles are quite different. Each food was consumed in its whole form along with its usual vitamins, minerals and other micronutrients.

The tests were conducted on successive Saturday afternoons in the fasted state, during the same month when I was doing the safe starches experiment and eating 100g of extra carbs from sweet potatoes per day. The fatty acid profiles of these three test meals are very different, as was their digestibility (particularly so in the case of the coconut oil).

The three foods had very different effects on my postprandial triglycerides, with avocados being the worst and coconut oil being the best. Of course the coconut oil would have produced a great deal of circulating ketones, which I did not measure. I believe the numbers below are roughly consistent with research I have seen comparing the effects different fatty acids. In addition, the data suggests that, as you might expect, the macadamia nuts were digesting slowest of all three fats.

The points in the graph below are the actual measurements and the curves show the best-fit solutions to my simplified triglyceride dynamics model. The curves were fitted in R using nonlinear least squares. The fits are not perfect of course, as the model is oversimplified at present and there will always be measurement error to contend with. In the next section I will discuss a possible enhancement to the model.

Triglyceride values and fitted curves based on the idealized computational model. Green: avocado; red: macadamia nuts; brown: coconut oil. Each meal was eaten at time=0 on successive Saturday afternoons in the fasted state. Each meal standardized to 75g fat.

Using the Model: Idealized Response Plus Inflammation?

I took a look at the data from the Sharman paper in light of my simplified model. Since the very low carbohydrate group in the Sharman study showed the lowest fasting and peak triglycerides, as well as the fastest clearance, we can assume they suffered the lowest chronic and postprandial inflammation. Therefore it is expected that the shape of the curve for the very low carbohydrate group would most closely approximate the ideal shape predicted by the simplified model. This is indeed the case in this data set.

As mentioned above, my presumption is that the idealized model will most closely reflect the dynamics of lipid absorption in a perfectly healthy, non-inflammatory state. Differences between actual data readings and the model predictions could be used to gauge the extent of both chronic and acute/postprandial inflammation. These results can be obtained with this data set by fitting the curve predicted by the model to the data points obtained from the low carbohydrate group. In this case, only the points along the rising left hand slope were used for the fit, in order to prevent the fitted curve from overshooting.

The graph below on the left shows the actual data from Sharman et. al. in red (baseline), green (low fat) and blue (low carb). The black line is the fitted model prediction. The graph on the right shows the differences between the actual readings for each series and the predictions of the model fitted to the low carbohydrate points.

The offset at time t=0 represents elevated fasting triglycerides, a chronic abnormality, while the differences in area under the curve, after adjusting for the baseline difference at t=0, could reflect different postprandial inflammatory responses.

Time on x axis in hours (test meal at t=1). Blue=very low carbohydrate, green=low fat, red=baseline, black=model fitted to very low carbohydrate using nonlinear least squares (nls()). Left chart shows original triglyceride measurements and model. Right chart shows original data points in each series minus model fit to the very low carbohydrate data points. Model was optimized for fit to points at t=1, 2 and 9 hours. Data from Sharman et. al. 2004.

The Sharman data are interesting because the three series, though quite different in appearance, were taken in the same individuals in a balanced crossover study design, and using the same standardized challenge meal. We don't need to calibrate the curves to account for changes in body size, genetics or any other factor. The only variable is the metabolic state of the individual resulting from the dietary intervention. This is why it makes sense to compare all three curves to the model prediction as fit only to the low carbohydrate data points.

We see that all three curves reflect an increase in "area under the curve" as compared to the idealized model. One hypothesis is that the low fat and baseline curves are higher because of blunted absorption of fat. However, when fiddling with the idealized model, it becomes apparent that blunted absorption would result in a more skewed triglyceride curve, instead of the more symmetrical ones observed here in connection metabolic disturbance. Therefore the model suggests this hypothesis is not correct. (Note for math nerds: the approximate model solution presented in this post has only three degrees of freedom and does not exhibit this skewing behavior. You need to work with the full solution with all four degrees of freedom to see it.)

The liver increases production of VLDL particles after meals (see e.g. Timlin and Parks 2005), contributing to the postprandial rise in triglycerides. This production is further increased in response to inflammation (e.g. exposure to LPS -- check out this awesome paper by Barcia and Harris to see why that might be happening). Now take a look at the chart on the right hand side above. Note that the curves are all approximately the same shape, but differ in height. Note that they all peak around t=5 (four hours after the meal), just as Timlin and Parks say they should if they reflect hepatic VLDL production. So basically that's why I think it's inflammation.

Based on the above, the constant term in the idealized model can be used to represent chronic inflammation. As before this would be calibrated to fasting triglycerides, with the caveat that they would need to be truly fasted numbers (i.e. 8-10 hour fasts may not be sufficient in individuals who present extended post-meal inflammation). In addition, a term can be added to reflect the release of VLDL triggered by the acute inflammatory response caused by test meal.

Concluding Remarks

Mainstream recommendations are that fasting triglycerides should stay below 150. Now I'm not sure a truly metabolically healthy individual will have fasting triglycerides of 150, but as a postprandial number, it seems to be a reasonable target. Based on the evidence presented above, I think a more significant factor may turn out to be the existence of an inflammatory tail, which would show up in the 4+ hour postprandial readings. However, at least as compared to blood sugar, the epidemiological evidence is quite thin, so it is hard to recommend any particular thresholds of concern.

From browsing around the scientific literature and my own experience with the test strips, there appear to be a variety of things that can cause elevated postprandial triglycerides. Here are a few:

• Fed vs fasted state (calorie excess or deficit)
• Exercise (beyond the effect caused by energy depletion)
• Type of fat in the meal
• Type of fat in baseline diet (particularly, lack of omega-3 fats)
• Baseline carbohydrate consumption
• Lipopolysaccharide and inflammation in general

Funny, this looks sort of like the list I laid out earlier for the factors affecting blood sugar.

A 2010 article ("Dietary cholesterol and egg yolks: Not for patients at risk of vascular disease") makes the case that eggs are a bad idea because their cholesterol content will lead to elevated postprandial triglycerides. That hasn't been my experience, but it would be pretty straightforward to test it. A test of two meals, calibrated for total fat and calories, with a non-cholesterol containing fat substituting for egg yolks in the control meal.

A variety of lines of evidence suggest that the polyphenols in olive oil could inhibit postprandial inflammation. The experiment here would be to compare the triglyceride response to extra light olive oil vs. a very spicy unfiltered extra virgin. The fatty acid content should be similar but the oils would differ greatly in polyphenol content.

It has been suggested that oxidative stress mediates these phenomena. A high dose of vitamin c or e, or a glutathione precursor (e.g. whey protein) may show an effect when consumed near the meal. It would be interesting to see how the triglyceride readings vary based on the timing of the antioxidant dose relative to the meal.

Overall I think triglycerides are a good target for self-experimentation. They seem to provide quick insight into metabolism and inflammation that is hard to get any other way. Experiments can be done quickly instead of waiting months for fasting lipoproteins to reach homeostasis. That said, there is some inconsistency in readings, suggesting significant measurement error and perhaps a hypersensitivity to your state of health, making single readings quite difficult to interpret. Likewise, readings taken far apart in time can be very difficult to compare.

In the mean time, it is interesting to see what kinds of insights we can get by building simple models of ourselves in our computers. If you'd like to try it yourself, take a look at CellML, an open standard for storing and sharing computational models. A free software implementation called OpenCell is available for Linux, Windows and Mac OS.

Butter Not Like a Statin After All?

I began my original butter experiment without a hypothesis about the effects of butter on cholesterol. I was tracking blood lipids so that I could identify any possible adverse reactions. After initially seeing a positive change, the numbers started to go the wrong way, with HDL dropping and non-HDL going up. At the same time, I had switched from the Kerrygold butter I usually used to a local brand from my farmer's market. It was supposed to be "100% grass fed," but it did not have the deep yellow hue of the Kerrygold, so I suspected it could be responsible for the adverse results. After a day or two using an Icelandic brand (Smjor), I decided for the remainder of the experiment to test only Kerrygold butter. A few days after the switch back, the numbers moved back in the positive direction. I tracked for a few more weeks and then reported my initial results here.

After the original post, I spent about three more weeks on the same diet (1/2 stick of Kerrygold butter per day), followed by three weeks of butter elimination (during this elimination phase, there was some increase in coconut oil as a partial substitute for the butter calories). Finally, I went back on a high-butter diet for an additional five weeks. During this last phase I decided to double down and eat a full stick of Kerrygold butter every day (about 112 grams of butter and 90 grams of fat). The diets were otherwise similar, though I was likely consuming more total calories during the high-butter phases. Data points were generally taken weekly, on Sunday afternoons (not fasted, but I don't believe it makes much difference, since I am not measuring triglycerides).

Based on the additional data I can report that the addition of butter to my diet may have increased my HDL cholesterol. However, it has not had a convincing long-term effect on my non-HDL cholesterol. The HDL remains elevated from before, but non-HDL readings are consistent with my average numbers over the past three years or so. My working hypothesis at this point is that the change in non-HDL I observed originally was a temporary disturbance of homeostasis. This may have been due to the butter or to another unknown cause.

I also believe this data suggests that different brands of butter can have different effects on cholesterol values. I will show this data below, though given the design of this experiment, I will not be able to prove it without a follow-up experiment. However, I am satisfied with my original concern, which is that the addition of a good brand of butter to my diet does not seem to have any measurable adverse effect on cholesterol measurements.

Results

The additional data points are shown in the two plots below. The four green points show the data that was dropped from the original data set because I was not using Kerrygold butter. The last five points are for the latest diet where I am eating a full stick of butter per day.

HDL measurements over time. Red points indicate control. Blue and green points indicate experimental group with consumption of at least 0.5 sticks of butter per day for at least 7 days prior to testing (blue=Kerrygold). No statistically significant difference in control vs. experimental measurements when green points are included. Trend line is for red points only.

Non-HDL cholesterol over time. No statistically significant difference between control vs. experimental measurements when green points are included. Trend line is for red points only.

The next two plots show the same data, zoomed in to show points from the beginning of the original butter experiment to the present time. The three red points are the latest butter-elimination phase.

Trend line for blue points.

Trend line for blue points.

Overall, I think there is some basis for concluding that added butter is having a positive effect on HDL, although the results are not statistically significant when the green points are included. However, it is also possible that some other factor is concurrently causing my HDL to increase.

The trend towards positive HDL outliers has continued in the butter dataset and is especially strong in the latest readings where a full stick per day is consumed, suggesting a dose-response relationship. My last three readings are all greater than 100, which is the meter's measurement limit for HDL cholesterol. Since I am computing non-HDL cholesterol as Total minus HDL, this measurement limitation will inflate the non-HDL points by an indeterminate amount at the same time as it underestimates HDL. The highest reading ever recorded in the control dataset is 83, while the butter dataset has 5 readings above that level and three above 100.

The non-HDL numbers do look like noise at this point, except for the original cluster of low readings when the original butter experiment began. There probably was some underlying cause for that, though it is clear to me that if the butter had anything to do with it, the effect was temporary.

Note also that there are several negative outliers for HDL in the non-Kerrygold butter set. This suggests that different brands of butter can have different effects on cholesterol measurements. This phenomenon may warrant further study.

What's the Take Away?

Okay, I hear what you're saying: "Greg, that's a heck of a lot of colored dots. So what?" I think the bottom line is, I'm eating a whole stick of butter every day and my lipid profile has either stayed the same or gotten a little better. Works for me.

Measurement Accuracy

During this experiment, I observed that the CardioChek PA seems to be very accurate at measuring HDL, but much less accurate at measuring total (and therefore non-HDL) cholesterol. One of the newly collected data points shown above is the result of a VAP cholesterol panel ordered by my doctor. The same morning as the blood draw for that test, I took three successive readings of HDL and total cholesterol with the CardioChek PA. The VAP test showed an HDL of 68 and non-HDL of 233. The three HDL measurements on the CardioChek PA averaged out to 68, same as the VAP result, with a standard deviation of only 1.7%. For non-HDL, on the other hand, the average was 191 with a standard deviation of 15%. Compared to the 233 non-HDL measured by VAP, the CardioChek PA was off by 18% even when three separate measurements were averaged.

Therefore, variance seen in my HDL readings over time suggests that real changes are taking place in my body. However, variance in my non-HDL measurements will be substantially the result of measurement error in addition to physiological changes. To put this in perspective, the standard deviation of all of my non-HDL measurements over the last 2+ years taken together is only 15.8%, barely higher than the meter's built-in measurement noise.

The CardioChek also sells test strips for direct LDL measurement. I will probably check those out at some point, as they may prove to be more accurate than the total cholesterol strips.

Next Steps

Ideally, self-experimentation would be based on biomarkers that can be measured accurately and cheaply, and that quickly reflect real health trends. Blood sugar measurements are very useful in that regard, because it only takes one finger prick an hour after a meal to obtain useful information about your body's response to carbohydrates. Over the course of this experiment, I've become skeptical of the value of cholesterol measurements as a target of self-experimentation. Given the meter's built-in measurement noise and the long half-life of circulating lipoproteins, it takes weeks or months for changes to be clearly evident. That makes it difficult to do useful experiments.

My most pressing health goal until the end of 2012 is a 405 pound deadlift. High cholesterol and extra calories from butter are probably going to help me achieve that goal, so I will be putting my cholesterol investigations on hold for the time being. In the mean time I do have a few other experiments to report on, so stay tuned.

Follow Up on Butter

Welcome readers. I wanted to post some follow up details on my experiments so far and say a bit more about the themes I will be writing more about. You'll notice that my headlines so far have been phrased as questions. This is not an homage to a popular quiz show -- they are questions because I don't know the answers. So with that in mind I wanted to revisit the results of my first experiment (Is Butter as Powerful as a Statin?).

Science is hard, and if you attempt it, you need to stay vigilant about not fooling yourself, because, as Richard Feynman has said, you're the easiest person to fool. So I'll be reviewing my initial results to see whether I've fooled myself. Of course when it comes to personal science, we should always be aware of this principle in reviewing work done by others. If someone has managed to fool himself (even for a moment), perhaps they might fool you too. I hope my readers will keep this in mind.

Commenter EricT highlighted the biggest weakness in my butter experiment.  I originally added the butter to my diet to see what sorts of subjective effects it might have (they were positive, but I did not track anything quantitative that I can present). I tracked my lipids at the time only in order to monitor whether they might move in an adverse direction. It looked like the Kerrygold butter caused beneficial changes while other brands did not. I therefore changed my hypothesis mid-stream, which is a substantial source of bias, effectively rendering my statistical analysis invalid (this particular weakness is not applicable to the second experiment on safe starches vs. HDL, though of course other weaknesses do apply). Meanwhile I am continuing the butter experiment and will present all additional data points in a couple of weeks so we can see whether the effect holds. In the mean time, by then I will have data from another VAP cholesterol test so I can have more confidence about what my CardioChek meter is actually measuring and how consistent it is. Stay tuned for that.  Meanwhile I have declared a "grain of salt" advisory on the butter post.

If the additional data points do not show a continuation of the effect seen previously, it would weigh strongly against the hypothesis that butter causes beneficial changes in lipids (though this is somewhat independent to the question of whether butter is good or bad from a cardiovascular perspective). Many alternative hypotheses could explain a negative outcome -- the effect could have been the result of the bias, a short-lived disruption of homeostasis, a systematic user error or environmental sensitivity with the CardioChek device, or an unknown and temporary cause coinciding in time with the applicable data points. Since I did not decide beforehand how long my experiment would last, the effect could also have been a statistical anomaly which only seemed "significant" because of the particular day on which I decided to compile my results (that is, you can't just wait until you get the result you want and choose that as the date on which to terminate your study).

In the mean time, I've been measuring another health marker that is related to gut health and the microbiome.  I am going to write about about that and then present some interesting and hopefully useful data over the next few weeks.

Do Carbs Lower HDL?

Attack of the Safe Starches

If you have been lurking around paleoland recently you will have heard the debate about safe starches (you may be sick of hearing about it in fact).  Last fall I ran an experiment to find out whether it might be a good idea to dip my toe into the safe starch swimming pool. I was tracking my lipids around this time with the CardioChek PA so I could have some idea of what was going on with my metabolism.  Here are the results.

I first experimented with a low carbohydrate diet in 2009 after reading Good Calories, Bad Calories by Gary Taubes. I lost 15 pounds in the first three months (though I had no idea I was carrying any extra fat).  I gained muscle without changing my exercise program.  My seasonal allergies went away.  My teeth got whiter, less sensitive and stopped collecting plaque.  I got fewer sunburns.  My joints stopped aching after exercise.  You get the idea.

Around this time I took an oral glucose tolerance test and found my numbers to be a bit high, though not in the pre-diabetic range just yet. I tend to get a high initial blood sugar spike, though the value quickly returns to normal. I thought this (along with the other general health improvements I experienced) was an indicator that a low carb diet was a good approach for me. I bought a cheap glucometer to play with and stuck with the low carb program.

Over the years my diet progressed to a paleo approach, while I continued to avoid starches. On most days I ate only eggs, meat, fish, some nuts and a good helping of green veggies. Given the safe starches controversy, I thought it would be interesting to try adding 100g or so of carbs per day in the form of sweet potatoes, just to see what would happen. There is a variety of chatter about the possibility that low carbohydrate diets can raise LDL, so I thought I might see a drop in non-HDL cholesterol with a bit of added starches. I did get a drop, but it wasn't to the non-HDL.

Results

Mean non-HDL did not change during the experiment (193 to 203, not statistically significant), but mean HDL dropped by a significant amount (67 to 57, p<0.001). I stopped the experiment after one month. You'll see in the plot below that the HDL went back up to the previous level after the daily sweet potato was dropped.

Boxplots show minimum, maximum and quintiles. Mean HDL: 67 (control), 57 (carbs).

The trendline is a linear model based on all low carb data points (before and after, but not including, the carb intervention).

Trendline based on all low-carb points.  Non-HDL cholesterol was not significantly affected.

Details of the Experiment

The data points were collected as previously discussed in the butter experiment. Data does not include any points during the butter experiment, as there was a significant change in HDL and non-HDL during that time. Statistical significance test is based on a linear model, calculated by ANOVA using R.

Sweet potatoes were eaten slowly so as not to spike blood sugar above 120, though spikes may have occurred on occasion. It usually takes me about an hour to eat one, though I can eat them very fast without a spike if I've recently completed a heavy workout.

There are a number of limitations to the interpretation of this data that are worth mentioning. Perhaps most significantly, this is a very short term test (1 month). It is possible that long term adaptations would have reversed the effects seen here. Second, there was only a single intervention period, which could have corresponded to another unknown variable which caused HDL to decrease. Since foods differ in micronutrient, mineral and toxin content, it is not possible to generalize from sweet potatoes to all carbohydrates.  I may have had a different result with white rice, white potatoes, taro, tapioca, etc.  Finally, of course this result applies to me only. There may be others who have a much higher carbohydrate tolerance, or who may even see the opposite result.

Is This Actionable Information?

So why am I bothering with this? First of all, simple curiosity. I have the ability to collect data that most people do not collect, and there is some interesting science concerning these molecules and what they may be able to tell me about my metabolism.

Does the decrease in HDL identified here represent an unhealthy change? Perhaps. Though HDL is commonly referred to as the "good cholesterol", I would not suggest that all decreases in HDL are unhealthy.  In fact I don't know how one would go about establishing the truth or falsehood of a statement like that. Conversely, we know of chemicals that raise HDL while simultaneously causing heart attacks.  My HDL was never "low" during the course of this experiment. However, together with certain other data I was also collecting at this time (which is a story for another day), I believe that 100 grams of starch per day is too much for me. At least in my current metabolic state, with my current lifestyle, exercise habits, sleep, stress level, etc.

Some smaller amount of starch is most likely "safe," and may be a good idea. These days I often eat a banana (about 20-30 grams of sugar+starch) after lunch, and I have not seen the negative effects caused in me by higher amounts of carbohydrate. As mentioned, I also have found that I can eat 100 grams of carbs immediately after a heavy workout with no significant change in my blood sugar (e.g. it might increase from 65 to 83 in response to a pound of sweet potatoes after two hours of powerlifting). These days, workouts like that occur about once a week (sometimes less), and I have not noticed any negative effects from eating carbs this infrequently.  Since many smart (and strong) people advocate carbs post-workout, I am willing to go along for now, at least while I do not have any contradictory evidence.

Published Research

Perhaps if I had looked this stuff up in the scientific literature first I would have been less surprised by my results. It turns out there is plenty of published research supporting the idea that carbs lower HDL. However, it is a bit surprising that the relationship continues even below 100g/day. I doubt there is published research on this relevant to long-term low carbers, so three cheers for personal science on that front. A study by a German team, published in January 2012 in the Annals of Nutrition and Metabolism, does a good job summing up the research that is out there.  Here is their conclusion regarding carbohydrates and HDL:

"There is convincing evidence that a higher carbohydrate proportion in the diet at the expense of total fat or saturated fatty acids intake lowers the plasma concentration of HDL cholesterol."

The paper is called "Evidence-Based Guideline of the German Nutrition Society: Carbohydrate Intake and Prevention of Nutrition-Related Disease." It is worth a read if you are interested and still awake. If instead you are sleeping (and not German), perhaps you are dreaming of a world in which the nutrition organizations in your country also use evidence as the basis for their guidelines.

Is Butter As Powerful as a Statin?

Notice: A Grain of Salt Advisory is in effect for this post. See comments and this new post for details. Subsequent data collection has changed the conclusions presented below.

Butter is an interesting food. For those who have graduated from an un-scientific fear of fat, it would seem to encapsulate almost everything that is good about cow's milk, with only small amounts of the questionable stuff (lactose and milk proteins). What's not to like? Certainly some very sensitive people could respond negatively to the trace sugars and proteins in butter, but for most people it seems like it should qualify as a health food.

Health professionals are usually quick to warn us away from butter because it is high in artery-clogging-saturated-fat. The old story goes that butter is bad for you because saturated fat raises cholesterol, and elevated cholesterol causes heart disease. I can't do an n=1 experiment on the effects of butter on mortality but with a home cholesterol meter it is easy to find out whether the first part of the old story is true. Does butter actually raise blood cholesterol levels? Whether that has anything to do with heart disease is a story for another day (I happen to believe there may be a small relationship between cholesterol and heart disease for some people, but it's complicated).

In fact, eating half a stick of butter a day significantly lowered my cholesterol. It lowered it about as much as the reported effects of many statins (sorry, haven't tried the statins so you'll have to trust the clinical trials on those). Read on for the details, but first a few words about butter.

What color is it?

These days, most consumers expect butter to be white and margarine to be a pleasing yellow. Never mind the fact that margarine was originally colored yellow in order to make it look like healthy grass-fed butter. For years, many states had laws prohibiting the sale of artificially colored margarine. Housewives were sometimes sold a small packet of food coloring with each tub so they could trick their families into thinking they were eating something that was good for them.

Butter with a high vitamin content is yellow, not white. Unfortunately, artificially colored butter is also yellow. In fact, butter and cheese were the first foods the U.S. federal government allowed to be artificially colored. This gave butter producers a leg up in competing against margarine, where artificial colors were usually not allowed. For most foods, added colors need to be identified on the label and/or ingredients list. Butter is exempt from this requirement (see Section 101.22(k)(3) of Title 21 of the U.S. Code of Federal Regulations).  Therefore, do not judge American butter by its color.

The best butter is made from milk produced by cows that live outside and eat their natural diet, grass. The "organic" label can be applied to corn-fed cows raised in feedlots. Therefore, the term doesn't really tell you anything useful about the quality of butter. It is about as useful as calling farmed fish "organic." Personally I don't pay any attention to that label when it comes to butter and I look for butter that I am confident comes from cows that live outside and eat only grass.

Butter throughout History

Anthimus, an exiled Byzantine doctor writing in northern Gaul in the 6th century, recommended butter with honey to patients suffering from tuberculosis. This is interesting because, for many years prior to the introduction of antibiotics, cod liver oil (which, like grass fed butter, is rich in vitamin A) was the standard of care for tuberculosis treatment.

The dentist Weston A. Price found butter to be a staple food in many healthy populations around the world.

According to MS Iceland Dairies, which produces the Smjor brand of butter, the Icelandic people prized butter as an important and healthy food.

"People usually had considerable stores of butter; it was used extensively, and it was considered a source of strength during the harsh winter months when people needed to consume a lot of fat. The recommended portion was approximately 1700g a week per person."

For the Americans in the audience: 1700 grams of butter is over 3.7 pounds. Per week.

Butter seems to be especially valued in cold climates, perhaps because it could be stored prior to refrigeration without going too far rancid. For example, see Tibet and their famous rancid yak butter tea.

Cholesterol for Breakfast

As I mentioned above, I don't believe there is a strong relationship between blood cholesterol levels and heart disease risk. In addition, if there is a relationship between blood cholesterol levels and total mortality, it appears to be non-linear and probably U-shaped, with increased mortality when levels are too high or too low. That said, my cholesterol has been a bit higher than I'd like over the last 10 years or so, and I've tried on and off to see if there is a reasonable way to change it without doing something crazy like eating laxatives or plant hormones. Over the past year I've been monitoring my blood lipids from time to time with a CardioChek meter and paying attention to what causes the numbers to change. Most of the data for this experiment comes from the CardioChek machine, though some readings are lab tests ordered by my doctor.

Notwithstanding the absence of supporting evidence, the 2010 Dietary Guidelines for Americans recommend that humans consume no more than 300 mg of cholesterol per day. Now a few times in the past two years I have substantially varied my egg intake, and therefore my total dietary cholesterol. Although I don't have enough readings to get good statistics on these changes, my observation is consistent with the published scientific literature. There is no association between dietary cholesterol, per se, and my blood cholesterol. My cholesterol generally stays the same or perhaps drops a little when I eat more eggs (e.g. going from no egg consumption to 4 per day). There may be some subset of the population that is sensitive to dietary cholesterol, but I do not seem to fall into that category.

The Thousand Calorie Omelete

This story is not about eggs or dietary cholesterol. It is about what happened to me when I added a large amount of butter to my breakfast. In January I added half a stick of butter (about 57 grams) to my morning omelete. I first got this idea from Seth Roberts, who reportedly has been eating half a stick of butter per day for some time in order to increase his mental performance.  Seth believes butter has been good for his heart. Dave Asprey also recommends eating lots of butter for brain health, among other reasons. My omelete consists of four pastured eggs, half a stick of Kerrygold Irish butter, and three tablespoons of coconut oil. Thats 1,076 calories, 110 grams of fat (75 grams from saturated fat) and 865mg of cholesterol. My diet otherwise stayed the same, though I was usually eating smaller portions of everything else.

The effect this had on my blood cholesterol was significant enough to inspire me to finally start a blog. Of all the interventions I've tried over the years, this has been by far the most powerful and the first one to show up clearly after the first couple of data points.

Here are the results. On average: HDL is up 19%, non-HDL is down 25%. Both results are statistically significant with high confidence levels, despite relatively small sample sizes (p<.001 and .01 respectively). Here are some boxplots to make it more convincing. The plots show min, max, mean, and the 1st and 3rd quintiles.

Boxplots showing min, max, mean and quartiles.
Mean HDL (control, treatment, change): 65, 77, +19%.
Mean non-HDL (control, treatment, change): 196, 148, -25%

The following plots show my pre-butter trends for HDL and non-HDL in red, with the post-butter readings in blue (thanks Seth for the suggestion).  This suggests that the new readings do not reflect the continuation of a pre-existing trend.  I did not add trendlines for the post-butter readings because there are not enough datapoints for them to make sense.  Note that 5 of the 6 post-butter HDL readings are the highest ever recorded.  4 of the 6 non-HDL readings are the lowest ever recorded.

What's the Best Butter?

When I was growing up, I was quite a picky eater, and was very specific about the brands of butter that I would allow my parents to bring into the house.  Most brands tasted downright rancid.  Even still, I don't think I ever tasted good butter until I was well into my adult years.  The bad stuf is not worth eating, and I now have a bit of experimental evidence to back that up.


My butter of choice is Kerrygold from Ireland. While Kerrygold seems to be somewhat coy about saying "100% grass fed" in its marketing materials, it is pretty consistently cited as the best butter that is readily available around the U.S.  Trader Joe's almost always has the best prices in my vicinity.  During the course of this experiment, I tried two other types of butter.  One was Smjor, mentioned above.  The other was a "grass fed" butter from my local farmers' market.  The Smjor was delicious and very yellow, so it looked like a good pick.  The local butter was white and not especially appetizing, perhaps understandable for cows living in New York in the winter and necessarily given some kind of supplemental feed.  The cholesterol readings I was getting from the Smjor and local butter were all over the map and were not consistent with the numbers I got with the Kerrygold.  Therefore I did what any good scientist would have done and threw those data points away so I could present results only for thousand calorie omeletes made with Kerrygold.

Details of the Experiment

There are 6 measurements in the experimental (butter) group and 16 in the control (non-butter) group. Additional measurements taken within 1 week were not used. The butter group includes all measurements made when I had been eating at least 1/2 stick (about 56 grams) of Kerrygold butter per day for at least a week prior to measurement. Datapoints when I was consuming other brands of butter were not included. The first five measurements were VAP cholesterol tests ordered by my doctor.  Otherwise, I measured total cholesterol and HDL cholesterol in a single panel with a CardioChek PA. I did not measure LDL directly, and I did not calculate it because my fasting triglycerides are not measurable with my meter (<50). Besides, the LDL calculation formulas are nonsense at best when triglycerides are under 100. I use non-HDL as a rough estimate of the atherogenic potential of blood lipids (basically a crude proxy for apolipoprotein B).

The increase in HDL and decrease in non-HDL were both statistically significant as determined by ANOVA using R (p<.001 and .01 respectively). However, I should point out that, for a variety of reasons, these are not high quality statistics. First of all, I have no formal training in statistics. Given that some data points were taken fairly close together in time, there is likely some serial correlation between measurements. I also made no particular effort to control for other variables. Finally, these results apply to me. There may well be people who respond differently and for whom butter is unhealthy. As with anything, it is always advisable to test things for yourself. So take this with a grain of salt. Or half a stick of butter, if you prefer.