There is an interesting new study out that looks at gut bacteria, a personalized-diet-creating algorithm, and blood sugar. Basically, they collected a bunch of data including biometrics (e.g. age, weight, hip circumference, A1c), family history, and, notably, gut bacteria (in a stool sample), then had participants use a continuous glucose monitor and keep track of their food, exercise, and sleep for week, then used that data to create an algorithm predicting blood sugar response, used that algorithm to create personalized “good” and “bad” diet plans that would either keep blood sugar levels down or spike them, respectively, pitted those diets against “good” and “bad” personalized diet plans created by a dietician with access to the same meal and glucose monitor data, and along the way looked at the way glucose response changes and is changed by gut bacteria.
I read a few articles and a blog post about this study before reading the full text of the study itself, and whether because I was reading quickly and missed a small but important word that went against the general emphasis of the article (I missed a “both” in this Atlantic article), information was omitted, the articles were just unclear, or actually misleading (it seemed like this ScienceDaily article’s repeated vague references to diets “working” were designed to encourage people to think of weight loss rather than blood sugar response), there were several ways that the study was not what it initially appeared.
Some interesting things they found that weren’t part of the main outcome of the study, but they noticed by looking at their data and their algorithm:
-Some people’s blood sugar response was very closely tied to how much carbohydrate was in the meal, others’ was not.
-For some people, eating fat with a meal greatly lowered their blood sugar response to the meal, for others it didn’t make much difference.
-Fiber in a meal increased blood sugar response to that particular meal, but having consumed fiber in the previous 24 hours decreased blood sugar response to later meals.
-The time passed since last sleeping and the sodium content of a meal increased blood sugar response, and proximity to exercise and water and alcohol content of a meal reduced blood sugar response.
One caveat to bear in mind: the full text of the study says that they only analyzed glucose response for certain types of meals:
Since real-life meals vary in amounts and may each contain several different food components, we only examined meals that contained 20–40 g of carbohydrates and had a single dominant food component whose carbohydrate content exceeded 50% of the meal’s carbohydrate content.
If subsequently following the personalized diets showed that the personalized diets performed better than “expert” diets over the course of an entire week, not just after this narrow category of meal, then this is not a big problem, or at least this is still as good as the “expert” advice.
One thing that articles about this study did not make clear or deemphasized was that the “expert” diet was also personalized, the “experts” had access to the weeklong meal and continuous glucose monitor data, and many foods that were on one person’s “bad diet” were on another person’s “good diet” in the expert diets as well as the algorithm diets:
The second, “expert arm,” was used as a gold standard for comparison. Participants in this arm underwent the same process as the prediction arm except that instead of using our predictor for selecting their “good” and “bad” diets a clinical dietitian and a researcher experienced in analyzing CGM data (collectively termed “expert”) selected them based on their measured PPGRs to all meals during the profiling week. Specifically, meals that according to the expert’s analysis of their CGM had low and high PPGRs in the profiling week were selected for the “good” and “bad” diets, respectively.
Both study arms constitute personalized nutritional interventions and thus demonstrate the efficacy of this approach in lowering PPGRs.
Post hoc examination of the prescribed diets revealed the personalized aspect of the diets in both arms in that multiple dominant food components (as in Figure 2F) prescribed in the “good” diet of some participants were prescribed in the “bad” diet of others (Figures 5G and S6). This occurs when components induced opposite CGM-measured PPGRs across participants (expert arm) or were predicted to have opposite PPGRs (predictor arm).
The Atlantic article in particular made it sound like it was only in the algorithm diet that surprising recommendations were made, but Figure S6 of the full study shows that, for example, both potatoes and pizza sometimes appear in the expert group’s “good” diet, and sometimes in the “bad” diet.
The Atlantic article also reported that the algorithm diet “performed as well as the two experts, if not slightly better”, but it looks like the “slightly better” was not statistically significant (especially given how few participants there were in this particular part of the study), since in the full text of the paper, it says the success of the algorithm diet and the “expert” diet were “comparable”:
Notably, for 10 of the 12 participants of the predictor-based arm, PPGRs in the “bad” diet were significantly higher than in the “good” diet (p < 0.05, Figure 5C)… The success of the predictor was comparable to that of the expert-based arm, in which significantly lower PPGRs in the “good” versus the “bad” diet were observed for 8 of its 14 participants (p < 0.05, 11 of 14 participants with p < 0.1, Figure 5C).
Other articles, instead of saying that the algorithm diet was better, just said that the algorithm diet did great without even mentioning that the personalized “expert” diet did about the same–for example, the ScienceMag article on the study simply said, “Segal and his team then used the findings to develop customized diet plans for a group of 26 people,” without explaining that about half of those people followed the “expert” diet and the other half followed the algorithm diet.
The study also makes the following claim for the superiority of the algorithm diet, that I’m not sure they actually can back up:
However, the predictor-based approach has broader applicability since it can predict PPGRs to arbitrary unseen meals, whereas the “expert”-based approach will always require CGM measurements of the meals it prescribes.
Since the two diets performed “comparably” in the tests they did, I don’t see how they can know this, and they don’t explain further. I guess to the extent that “experts” refuse to predict glucose response to the foods they don’t have data for because they know they can’t have confidence in the result, while the algorithm plunges ahead anyway, that’s trivially true? They’d need to actually introduce new, untested foods and see how their algorithm performs.
Another thing the articles I read do not make clear is that the algorithm to come up with a personalized diet does NOT simply require a stool sample (to get microbiome data) and some other profile data from the participant. It also requires the same time period of analysis with a continuous glucose monitor (CGM)–given that this is the case, I’m very skeptical of claims that it can predict response to new foods.
Participants then underwent the same 1-week profiling of our main 800-person cohort (except that they consumed the meals compiled by the dietitian), thus providing the inputs (microbiome, blood parameters, CGM, etc.) that our algorithm needs for predicting their PPGRs.
In a way I’m NOT surprised–it seemed pretty amazing that they could come up with a personalized diet without testing subjects’ responses to a bunch of different foods. But it makes it so much less amazing that I start to wonder what the point is of collecting microbiome and profile data–couldn’t they make an algorithm based on type of food, amount of food, and blood sugar response? How much does that other data really help them make such an algorithm, vs explaining WHY they react the way they do? Since the “expert” diet performed “comparably”, and that was based on just meal data and blood sugar data, it seems likely that it didn’t help that much for the specific purpose of coming up with a personalized diet–though it did help with some of the other stuff they found out. Was the microbiome data really to see what bacteria are associated with high/low blood sugar, and to fuel their other work in future studies? (This lab has had some other highly publicized gut microbiome studies.)
I first found out about this study via the Skepchick blog, but while the Skepchick post corrected some misinformation, it also contained some out-of-date information or misinformation of its own [caveat: this is based on the transcript; I didn’t watch the video]. I’m mostly going through the quotes in order, but the first inaccurate one takes the longest to rebut, so I saved that one for last.
The food you eat has a huge impact on your blood sugar, which is why people with diabetes need to know how their blood sugar will change when they eat certain foods. Past research has led us to create the glycemic index, a measure of just that. White bread, for instance, is considered a high GI food because it causes the blood sugar to spike. Kidney beans would be an example of a low GI food.
Past research, yup. Glycemic index offers a very incomplete picture of how the food will affect your blood sugar; glycemic load takes into account both the glycemic index of foods and the total amount of carbs, but of the two, what’s most important by far for diabetics to know is the total amount of carbs, and carb counting/carb exchanges are the currently favored tool. The scientists behind this study describe carbohydrate counting as the “current gold standard” for predicting the impact of food on your blood sugar.
This new research turns that knowledge around, though, showing that different foods can have dramatically different results on individuals’ blood sugar levels.
It is true that previous studies were smaller, but this is a large-scale confirmation of that, not unexpected new knowledge. See, for example this quote from Dr. David Ludwig from a Science Magazine article on the study: “We’ve known for decades that one person will experience a much greater rise in blood sugar than another for many reasons unique to the individual. This provides interesting quantitative data about the individual factors, but the basic point isn’t new.” There are basically two new parts–that gut bacteria both influence and are influenced by blood sugar, and that they’ve created an algorithm that can predict people’s blood sugar response to foods and create an individually tailored diet that performs as well as a personalized nutrition plan created by a dietician. The authors acknowledge that it’s long been known that different people respond to the same food differently in the study’s introduction: “However, the few small-scale (n = 23–40) studies that examined interpersonal differences in PPGRs found high variability in the response of different people to the same food (Vega-López et al., 2007, Vrolix and Mensink, 2010), but the factors underlying this variability have not been systematically studied.”]
Sure enough, different gut bacteria reacted in different ways to the various foods, to the point that some people had a bigger blood sugar increase after eating sushi compared to eating ice cream.
Why is this so surprising? Sushi has kind of a health halo and ice cream decidedly does not, but sushi typically contains white rice, which is quite high in both glycemic index and glycemic load. Both ice cream and sushi contain fat and protein, which can lower glycemic response. Perhaps it seems surprising to some people because of the health halo, perhaps because of the idea you still sometimes see floating around that “simple carbs” (table sugar, honey, corn syrup, etc.) are always significantly worse than “complex carbs” for diabetics? (Now that white/refined carbs are widely understood to be “bad carbs”, I sometimes see foods such as white bread or white rice referred to as “simple” carbs, but words mean things, and that is not what that means.)
The scientists who conducted this study were perhaps not so surprised by this: “For foods with a published glycemic index, our population-average PPGRs agreed with published values (R = 0.69, p < 0.0005), further supporting our data (Table S1). For example, the average PPGR to rice and potatoes was relatively high, whereas that for ice cream, beer, and dark chocolate was relatively low, in agreement with published data”.
“The good news is that this means that in the future, doctors will be better able to personalize diet recommendations for diabetic and pre-diabetic patients. The bad news is that those patients are probably going to have to poop in a jar to make that happen.”
Eh… Yes and no. They (or rather, dieticians) can personalize diet recommendations now, by having patients keep a food journal along with checking their blood sugar before and after they begin eating, in order to see if a type of food affects them out of proportion to the amount of carbs it contains. (“Better” is not clear here, since the algorithm didn’t result in a significantly different outcome from the personalized “expert” diets.) But they don’t typically have a continuous glucose monitor helping them, like in the study–most non-insulin-dependent diabetics and prediabetics are lucky if their insurance pays for more than one test strip per day. So while collecting enough data to find out which foods spike your blood sugar can be a challenge, if you know that whole grains spike your blood sugar, and lactose and fructose just don’t seem to affect it much, you don’t need to wait to find out what balance of gut bacteria is causing it, you can go ahead and put more milk and fruit and less oats in your oatmeal.
OK, here’s the long, involved one:
[Type II diabetes is caused] by your cells not responding to insulin correctly because of obesity, inactivity, poor diet, and a bit of genetics.
This really downplays the role of genetics, to the point that it’s not just a difference in emphasis, it’s incorrect. The ADA says
Type 2 diabetes has a stronger link to family history and lineage than type 1, although it too depends on environmental factors.
Studies of twins have shown that genetics play a very strong role in the development of type 2 diabetes.
Lifestyle also influences the development of type 2 diabetes. Obesity tends to run in families, and families tend to have similar eating and exercise habits.
So yes, eating and exercise (and sleep, stress, and probably gut bacteria) affect development of Type II diabetes, but genetics “play a very strong role”–it’s not “a bit of genetics”.
Most people with BMI >25 will never develop diabetes, and some thin people will get it, according to an ADA article about diabetes myths that many people believe. A doctor interviewed by the Wall Street Journal is more specific, saying 20% of thin people will get diabetes and 75% of BMI>30 people won’t–i.e. 20% of thin people vs 25% of fat people. (Though I’ve also seen in other places that the overall prevalence of diabetes is 11%–I guess that is due to the difference between “has diabetes now” and “will ever get diabetes”. I don’t know if there are any numbers for how much earlier, on average, fat people might get diabetes than thin people–delaying diabetes is certainly beneficial, but it also doesn’t fit well with the narrative of “if you just do the right things, you will NEVER GET diabetes. And there are certainly things you can do from a HAES perspective to delay diabetes.) Yet according to the ADA, over half of Americans believe that if you are overweight or obese you will eventually get diabetes. Knowing that most fat people won’t get diabetes, and some thin people do, sure does interfere with just world theory protecting you, though. (Hat tip: I first became aware of this through these posts by Living400lbs.)
Watson’s summary also omits a huge factor in the development of Type II diabetes that more than half of Americans are unaware of: age.
Some of this confusion surely comes from the fact that the rise in BMIs and the rise in rates of diagnosis for diabetes are often talked about in the same breath, and it is strongly emphasized in health articles that having a high BMI increases your risk of diabetes. But I think some of it also comes specifically from the frequently-cited statistic that ~80% of Type II diabetics are “overweight or obese” in the US. (Keep in mind that ~70% of US adults are “overweight or obese”.) (Caveat: the percentage of BMI>25 diabetics has changed over the years, as overall population BMIs have changed, so I think it might be closer to 85%, but 80% is the number I usually hear in popular articles.) People get confused and think that means that 80% of people with BMI>25 will get diabetes.
[Interestingly, although Type I diabetes is thought of popularly as a genetic disease, environment actually has a large role in its development, though how is poorly understood. Still, if one identical twin gets Type I diabetes, the other twin has a 50% chance of also getting Type I.]
As for the “poor diet” bit:
The Harvard School of Public Health says, “Numerous epidemiologic studies have shown a positive association between higher dietary glycemic index and increased risk of type 2 diabetes and coronary heart disease.” So the ‘poor diet’, to the extent that it means specifically higher glycemic index, has evidence behind it. (A “good” diet with a 90s-style low-fat emphasis might also be a fairly high glycemic index diet.)
Inactivity does affect your insulin sensitivity, though again, it’s not enough to cause diabetes without some degree of genetic vulnerability. No doubt our current environment has a significant role in the current prevalence of diabetes, including its influence on the oft-discussed diet and exercise (but don’t forget its influence on sleep!), and sure, there are are things you can do to lower your A1C if it is creeping up towards diabetic levels, but genetics load the gun.
The quote from the Harvard School of Public health goes on to say, “However, the relationship between glycemic index and body weight is less well studied and remains controversial.” So Watson is on firmer ground when she says that this study doesn’t really have anything to say about body weight–the main point of her post. I think that the focus on body weight in the press may have come from previous studies that linked gut bacteria and body weight. The study’s introduction mentions one: “Pioneering work by Jeffrey Gordon and colleagues previously showed that it associates with the propensity for obesity and its complications, and later works also demonstrated associations with glucose intolerance, TIIDM, hyperlipidemia, and insulin resistance”. The results section also notes that several bacteria previously associated with BMI>30 were also associated with increased blood sugar response in this study. So the study does allude to the association between certain gut bacteria and body weight, but that is not the focus of this study and it did not generate new information about that.