- The carbohydrate content of a food alone does not accurately predict insulin response. Protein and fibre content of food also influence in insulin response.
- The food insulin index data indicates that dietary fat is the one macronutrient that does not does not require a significant amount of insulin.
- Net carbohydrates plus approximately half protein correlates well with observed insulin response.
- This knowledge can be used to help select low insulin foods and more accurately calculate insulin doses for diabetics.
Back before the GFC I used to dabble in share trading. I don’t know much about financial systems, but I spent a good deal of time designing and testing “trend following” trading systems.
One of the pitfalls for newbies is to design a system with excessive “curve fitting”. That is, to design a complex system that would work fantastically on a specific set of historical data. If you ran an overly curve fitted system on another set of data or tried to trade it in real time it would fail because it was too finely tuned to the discrete set of historical data.
“Everything should be as simple as possible, but no simpler.”
Another lesson from trading is that you should be able to describe simply why a good system works. My trading system scanned the market for stocks that were moving up quickly over a number of time periods with minimal volatility so that I could place a close ‘stop loss’ that would take me out of the trade quickly if the trend turned.
When the GFC hit things got too volatile and I got out of the market. It was no longer fun. However the skills I learned as an amatuer a quantitative trader (along with my day job running multi criteria analyses to identify motorway alignments, road investments and the like) have given me an interesting angle on nutrition that I hope people find useful.
On the Optimising Nutrition blog I have tried to describe a system to manage nutrition that makes sense to me. I want to document the things that I wish someone had shown us when we started out trying to understand diabetes and nutrition.
If we want to understand and predict the behaviour of insulin, the master regulator hormone of the human body, we need to first determine what we know that is accurate, significant and useful that we can use.
Kirstine Bell’s PhD thesis Clinical Application of the Food Insulin Index to Diabetes Mellitus (Sept 2014) details the results of the latest food insulin test data for more than one hundred foods. It also evaluates the relationship between insulin demand and protein, fat, carbohydrates, glycaemic index, glycaemic load, indigestible fibre, individual amino acids and blood glucose.
Previously I have discussed in a moderate amount of detail how to calculate how much insulin may be required based on the carbohydrate, protein and fibre ingested. Given the importance of this issue, this article looks in more detail at what can be learned from the test data included in this thesis about the relationship between these parameters, with a view to better manage blood glucose and insulin demand. You will see that I have tried to look at the issue from a number of different directions and have also included a more rigorous statistical analysis.
Most people know that carbohydrates require insulin. As shown in the chart below, carbohydrates goes some way to explaining insulin response. However it is far from a perfect relationship (R2 = 0.44, r = 0.67, p < 0.05).
Taking indigestible fibre into account (i.e. net carbohydrates) improves the relationship (R2 = 0.48, r = 0.69, p < 0.05). The best correlation is achieved when we subtract all the indigestible fibre from the total carbohydrate value. However we can see from the cluster of data points on the vertical axis there is something going on that is not explained by carbohydrates alone.
The importance of dietary fibre should not be discounted, especially when trying to reduce insulin demand. Some recommend that diabetics limit total carbohydrates, rather than considering net carbohydrates, or non-fibre carbohydrates. The danger with a total carbohydrates approach is that people will avoid fibrous non-starchy vegetables that provide vitamins and minerals that cannot be obtained from other foods (unless you’re consuming a significant amount of organ meats), as well as feeding the gut bacteria which is also important to help improve insulin sensitivity and the body’s ability to digest fats. 
The food insulin index data indicates that foods that are largely comprised of fat have a negligible insulin response (R2 = 0.38, r = 0.631, p < 0.001).
To put this another way, the chart below shows the sum of carbohydrate plus protein (i.e. the non-fat content of foods) versus the insulin index (R2 = 0.38, r = 0.62, p < 0.001) indicating that:
- the greater the proportion of fat in a particular food the less insulin is required; and
- the more carbohydrates and / or protein ingested the more insulin is required.
Hence, it appears that to reduce insulin demand we need to reduce carbs and / or protein!
The figure below shows a similar chart for the glucose score (i.e. the area under the curve of the blood glucose rise over three hours after ingestion of the food). Again, this indicates that the blood glucose response is lowest for foods that contain a higher proportion of calories from fat (R2 = 0.45, r = 0.68, p < 0.001).
While it appears that insulin demand is triggered by carbohydrates and protein, what is not clear is the relative degree to which carbohydrates and protein contribute to insulin demand. Are they equivalent or does protein cause a smaller insulin response?
Another observation from trading is that you can learn a lot by considering outliers. You have to decide whether the data points that don’t quite fit the trend are garbage or ‘black swans’ need to be accounted for in the system.
In the carbohydrate vs insulin relationship the outliers are the high protein foods that trigger a higher insulin response than can be explained by considering carbohydrates alone.
As shown in this plot, high protein foods are typically lower in carbohydrates which produce the greatest amount of glucose. Choosing higher protein foods will generally reduce insulin (R2 = 0.10, r = 0.47, p < 0.001).
Increasing protein will also typically lead to a spontaneous reduction in intake due to the thermic and satiety effects of protein.   Protein is critically important for many bodily functions. It is vital to eat adequate protein.
However protein in excess of the body’s needs for growth and repair can be converted to glucose. The fact that protein can turn to glucose represents a potential ‘hack’ for diabetics trying to manage their blood glucose as they can get the glucose required for brain function without spiking blood glucose as much as carbohydrates.
Choosing higher protein foods will generally lead to better blood glucose control. Although high protein foods still raise the blood glucose somewhat, particularly if you are not insulin sensitive, however the blood glucose response is gentler and hence the pancreas can secrete enough insulin to balance blood glucose.
For most people, transitioning to a reduced carbohydrate whole foods diet will give them most of the results they are after. However for people with Type 1 Diabetes or people trying to design a therapeutic ketogenic diet, consideration of protein may be important to further refine the process to achieve the desired outcomes.
For a healthy bodybuilder the glucogenic and insulinogenic effect of protein might be an anabolic advantage, with the post workout protein shake providing an insulin spike to help build muscle.
However for someone struggling to lose weight on a low carb diet, considering the insulinogenic effect of protein might just be what they need to reduce insulin and normalise blood sugars and thus enable them to reach their goals.
The glycaemic index is a reasonable predictor of insulin demand in terms of correlation (R2 = 0.54, r = 73, p < 0.01), however the ‘elephant in the room’ again is the high protein low carbohydrate foods (e.g. white fish, low fat cheese, lean beef etc).
The other issue is that the glycaemic index is an empirical measurement that has to be measured in humans “in vivo” and can’t easily be calculated based on commonly available food properties. And again, the glycaemic index does not deal with the insulin response from high protein foods.
The same issues apply to glycaemic load. There is a reasonable correlation between glycaemic load and insulin demand. However it still does not explain the insulin effect of high protein foods (R2 = 0.57, r = 0.75, p < 0.01). And you have to run these tests in real people “in vivo”.
Like the food insulin index, the glucose score is measured “in vivo” based on the area under the curve of a healthy person’s glucose rise due to a particular food.
Glucose score is interesting in that it actually achieves an excellent correlation with insulin demand (R2 = 0.75, r = 0.87, p < 0.001), however there is still a disconnect when it comes to high protein foods.
It seems that some foods that do not raise blood glucose significantly over three hours still elicit an insulin response. High protein foods digest slowly although they do still require insulin to metabolise. In a normal healthy person the body’s insulin response to protein is balanced by release of glycogen from the liver, with blood glucose being kept in balance by insulin and glycogen. 
In a normal person the insulin keeps up with this slow blood glucose rise and hence we do not see a pronounced blood glucose spike due to high protein foods.
The interesting outliers here are processed low fat milk products that seem to require more insulin than would be anticipated by the blood glucose response. On the other side of the trend line we have brown rice, pasta and other less processed whole foods which raises the blood glucose but does not require as much insulin as might be expected.
Accounting for fibre (i.e. net carbs rather than total carbs) goes some way to help anticipate the effect of processing. However the effect of processed foods is an interesting area for future study that is beyond the capacity of this dataset to address.
I ran a number of correlation analysis and could not find an explanation of why a certain food sat above or below the trend line, whether it be carbohydrates, sugar, fibre or protein.
The sugar content of a food is not a particularly useful predictor of insulin demand (R2 = 0.10, r = 0.32, p = 0.001) compared with net carbohydrates (R2 = 0.48, r = 0.69, p < 0.05). Quitting sugar is only part of the solution. Most people struggling with diabetes or obesity should ideally consider their total carbohydrate intake.
Kirstine Bells’ Clinical Application of the Food Insulin Index to Diabetes Mellitus documents the development of a number of formula to explain the relationship between food properties and the food insulin index response. The aim of this her thesis was essentially to build an improved glycemic index to predict insulin response rather than only considering changes in blood glucose.
The chart below shows the best relationship developed using a stepwise multiple linear regression analysis of the various parameters to forecast insulin demand documented in Clinical Application of the Food Insulin Index to Diabetes Mellitus. 
The correlation is excellent (R2 = 0.78, r = 0.89, p < 0.001). However this relationship relies heavily on the glucose score (GS) which has to be tested “in vivo”.
If we strip out the glucose score then the best relationship achieved in the thesis is the one shown below using carbohydrates and protein with a correction factor (R2 = 0.46, r = 0.68, p < 0.001).
The problem with this approach is that it assumes that high fat foods have some insulinogenic effect. However we have seen above that high fat foods have a negligible insulin response. This formula also does not account for indigestible fibre which should be subtracted from the total carbohydrate count. And according to this formula a food with zero carbohydrate and zero protein would still have a significant insulin index response of 10.4, which does not make sense.
simple is true
If we take out indigestible fibre (net carbs), assume that fat has a negligible insulin response and refine the protein factor to maximise the correlation with the test data, we end up with this chart which has an improved correlation compared to the model above (R2 = 0.49, r = 0.70, p < 0.001).
This approach also does a good job of predicting blood glucose (R2 = 0.59, r = 0.77, p < 0.001) as shown in the chart below.
Individual foods can be ranked and prioritised based on their proportion of insulinogenic calories using the following formula:
Foods with the lowest proportion of insulinogenic calories will have the gentlest impact on blood glucose and have the lowest insulin demand, a consideration which will be very useful for people who are insulin resistant (i.e. Type 2 Diabetes or Pre-Diabetes) or not able to produce adequate insulin themselves (i.e. Type 1 Diabetes).
Diabetics and people wanting to reduce the insulin demand of their diet can track the total insulin load (as opposed to carbohydrate counting) using the following formula:
The total insulin load can be reduced by decreasing carbohydrates, increasing fibre, moderating protein to the body’s optimum requirement and increasing fat until target blood glucose are achieved.
can we design a “perfect” system?
There is still quite a degree of in this real life data. This could be due to measurement error in the macronutrients, food quantity, individual characteristics of the people that the food was tested on, or something else.
This approach considering the insulinogenic effect of protein and carbohydrates does however help to better predict insulin demand than carbohydrate alone.
The fact that there is still a high degree of variability in the data and hence limited ability to accurately predict the insulin response to food can be mitigated by keeping the overall insulin load of the diet reasonably low.
Dr Richard Bernstein talks about the ‘law of small numbers’ whereby the compounding errors in the calculation of insulin requirement and the mismatch of insulin response with the rate of digestion misalign means that it is impossible to accurately calculate insulin dose.
The only way to manage the high level of variability is to reduce insulin demand to manageable levels. This is especially beneficial for people who are injecting insulin, but also relevant for the rest of us.
Building on the analysis of the food insulin index data, the key assumptions that underpin this system are:
- carbohydrates require insulin,
- indigestible fibre does not require insulin, and
- the glucogenic portion of protein that is not used for growth and repair and not lost in digestion also requires insulin.
In order to reduce our insulin load we should do the following, in order of priority:
- Reduce insulin load until you normalise blood glucose levels (i.e. reduce digestible carbohydrates and moderate protein if necessary),
- Increase nutrient density as much as you can while still maintaining good blood glucose levels (note: this will likely also include fibre from non-starchy veggies which will also increase fibre which reduces insulin and slows digestion),
- Reduce dietary fat if you still need to reduce body fat levels, and
- Implement an intermittent fasting routine to improve your insulin sensitivity and to kick-start ketosis.