Rotate your injection or infusion sites. That’s a key message that anyone living with type 1 is delivered very early in their life with the condition. And for good reason. Nobody wants to have to live with scar tissue or lipohypertrophy affecting their insulin absorption.
But within the advice about location lies another reality, before we get to any tissue damage. Different locations on the body absorb insulin at different rates. In fact, the variation can be quite substantial. Here, we dig into how that differs and more importantly, what that means when using an AID system.
Absorption variation by location
Many papers have been written that discuss variation in absorption, but perhaps one of the most useful is this one.
It’s useful because it breaks out the effects of administering insulin in different locations, both in terms of absorption timing and what that means for the pharmacokinetics.
Why do pharmacokinetics matter? Variation as a result of the site you use might affect the time insulin takes to hit maximum concentration and the maximum concentration, which would affect the “power” of it to reduce glucose levels.
As the document states, with regular human insulin, the Tmax (time to max concentration) in the abdomen can be up to twice as fast as other sites they mention, which includes the thighs and deltoids (upper arms). But we’re not really interested in Regular insulin. We want to know about Rapid-Acting insulin.
Fortunately, Break etc al produced this paper comparing Regular and Lispro insulin back in 1996. The upshot is that they did see differences in absorption between sites, but they were not statistically significant. However, let’s take a look at them.
Firstly, the graph below shows the glucose infusion rate for the three different sites the study used. Now it’s worth mentioning something important about the study. It was a euglycemic clamp study done on “healthy” volunteers, which likely means “without T1D”. Perhaps more importantly, it was a very small study with only 12 volunteers. Which makes variability in the results that much more difficult to manage. The graph below is taken from the paper, and shows the difference in peak glucose infusion rate for the three sites. From this, we infer that there were differences in peak insulin concentration across the three.
The table below shows the max concentration of insulin and the time to that peak. It also shows the 95% confidence interval for duration of insulin action and onset interval.
Location | Difference in Tmax | P | Difference in Cmax | P | Onset Interval | Duration Interval |
(minutes) | (pmol/l-1) | (minutes) | (hours) | |||
Abdomen | 0 | – | 0 | – | 20-40 | 5.25-6.5 |
Arm | 16.2 | 0.612 | -194 | <0.001 | 25-50 | 6.5-8.0 |
Thigh | 12.5 | 0.696 | -131 | 0.005 | 25-50 | 6.25-7.75 |
What’s interesting about this is the p-values. While the differences from the abdominal concentration had very strong p-values, the differences in time to max concentration were very weak, suggesting significant variability in the results.
What’s clear is that there is a lower peak concentration in the arm and thigh, and additionally, using the glucose infusion data, we see that the peak seems to last longer.
Having said that, the difference in time to peak concentration in the arm and thigh is noticeable, and if you think about it, a ten minute difference is the equivalent of the reported difference between Lyumjev and Fiasp, or Fiasp and Novolog, so while it doesn’t seem enormous, we know it can have an impact.
The other point of note in this table is the difference in the duration interval. Both the Arm and Thigh exhibited far longer tails than the use of the abdomen.
The paper also shows that the Area Under the Curve (AUC) is pretty much the same across all three sites, suggesting that the same amount of insulin is being absorbed.
All taken together, we can see that this means that injection or pump use in different locations has somewhat different characteristics.
Applying this data to Automated Insulin Delivery systems
Now that we’ve dug into this data, it seems fairly clear that the differences we see here could have a noticeable effect on the operation of an Automated Insulin Delivery system. Firstly, we need to consider the three different types:
- Machine learning systems
- Systems with set insulin dynamics
- Systems with access to change insulin dynamics
For each of these, the impact of the above will differ.
In a system that uses machine learning to work out insulin duration and peak time based on your own pharmacodynamics, there is a risk that changing site away from the abdomen may throw the calculations out, as it is likely to expect that the insulin will act more quickly and that the duration will be shorter than it is. It will then learn about the changes, but in the interim period there’s a risk that the results are suboptimal.
Similarly, where insulin dynamics are a set configuration, placement of a pod or cannula on your arm after using the abdomen may also throw the calculations out, but in this case, they will always be out, and are likely to lead to other changes being made to the user profile in order to manage the differences.
Finally, in the case where it’s possible to adjust insulin dynamics, what you thought were profile problems may be better fixed by changing the insulin dynamics.
In the two images below, I’ve adjusted the peak and duration to account for the differences described above. The first image shows the “normal” dynamics. The second one is adjusted according to the above table for the arm.
What you can see is that whilst the area under the curve should be more or less the same, there’s a longer, fatter tail with a longer duration and later peak. Additionally, as you’d expect, the peak concentration is lower.
Whilst in the diagrams above these differences don’t look to be all that large, the effects of them would be to reduce the amount of glucose lowering ability early in the bolus lifecycle and also to extend the amount of time that the tail has an effect and the size of that effect. Essentially the difference between the two graphs is likely to be an initial higher post prandial glucose level and subsequent hypo, if the system believed it had image one, when in reality it had image two. If a machine learning algorithm was expecting the first graph and got the second, it may well misinterpret the outcomes to be an incorrect insulin sensitivity or incorrect basal levels and potentially make changes accordingly.
Conclusions
As we said earlier in this article, the data associated with the assumptions in the previous section comes from a small sample size with a statistically insignificant set of data in terms of time to peak, however, it potentially provides another tool to modify configuration of oref1 based systems where you move sites around your body and end up with unexpected results. It might also help where systems are learning about your use and you observe similarly unexpected events associated with site location variation.
Ultimately, the data in this area is fairly sparse, but it doesn’t seem unreasonable that for automated insulin delivery, this maybe an area that’s worth further research. Given the popularity of patch pump based AID systems, their prevalence in the open source world and growth in general diabetes care, and the flexibility they provide, it certainly seems to be something to look into more deeply.
Could switching to Fiasp or similar when using sites away from the stomach help with this?
I don’t think there’s really any good data out there. It would require experimentation.
Also, I’m not sure why you would want a slower action when you were using the abdomen to reduce system changes, when the outcomes are likely to be better with a faster insulin.
I’m tempted to try it if I can get a prescription. I’m currently on Omnipod 5 with Novorapid and whenever I rotate away from my stomach, I have higher sugars in general and more post-meal spikes (and Omnipod 5 isn’t good at keeping you below 7 in the first place).