As many of us do going in to the new year, I made a decision to be more active, eat more healthily, etc. The usual “Christmas was a bit big and now so am I” type of decision. As part of the activities part of that decision, I decided on hot yoga, for various reasons. Let’s be honest? Who doesn’t like to spend 90 minutes in a room at a temperature of 41C with 30 other, people sweating profusely and generally exhausting oneself?
Clearly, using an insulin pump and a closed loop in this environment would be a challenge, so for each session, I remove the pump, check glucose before entry and then 90 mins later when I come out, and dose insulin at the end accordingly. This usually causes an increase in my glucose level over the 90 minute period of between 3 and 4 mmol/l (54-72 mg/dl). There’s also a significant heart rate level throughout the session as shown below. The yoga session is the area in the red box.
After doing this the first time, I noticed that the values on my Dexcom G6 post session were signficantly higher than those that the Ascensia Contour Next showed up.
Following further sessions, I determined that the G6 values were showing up as between 3mmol/l (54 mg/dl) and 5mmol/l (90mg/dl) higher than those of the finger prick testing. I also tried the blood tests using different finger stick devices, but the values remained constant.
Now that wouldn’t be great under normal circumstances, but with a hybrid closed loop attached, it obviously presents significantly greater challenges, especially as the data that’s being produced doesn’t appear, looking at the output being received, to be showing any sign of error or noise.
Given this discrepancy, I took to Twitter to see what people thought. An interesting point was made about the operating temperatures of devices, so I thought this warranted investigation. It transpires that the Dexcom G6 transmitter operating temperature range is shown below:
For those living outside the US, that’s 10C to 42C. So while the temperature I’m operating in doing hot yoga is high, it’s not outside this range (class temperatures are 105F/41C) with the room set at 40% relative humidity, although, once you have 30 people sweating into it, I suspect that the humidity is substantially higher. So one would expect that the sensor/transmitter combination should still operate within expectations.
There is also a statement in the user manual, relating to diathermy (electrical heat) treatment that says:
heat could damage the components of the G6, which may cause it to display inaccurate G6 sensor glucose readings (G6 readings)
Which I can understand, but we should be cognisant that the operating temperature in this case is within operational conditions.
As a comparison, we also have the Libre, which has the following characteristics:
Again, we see a “hot” extremity value that’s higher than the class temperature.
Given these operational characteristics, what was happening to the sensor readings in this environment?
Sensor readings compared to blood glucose
Having observed the same thing happening on the Dexcom over 4 different classes, a Libre was attached to my other arm for comparison, to see whether this is a specific issue with Dexcom’s G6 or a wider issue with glucose sensors and the conditions of hot yoga. Or, and this is much harder to corroborate, an issue specific to me. The data is shown below in tabular and graphical form.
Within this image, it’s possible to see that there’s a significant difference in the rise identified by the Dexcom, compared to what is seen in blood testing.
In fact, in this example for the 2mmol/l rise in the blood glucose, there’s a 4.4mmol/l rise in the Dexcom G6 reading, and the CGM reading ends up around 16% higher than the finger prick. This example therefore shows a 120% greater “rise” in the Dexcom reading compared to the rise in the blood reading, and the variation of the glucose reading after the rise is one of the lower ones that I’ve seen.
This then takes a couple of hours after the event to normalise back down in line with the glucose reading.
Now a single example doesn’t make for a trend, however, in nine out of ten hot yoga sessions, I’ve seen this type of difference showing between blood glucose tests at the end of the session, and the Dexcom readings on reconnecting the transmitter to the phone (this is with the system using the Dexcom G6 native algorithm).
It’s also worth noting that the data being output by the Dexcom G6 throughout this period contains “Clean” noise data. In other words, the rise that the G6 is seeing appears completely normal to it. I don’t know whether there is additional data that the G6 can output as part of its iCGM designation, but I’d expect that that wouldn’t necessarily show any major outliers.
Finally, the time taken to return to “normal” seems to vary a bit. In the examples shown on the page, it’s taken about 1.5 hours, however there have been one or two examples that have taken longer. It’s not clear whether this is related to the specific sensor being less well auto-calibrated, and therefore should also be considered as a potential issue.
Freestyle Libre2 (EMEA edition)
Alongside the Dexcom G6, I’m using the Libre2 for a different purpose. But it also allows me to see whether the phenomenon that has been observed with the G6 is also there in the Libre2.
So far, the initial impressions of Libre2 suggest that it suffers less with variation compared to the G6, but still experiences it, and this is patchy data that is only from two sessions.
As there’s no backfill using the patched Librelink app for Libre2, you can see the gap while the hot yoga took place.
Compared to the Dexcom G6, the Libre2 experienced a rise of 3.3mmol/l compared to the blood rising 2mmol/l. In terms of “rise”, the Libre2’s was 60% larger than that of blood. This is also not a great outcome, but is perhaps less of an issue than that seen with the G6. It’s also worth noting that the absolute value was only around 10% above the blood reading.
While this is not a systematic review of a large number of data points, the overall impression here is that the G6 is struggling in conditions that it should be operating normally in, while the Libre2 is also affected, but less so. Whilst admittedly, there aren’t enough data points to provide a fair representation of both sensors, it does beg the question “Why?”. When you look at the two results side by side, you do have to ask this.
In the above image, the straight line shows the yoga period, and when the dots restart, we see the difference between the Libre2 and Dexcom at this point.
When this question was raised on Twitter, a number of theories were put forward, including the metabolic effects of the type and duration of exercise and also the heat involved. There’s also the question of how the G6 and Libre2 differ in terms of their hardware architecture and importantly temperature sensing.
Essentially, the Libre uses a thermistor to check skin temperature and at least one other to check environment temperature, while the Dexcom uses the sensor filament as one part of the temperature check, and may or may not have an air temperature check as well.
It’s possible that these two different approaches result in the variation that is observed. It’s also possible that the differences in sensor construction mean that in circumstances where there is very high vaso-dilation and thus skin temperature, which may be expected in a hot yoga class, the performance of the sensor substrate may vary.
Either way, the difference between the two is obvious:
Whilst this data is only from a small number of points, there seems to be a clearer overcompensation by the Dexcom G6 during the circumstances around hot yoga at 41C. It’s hard to define what’s causing it, but given the temperatures are within the “normal” operating conditions for the G6, it’s not necessarily something that would be expected.
What’s perhaps slightly more concerning is that the sensor is not giving off any indication of “wrongness” of the data, in that there’s nothing in the standard instruction set that suggests that this data is everything other than clean. There may be additional iCGM fields that I am not aware of, but unless there’s some sort of warning that the temperature may be of concern in those settings, I expect that the iCGM output would mirror that of the data that xDrip receives from the native algorithm, and that’s that it is operating normally.
Where this is of concern is in its use with APS systems. If there is no evidence that there are any issues with the data, a commercial APS system (or a DIY one for that matter) will dose insulin based on the data it has. Given that the system has overstated by up to 5 mmol/l in the observations that I’ve made, the effect on dosing that this could have would be quite severe, especially in a model where the algorithm can bolus to deal with highs.
In theory, the higher targets of commercial systems should provide some protection, but even then, a commercial algorithm targeting a value of 6.6mmol/l, with a sensor reading 3-5mmol/l over the real value doesn’t leave a lot of room for error. Yes, if a user is aware of it, they can set a post exercise “exercise target” to ameliorate this, as would a DIY user, but if not, you’d have to hope that the higher settings and shortish duration (1-2hours) of erroneous readings would keep someone safe.
Ultimately, this highlights the other side of technology. There are likely to be unexpected situations where it doesn’t quite work according to plan, and the ability to be aware of what’s going on is still important in dealing with diabetes.
A point that everyone moving onto a technological closed loop needs to be reminded of.