It’s been two weeks since changing over to the DanaRS pump and CamAPS to give them both a run for their money and see how I got on, and in doing so, there are some very clear learning points that have come out of it for me, and potentially for anyone else thinking about changing to a commercial system having used a DIY one.
1. Follow your own rules
First up, whenever someone moves to a DIY system, we always tell them to familiarise themselves with the pump (and pumping) and get that set up correctly before looping. I’m guilty here of not following this rule, jumping onto the DanaRS and CamAPS and getting some, well, not necessarily great results.
With issues with various cannulas coming to light over the past two weeks, it wasn’t particularly sensible to switch straight into the Auto-mode setting. As a result, I suspect that the repeated highs occurring at roughly the same time caused by issues with the steel cannulas not remaining well set overnight may have caused the algorithm learning capacity some confusion, and most likely delayed it’s capacity to learn my activity. Under normal circumstances, one off issues wouldn’t have an effect, but unfortunately what happened was a pattern.
What’s the takeaway here?
Before switching on any kind of Automated Insulin Delivery (AID) algorithm, DIY or commercial, especially one with learning characteristics, make sure you’ve figured out all aspects of the pump and more importantly, its sets and have them working smoothly.
2. The algorithm isn’t just learning about you, you’re learning about it
As the CamAPS team will tell you, over the first few weeks of use, the algorithm learns about you. It figures out whether you over or under bolus for meals, for example, and if your pump entered ISF and ICR are vaguely accurate, then corrects for your normal habits.
But as a DIYAPS system user, I had built up habits that, in order for me to not end up with wildly fluctuating glucose levels or prolonged post-prandial highs, I had to unwind, so that the algorithm and I could meet somewhere in the middle.
The key things here for me are around breakfast and lunchtime, and how I have dealt with those using OpenAPS. For a good couple of years, I’ve relied on the ability to automate “Eating Soon” and the availability of the “UnAnnounced Meals” combines with the SuperMicro Bolus functionality to manage around meals.
CamAPS (while it is learning at least – I don’t know whether it is able to automate some sort of “Eating Soon” extended bolus – I suspect not) doesn’t have this pre-emptive adjustment of insulin ahead of a meal (at least not while it’s learning from what I can tell), and doesn’t bolus, so you have to revert to the more traditional model or pre-bolusing. It took me far longer than I would have liked to figure this out, but for people coming off MDI or standard pump use, this would be standard and expected behaviour.
As a result, I had issues with post-prandial highs that frustrated me, but moving back to a more traditional bolusing model has seen this fall away and post-prandial highs have been significantly reduced.
This highlights the points being made in one of the sessions at ATTD 2020 relating to using any sort of AID system, DIY or commercial. It’s not just turn on and go, there are probably changes required from a user perspective that shouldn’t be understated.
Timing of announcing the carbohydrates doesn’t seem to have made a huge difference either, but maybe we’ll see more about that over the next couple of weeks.
3. It’s the physical bits that screw things up
I’ve already mentioned the issues with cannulas and it taking me a bit of time to have them functioning in a way that they worked properly for me.
The other physical thing is the Dexcom sensors, and after the sensor change this time around, it flatlined for around six hours for me, which wasn’t all that helpful. Fortunately, it recovered to a reasonable level and I didn’t lose the sensor, but it was a valuable reminder of where the issues lie in all AID systems.
It also means that the summary data (captured off receiver into Nightscout) for the last two weeks contains 72 points that it considers to be lows below 3mmol/l that are simply incorrect data.
Ah well, these things happen, and are, essentially, the key issue with anything physical. It’s the bit that’s hardest to get right.
The first two weeks overall
Given all that I’ve written above, you might be forgiven for expecting that CamAPS might not have done too well. But that would not be true.
If we look at the aggregate Time in Range data for the period, we see that it did pretty well, given the challenges thrown at it:
78.4% isn’t bad for a system that’s learning about the user as it goes along, whilst in closed loop, and the reality is likely to be that that is likely to be 80.3%, given the flatline that occurred at the Dexcom change. Likewise, the percentage time low, adjusted for that error is around 7.8%.
Whilst this is too much time low, a major part of that can probably be explained from manual intervention in relation to the set issues, and it’s not a true reflection of the algorithm in action.
And looking at it from the perspective of someone who fits into the average Hba1C in most of the world, and was using a pump prior to CamAPS, they’d consider this a huge improvement during the two weeks of starting up with the system.
To put that into context of my normal life, the last two weeks of February using OpenAPS as normal showed the following outcomes:
Why have I taken the last two weeks of February here for comparison? Because during early March I was testing some personal code that is not in the public domain, so it’s not really a fair comparison.
The key point here is that while TIR was 86.9%, the amount of lows during the equivalent period was around half that of the first two weeks with CamAPS. As I’ve already mentioned though, much of the variance can be attributed to me trying (badly) to deal with set issues.
What’s perhaps more important is that the system appears to have settled down now, so the next two weeks are the ones to watch, and I look forward to seeing the outcomes.