Which seems like a suitable metaphor for a field test of an experimental algorithm that doesn’t require mealtime bolusing and uses stepcount to mediate how aggressively the algorithm works.
Boost version 4.1.3, with some fairly big changes and all the stepcount goodies was getting out through the rigour of numerous half marathons, random food, cider, wine and heat. How would it fair?
Turns out that the answer to that is “pretty well, actually, thanks for asking!”.
With plenty of walking and dancing (well jumping up and down on the spot), which averaged more than 20,000 steps per day, we saw the below figures.
I take these as not at all bad, but they are a little bit of a lie.
Friday and Saturday looked like this (firstly time in tight range and secondly time in range):
Sunday, however, looked a lot more like this:
Sunday started off nicely, and then my first pump site change didn’t go as planned. Much manual bolusing ensued as it became apparent something was wrong, and I was trying to have a meal in one of the few places that you pre-book and sit down at, so I wasn’t stopping on that.
As a result Sunday had the extreme highs and resounding lows of a day with a knackered set (it had gone into a blood vessel). It also demonstrated where attempts at fully automated systems hit blockers (and potentially issues if you’re using AI).
Still, can’t be helped.
Additional steps adjustment
As I’ve talked about previously, activities affect my sensitivity. Historically I’ve only really looked at resistance training though.
Glastonbury this time enabled me to identify and additional pattern that could be rolled into APS systems, and lends itself to machine learning spectacularly well.
As discussed in one of the many discord channels, after a period with higher than average aerobic exercise, many people become more sensitive, and some people have developed models to try and factor that in to adjustments.
What I noticed this time around was that effectively doubling my daily steps resulted in a reduction in profile needs of around 15%.
Given we have the tools to take this into account, it provides a way for a manually set variable, and potentially something that might fall into AAPS automations, allowing profile changes based on previous periods’ step counts. Something for those of us who are playing with the code to consider.
It’s also something that, with the use of Machine Learning, we might find very effective.
Compared to last year, the evolution that I’ve made in Boost made it much easier. Although, overall, the numbers appear to be more or less the same, the big difference over the full five days was a lot less time spent low, and this time around it being clearly down to managing a failed cannula and forgetting to reduce my profile after walking 13 miles. The latter a rookie error.
In general then, the step adjustment does seem to help when a lot of steps are involved, and I was very pleased with the outcomes!