Over the past couple of weeks, I’ve been very excited to be using the HAPP open loop artificial pancreas system as derived from OpenAPS. It’s essentially the OpenAPS code ported to run on an Android phone (with much of the work done by Tim Omer), and we are working on trying to get it talking to various pumps. It was the main reason I wanted to get the xDrip platform up and running.
It’s been an interesting and enlightening experience handing control of my insulin over to an algorithm (all be it with my agreement on its decisions) but the outcome of it has rather shocked (and pleased) me.
It all works very simply. You tell it your high and low limits, and your target glucose level. You then give it your pump basal rates over 24 hours, your Insulin Sensitivity Factor, Insulin Carb ratio, Insulin Duration and a carb absorption rate (a variable that needs a bit of work) and then it used the data from the CGM to tell you how to adjust your basal, if you need additional bolus, etc. This is the product that Tim Omer has been seen in the news talking about. I’ve got it all linked up to my Nightscout implementation so that I can generate a decent set of reports to see how it is all working.
You can limit the maximum basal you allow it to apply and really you should configure your pump to have a maximum temporary basal in line with what is set in HAPP. I find that I end up setting Extended Boluses a lot because I can’t set my temporary rate high enough. It also tend to correct with higher basal rates rather than using bolus corrections.
This is what it looks like. There’s a lot of information on the screen, but essentially you have current CGM data at the bottom, with predictions showing, a number of note items in middle section and then a breakdown of blood glucose level, change since last reading, anticipated result in 30 mins and Carbs and Insulin on board.
The results from the first two weeks of using it have been really interesting. Below are my AGP, hourly stats and distribution. This hasn’t been a particularly low carb period and has included a day where I ate Pizza, Hummingbird cupcake, and various cookies. These numbers, in light of that, are really quite remarkable! The target range being shown is aligned with that of the studies done by UK HCPs and ranges from 4.4 to 10. I have to admit I’m really impressed. The biggest gain being in the narrowness of the 10-90th range and the resultant standard deviation falling below the 33% of median, let alone mean!
All in all, with the experience of the last couple of weeks, I’m going to keep on using it. It’s been really impressive and has provided some interesting insight into managing myself.
One of the most difficult parts of this experience has been trusting that the algorithm knows what it is doing. The learning experience is really that it does everything far more slowly and calmly than I would, and instead of hammering in a rage bolus to bring down a climbing blood glucose, suggests a significant hike in basal to account for it. It means the drop out is a much slower incline. In spite of that, its prediction capability is surprisingly good and as you can see, it’s done a remarkably good job!
I’ll update further as I use it more, and hopefully, when we finally get a closed loop version up and running. If you’ve any questions, feel free to comment and ask!