We’ve just lately seen the Libre2 and Libre3 permitted to be used with AID methods, so I believed it will be an fascinating time to have a look at what occurs in one of many out there apps that purport to take the Libre 1 minute information and make it out there at 5 minute intervals to be used with open supply AID instruments.
Right here we have now information produced by Diabox. The graph reveals the variation of 1 minute information (blue line) after which Diabox’s interpolation to create a 5 minute dataset (orange line). The info has been captured utilizing a mix of Nightscout and xDrip. There’s clearly fairly a dramatic distinction.
That is solely a small dataset, taken from a six hour interval, nevertheless it permits us to see a number of the challenges in utilizing the minute-by-minute information that’s captured in some third get together apps.
The minute-by-minute information seems to point out loads of microspikes whereas the 5 minute information is clearly utilizing some type of averaging or smoothing to get rid of this jumpiness. What’s not clear from any of the documentation is how it’s being carried out.
What information are we really ?
I requested Bubblan to supply some suggestions on the supply of each datasets. The suggestions was that Diabox itself pulls uncooked information, however depending on the way you show it (Libre2 patched app on xDrip or information from Nightscout) you might be more likely to see various things.
The 5 minute information ought to be the identical as one minute at 5 minute samples.
If, nonetheless, you feed it via xDrip, utilizing the patched Libre2 app possibility, xDrip itself smooths the info, which might clarify why the 5 minute information factors don’t match the one minute ones each 5 minutes.
Should you’ve used the official Libre2 app, usually you don’t see the microspike impact evident within the graph, which is one thing that appears to be smoothed by what we’ve come to know within the open supply world because the OOP algorithm. From what I can see, straight out of the field, this isn’t what Diabox does.
That’s to not say that Diabox doesn’t supply smoothing. It very a lot does, with two choices.
Savitzky-Golay makes use of a neighborhood polynomial to determine the match of the info inside the smoothing. Borrowing a picture from Wikipedia, this may be seen in operation:
![](https://i0.wp.com/www.diabettech.com/wp-content/uploads/2023/11/Lissage_sg3_anim28229.gif?resize=610%2C460&ssl=1)
Noise estimate correction I assume is looking again over a sequence of knowledge factors and estimating the noise, then making use of a correction to the subsequent information level.
For minute by minute information, utilizing both of those might be preferable to the smoothing free information proven within the preliminary graph.
5 minute information
The 5 minute information is smoothed by xDrip. The method would have to be confirmed through xdrip documentation . Enterprise fundamental and weighted averaging over a sequence of knowledge factors didn’t reveal something, so I assume this makes use of a extra superior smoothing method. It’s definitely not taking the info level at a cut-off date as the worth it makes use of.
Why can we care?
If you’re utilizing an Automated Insulin Supply system, it’s best to care.
Firstly, realizing what information you’re really utilizing for something open supply might be a good suggestion, and as this reveals, it’s not all the time instantly apparent what that information is. The extra purposes, the higher danger of lowered readability.
Secondly, think about an algorithm making a call based mostly on the microspikes proven within the uncooked one minute information in comparison with the 5 minute smoothed model.
We already know that almost all of economic methods implement their very own smoothing algorithms, and that each AndroidAPS and iAPS have additionally carried out related. This helps to keep away from the misguided increased readings that might lead to overdelivery of insulin.
It’s additionally useful if you happen to’re dosing from an app, for related causes.
The takeaway is that if you happen to’re utilizing any software like this, it’s value contemplating enabling the smoothing performance. It might not offer you precisely what the sensor is studying proper now, however doing so is far much less more likely to lead to by chance delivering an excessive amount of insulin. Which is extra vital to you?