Thursday, June 11, 2015

My position on releasing a third party open source Libre app.

I keep getting messages (about twice a week on average, sometimes more) about releasing or eventually helping develop an eventual open source Libre application (or even a closed source commercial one). At this point, I will neither do it, nor provide direct explicit assistance.

Let's recap the technical side: Abbott uses a custom but simple TI chip. What they are doing is conceptually not far from what a sample humidity/temp sensor would do, except for the fact that they are "driving" their sensor. The data provided is, leaving aside the standard signal processing done on the TI chip, as close as we'll ever get to pure raw data and therefore quite different in nature from the cooked secondary raw data the Dexcom provides. That means that a clean application will have to replicate every process Abbott applies to the data and every behavior deriving implicitly from any model they might use.

I posted this a while ago. This is the result of one of several pseudo-CGM runs I logged and interpreted compared to the official interpretation of the minute by minute data delivered by the sensor.
"Abbot" is the official interpretation of the data.
"public" was a direct interpretation based on what was publicly available at the time.
"private" was what I would call a method based on 'constrained linear extrapolation'.
"private experimental" was a method based on a more sophisticated model, not unlike what is found in the literature.

This looks nice and has consistently looked nice when the Libre worked well. It did blow on occasions, especially when the Libre itself was asking for a few minutes of respite. It also blew on one occasion on what seemed to be a NFC read occurring during a sample write to the FRAM. It blew consistently in temperature change situations (but so does the Libre on occasions).

So, why not release it and bask in the 'glory' of having released the first Libre application?

Several reasons.

  1. technically, on a custom micro-controller, I will never be able to interpret every 'flag' with certainty. We are talking bits here and bits potentially have wildly different meanings in different locations. While a "sampling complete" posted flag is fairly standard, there is always the potential that one rarer flag will be missed or misinterpreted.
  2. any interpretative model makes assumptions: published derivative based models are very sensitive to noise in the signal for example. Imagine that a single data point is a bit funky, such as data point 3 in the sequence 142 - 152 - 146 - 168. BG is clearly rising but, for some reason, point 3 was under measured. If you use simple numeric differentiation, you'll end up with +10, -6, +22. Relying on +10 -6 may lead you to decide the situation is stable or rising slightly. Relying on +10 and +22 can lead you to predict a very sharp increase. Data point selection matters: you can't miss an error because of a technical issue (point 1) and you can't use an unvalidated model just because it happens to work most of the time.
  3. it would be in the hands of real people  - now, let's be politically incorrect here. Real people means everyone including people on the extreme left of the median in terms of cognitive abilities. (yes, I weaseled out of that one as those are unlikely to understand what I just said and be vexed by it). Blowing up from time to time on concerned and involved people who understand the limitations is OK. Blowing up on others can be catastrophic, especially if you don't blow up 99% of the time and instill fake confidence over time.
  4. Abbott is a big aggressive company. You only need to read some specialized boards to realize there is a wide gamut of opinions even among Abbott's employees (or people claiming to be). I've had a few contacts, direct and indirect with them, mostly on the confidentiality issues and have seen/read uninformed lies, informed lies and open, honest truths. The Abbott patents are so long and wide ranging you can't be sure lifting your right finger while scanning your ISIG is legal. Some documents of the patent disputes between Dexcom and Abbott are available on the net and the details discussed just blew me away. You get the feeling that you need more highly specialized lawyers to argue for months about irrelevant points than you need MSP 430 developers, MDs to develop a sensor from scratch. Releasing anything some lawyer inside Abbott thinks is protected IP _could_ mean trouble. That's why I gave links to published papers and patents on the blog. To some extent, they help. Going into deeper details might be an issue. Dexcom has been, on the whole, extremely tolerant of the Nightscout project. Abbott may behave totally differently.
  5. I don't personally care about a phone application, I care about a full CGM application and convenient data collection adds another layer of obstacles.
This being said, investigating the Libre is fun and, if you take the time to dig deeper into the details, you learn a tremendous amount of useful things!

No comments:

Post a Comment