Autotuning Diabetes Settings in the Cloud: Part 2

The AutotuneWeb system described in Part 1 is made up of 4 services in Microsoft Azure working together as illustrated below:

AutotuneWeb Services

The App Service runs the website that users access. Written as an ASP.NET MVC application, the user enters their Nightscout URL and the application:

  1. Extracts the profile information, including the insulin:carb ratio, insulin sensitivity factor and basal rates
  2. Converts that profile information to the format expected by Autotune
  3. Displays the converted profile and the command line required to run it to the user

At this point we’ve covered the initial requirements to simplify access to Autotune. The remaining step is to actually run Autotune automatically and email the results back. To handle this we need a few more services:

The SQL Database stores details of each Autotune job that is requested, including a unique identifier and the email address to send the results back to.

The Storage Account holds the Autotune-format profile information and the results produced by Autotune.

The Batch Account processes Autotune using the profile stored in the Storage Account, places the results in the same Storage Account.

When Autotune has finished running, the App Service is notified by the Batch job. The MVC application then retrieves the Autotune results from the Storage Account and reformats it into a more easily-readable format using the same Razor engine used by the MVC views. The email address to send it to is retrieved from the SQL Database, and the email is finally sent using SendGrid.

Because Autotune is run using Azure Batch, the service can scale quickly based on the amount of jobs that users are requesting at any one time. An initial version that used the single VM I had set up originally to run only my own Autotune commands was quickly overwhelmed when more than 2 or 3 jobs were queued at once. By switching to Azure Batch I eliminated this problem and also reduced cost, as I could use low priority VMs for 20% of the cost of a regular VM. The scaling is handled automatically by a formula on the Batch Account pool that increases the number of VMs when there are more jobs waiting to run and removes VMs all the way down to zero when there is nothing waiting.

Want to run it yourself?

If you have a Nightscout site that you want to run Autotune against, you’re welcome to use AutotuneWeb to do it for you right now.

All the code for AutotuneWeb is available on GitHub if you want to run everything yourself or just see more detail about how any of it works. PRs welcome!

What’s Next?

I used a SQL database originally because that’s the technology I’m most familiar with. However, it’s probably overkill for what is required by this application and could be more efficiently replaced by a Table in the Storage Account. This is my first “cloud-native” application, so please let me know if there’s anything more I could do to make it more “cloudy”!

Autotuning Diabetes Settings in the Cloud: Part 1

Three years ago our first son was diagnosed with Type 1 Diabetes (T1D). Over the next few months we quickly got used to the process of calculating the amount of insulin required for meals and high blood sugar, and adjusting those calculations on a trial-and-error basis. As a software developer I felt sure there should be a better way.

Thanks to following Scott Hanselman on Twitter I stumbled across some mentions of the OpenAPS project – an open-source artificial pancreas system. Simply amazing stuff! The basics of the system are:

  1. A CGM system that monitors your blood sugar and sends a reading every 5 minutes to…
  2. A processing system running on a Raspberry Pi, Intel Edison or similar that runs all the same calculations as we were doing manually, and sending the results to…
  3. An insulin pump that delivers (or stops delivering, depending on the results of the calculations) the required amount of insulin

This seemed great, and as soon as we could get our hands on a compatible insulin pump we got up and running with the AndroidAPS system that uses the same core logic but simplifies the physical setup by using a standard Android phone as the “brain” and bluetooth communications with the CGM and pump.

However, there’s no clever machine-learning-y stuff going on with those calculations, it’s still essentially performing the same calculations as we always were, it’s just doing it for us automatically and reliably. While that’s definitely a step forward, if the ratios (the insulin:carb ratio that determines how much insulin should be required when eating, and the insulin sensitivity factor (ISF) that determines how much insulin should be required to bring down a high blood sugar level) and basal rates (the amount of insulin continuously delivered in the background) are incorrect then the results aren’t going to be as good as we’d hope for.

The process of refining what those settings should be normally seems to be done simply as trial-and-error, which can get very frustrating! Luckily, one part of the OpenAPS project is Autotune, a system that can run on historical data covering blood sugar, carbs and insulin and recommend changes to those ratios and basal rates based on hard data. Definitely much more appealing to my analytical nature!

The problem is, as a Windows and AndroidAPS user, I found it pretty difficult to use. The details of the ratios and basal rates currently in use weren’t stored in the format that Autotune was expecting, and I had to go back to the documentation each time to work out the format it needed and the details of the command line I had to run. I also had to run it on a Linux VM, just making it that bit more difficult and meaning I only ran it infrequently rather than as a regular process.

To try and make the process simpler for me, and hopefully for others in a similar situation, I started the AutotuneWeb project to:

  1. Automatically convert the profile information stored in Nightscout to the format required by Autotune
  2. Give me the command line details to run, and even,
  3. Run the command for me on my Linux VM and email me the results

I’ve been through a few iterations on how to get this to work, so in Part 2 I’ll run through the various parts that make up this system and how I’ve currently got it running in Microsoft Azure for minimum cost.