We closed the loop with #DIYPS and now have a DIY closed loop artificial pancreas running right now (for me, aka n=1.)

Here’s what a closed loop #DIYPS artificial pancreas looks like:

(A raspberry pi powered with a battery (shown) or via an outlet; a Carelink USB plugged into the Pi and in close proximity to the pump; a Medtronic insulin pump (an older version, thanks to a family from CGM in the Cloud for sharing it!); and of course my CGM and Android uploader phone that’s sending my data into the cloud/#DIYPS)

Here’s how a closed loop artificial pancreas or “#DIYPS closed loop” works:

Over the past year, we’ve been using #DIYPS to process BG, insulin on board and carbohydrate decay in real time. This yields real-time predictive alerts of three kinds: more insulin needed (X.X units of bolus insulin); less insulin needed (temp basal to zero insulin for X minutes); or carbs needed. Using this algorithm and my data entered (insulin taken or reduced; carbs ingested); we’ve seen significant results not only for the first 100 days, but also sustained over an entire year for improved average glucose, improved a1c, and improved time in range.

Closing the loop with #DIYPS

Three weeks ago, we mentioned at D-Data Exchange 2014 that we were planning to close the loop with #DIYPS…by August 1 (our wedding date :)).

The FDA was even in the room.

Two weeks later…we closed the loop.

It was this simple:

• Thanks to Ben West’s awesome work, we plugged the Carelink USB into the Raspberry Pi. Thanks to a generous family from the CGM in the Cloud Facebook group, we acquired an old 522 insulin pump that could talk to the Carelink USB. (My current pump is too modern to work with this setup.)
• Once we established that the Pi/USB setup could read from the pump, we started connecting #DIYPS to the system to close the loop.
• Normally, #DIYPS provides 3 types of recommendations (see above). We reconfigured it so that while “looping”, it only enables the pump to do temporary basal rates instead of boluses. If #DIYPS says I need more insulin, it tells the pump to begin a temporary basal rate for 30 minutes above my normal basal rate. If #DIYPS says I need less insulin, it tells the pump to begin a temporary basal rate lower than my normal basal rate for 30 minutes, effectively reducing my net insulin.
• When something changes and I need to do more or less insulin, #DIYPS tells the pump to increase or decrease the temporary basal rate, or resume my normal basal rate.
• Note: there is a maximum basal rate that the pump can provide. This is a safety check so that #DIYPS cannot recommend a seriously dangerous basal rate. In fact, if I need more insulin than #DIYPS closed loop can safely provide using a 30 minute temp basal, the system will alarm and alert me to that state (although I usually already know about it). Same for if I was dropping fast and needed carbs in addition to reducing insulin; it would alarm and recommend carbs above and beyond reducing insulin.

I’ve been ‘looping’ (using #DIYPS closed loop) for a little over a week. Here are some of my overnight results:

As you can see, the outcomes are awesome. In fact, they are better than I expected!

#DIYPS closed loop results are significant

Here’s a break down that shows on an example night (Sunday night, 12/14 through Monday morning, 12/15) of how #DIYPS performed:

You can see that around 10:30 (not pictured, but in the graph), I was running low and needed some temp basals to zero, then later some higher temp basals to accommodate the resulting rise in BG. Later, my normal basal rates sufficed until another point overnight when I needed less insulin to prevent a low.

While this is 1/9 or so example nights, this is pretty much how it works every night. BGs drift up? Temp basal higher (usually 1.6 or 1.8, which is <=.5u above my normal). BGs drifting low? Temp basal lower or in extreme cases to zero to prevent a low.

What happens if the closed loop artificial pancreas fails/stops working?

If the loop stops working*, #DIYPS continues to work normally (processing all data and pushing recommendations), and my pump goes back to my normal basal profiles after any existing temp basal rates expire (which would be a max of 30 minutes). No big deal and no more dangerous than I would be walking around with my pump acting normally.

*I’ve found numerous ways to make the system stop working: no battery power; being too far away from the USB stick so the pump can’t communicate; no cell signal to provide wifi as a hotspot to the Raspberry Pi; if the pump battery goes low, RF gets turned off and the USB can’t talk to the pump; if there’s no cell signal or wifi at all, I can’t upload CGM data to #DIYPS anyway; if the SD card gets corrupted inside the Raspberry Pi; etc.

So, there’s many points of failure. But it doesn’t matter. #DIYPS closed loop is safer for me (disclaimer: just me! I’m the only one doing this as a personal experiment to my own body) than if I just had a pump and a CGM, because it can reduce insulin and prevent lows. And it can increase insulin slightly and prevent rebounds.

After about a week’s worth of data, here’s the results of #DIYPS (what I’ve been using for the past year) versus the week of closed loop #DIYPS:

The target BG is especially awesome, because we have the closed loop targeting a BG of 115!

It’s also worth noting the reduction in alarms. Remember, I sleep through CGM alarms, so we built #DIYPS. #DIYPS makes a lot of alarms – sometimes upwards of 40 a night (those are really bad nights!). With the closed loop nights, there are very few alarms..and they’re usually alerts about things #DIYPS closed loop is taking care of on it’s own (see all of those no-hitters above!).

What’s next with #DIYPS closed loop

We’re going to continue and iterate on the closed loop and keep making it better. I’m definitely going to keep using it overnight – it’s like having a working pancreas, the kind where you can go to sleep without fear.

We’re also adding this to the ‘things we’ve learned while working on #DIYPS’ list:

• Calculating the carbohydrate absorption rate for an individual – how many carbohydrates your body absorbs in an hour
• Tracking “carbs on board” or “carb decay” – how many carbohydrates are still being digested in your body and impact BG levels
• “Net” IOB – tracking net IOB by incorporating “negative” boluses from temporary basal rates to zero
• Preventing post-meal BG spikes/drops with early meal boluses and better understanding of insulin activity
• … and you can control overnight BGs solely using simple adjustments of insulin rates.

Thanks to everyone who made it through a really long post, and for the MANY people who have contributed in various ways to helping us close the loop with #DIYPS! Also, let us know what questions come to mind that we haven’t managed to answer above about how the system works, results, etc.

We’ve been quiet on the blogging front (yes, I admit, I blogged!) because we’ve been busy closing the loop. (We’re always active on Twitter, though – make sure you watch #DIYPS on Twitter for more real-time updates in between blog posts, and you can follow @ScottLeibrand and @DanaMLewis.)

That’s right.

Closing the loop.

It feels good to say that :).

We’re submitting a research proposal to Medtronic to get access to their research device that will enable us to send commands to my insulin pump. These commands will be generated from #DIYPS. Because that would remove the “human in the loop” component that’s a current part of #DIYPS (i.e.  me pushing the pump buttons), that would move #DIYPS from being not an artificial pancreas to a closed loop artificial pancreas system.

This is not what we imagined we would do when we set off nearly a year ago to do this work (make a louder CGM alarm system, enable remote monitoring, etc.), but this is awesome!

And, in the true #wearenotwaiting spirit, we’re not waiting for the research device to start this phase of our work. We got started with this phase of our research by getting a Raspberry Pi and hooking it up to the Carelink USB and sending commands to a test Medtronic pump.

(Note: this pump is not hooked up to my body! There is no insulin in it! We’re just testing to see how data can be read off the pump & fed into #DIYPS. There will be many tests with a pump not connected to my body and many safety checks and stages before we would get anywhere near to allowing this to run under waking, close observation let alone under observed sleeping conditions. Stay tuned for more on this later.)

We’ve been having many discussions with the other developers of Nightscout, and end users in the CGM in the Cloud Facebook group around this topic. Scott and I are both passionate about the open source part of our project and being open and transparent every step of the way. We’ve been learning a lot from #DIYPS and have some key features that we’re putting back into the overall Nightscout platform as a result. We’re calling this “#DIYPS-Lite”, or the “wip/iob-cob” branch of Nightscout. (This is not what the main community is using, it’s still an advanced test branch, but enables more advanced users to implement and take advantage of some of these features, like net IOB and COB.) We’ve also been contributing to the roadmap of Nightscout; as of this past weekend, we helped map out the priority list to show what are priorities of some of the core developers (including Scott and I); what’s currently being worked on; and what’s on the ‘backlog’ as a project we’re all interested in and is important to the community, but no one’s had a chance to jump in and start it yet.

Back to the closed loop update

We won’t be releasing any closed-loop systems to the public any time soon, because that would mean distributing an untested medical device system that controls administration of a potentially lethal drug (insulin), which the FDA (and anyone in their right mind) would (rightly) frown on.

Instead, we want to open up our process of developing such systems, so that other researchers (including independent ones like us, and those inside traditional research labs and companies) have a roadmap to follow for how to safely put together and test an open-source artificial pancreas system.  The more people and companies we have developing such systems, sharing their results, and learning from each other’s work, the closer we’ll be to having a system that could indeed be proven safe and effective to the FDA’s satisfaction, and eventually released to the public.  In the end, we suspect the companies and research labs developing APS systems will be the ones to commercialize them, but we are confident that our efforts will accelerate those timelines and result in a more dynamic, innovative T1D medical device ecosystem that will provide a bridge to a true cure.

We are hoping that we can show the world how open source innovation and new regulatory paradigms can deliver safe and effective results for people living with T1D faster than traditional medical device development and traditional regulation.  By doing so, we can change how all successful medical device companies approach interoperability, and get to a world where the latest ideas, algorithms, and software can be made available to everyone with T1D, regardless of what hardware they’re using.

The Do-It-Yourself Pancreas System (#DIYPS) is a tongue-in-cheek name that recognizes a serious fact: people with (type 1) diabetes (PWDs) have been acting as their own pancreas for years.  What a healthy pancreas does using subcellular machinery in islet beta cells, and a full closed-loop artificial pancreas system (APS) attempts to do automatically using medical technology, are the same things that PWDs have been doing, on their own, with whatever help they can get.  We recognize that for many PWDs the available help is not yet enough, so we are not waiting.

Medical device companies are working on new Artificial Pancreas Device System components, and researchers are conducting clinical trials of artificial pancreas systems incorporating various levels of closed-loop technology to do anything from automatically suspend basal insulin when BGs are below a pre-set threshold (available now in the Medtronic MiniMed 530G) to a fully closed-loop bionic pancreas system that automatically doses insulin (and even glucagon, in one trial) to provide full 24/7 BG control.  However, full closed loop systems are still years away, so there is a serious unmet need for better tools to help people with diabetes manage in the meantime.

The #DIYPS is designed for immediate deployment, using a device-agnostic web-based interface, and optionally ingesting data from any of a user’s existing medical devices using commodity off-the-shelf hardware (such as a Moto G Android phone connected to a Dexcom G4 via USB).  In fact, interactive use does not even require automatic data upload capabilities: once the system is configured for their insulin sensitivity and other user-specific parameters, a user with any sort of CGM or meter can manually enter BG levels, boluses/injection amounts, and carb counts and get an immediate estimate of their resulting BG levels.  In addition, the system will estimate insulin on board (IOB, meaning insulin still working in the body) and unabsorbed carbs, so a later follow-up assessment only requires updating the BG data to figure out if any additional action is required.

The #DIYPS can be used immediately to empower PWDs to prevent many instances of out-of-range BG levels.

The #DIYPS is not a substitute for a true closed-loop Artificial Pancreas*.  It is not a medical device, and does not dose insulin.  However, it does fill two currently unmet needs.  Firstly, none of the CGMs on the market currently support robust enough alerting features to awaken many patients when they experience low BG while sleeping.  Such nocturnal hypoglycemia unawareness is a serious risk for many people with diabetes, and better tools are required to deal with it.  The MiniMed 530G’s Low Glucose Suspend feature is a huge step forward there, but even it doesn’t help with the converse problem: at other times users will end up going an entire night with high BG, and never wake up to their CGM’s high BG alarms.  The #DIYPS’s alert system certainly cannot replace a CGM’s alarms (it’s not a medical device, and various interruptions in connectivity can prevent it from alarming), but it does have the ability to send notifications that can trigger alarms (i.e. on an iPad or phone) loud enough to wake the deepest sleepers. It can also function as a de facto remote alert or monitoring system, allowing family or loved ones in different locations to check and make sure a PWD is responding to alarms. Until device manufacturers get FDA approval for loud enough devices, this alone is enough reason for many people to want to put together their own alerting system like the #DIYPS – and was the original reason for building it.

Secondly, it is currently completely up to patients (with occasional help from their doctors) to determine how much insulin is required at any given time, and decide when additional carbs are required to prevent a low.  Performing such calculations manually every few hours, without ever being able to take a break, is a major problem for the quality of life of many people with diabetes.  There are very few effective tools available today to help, so such calculations are usually done in the user’s head, and are very prone to human error.  (The few tools available are also often not in one system, creating barriers for effective use.) In addition, many factors such as stress, sickness and varying activity levels can impact BG levels.  This combination of factors often results in frequent (often daily) instances of low or high BG levels.  The #DIYPS system provides a framework to support the user in making the best decision in every situation, while reducing the cognitive load on the user, and reduce the risk from harder-to-calculate factors like stress or varying activity levels.  It also helps the user to pay more attention and take action when they are at risk for hypoglycemia (predicted based on CGM, IOB, and carb data input into the system), while allowing them to continue living their life normally, with less worry about managing diabetes.

Finally, the #DIYPS system is designed to be device and data source agnostic, and to integrate with frameworks like Tidepool to allow users to do things that have never been possible before.  As new and better medical devices are released, our hope is that all of them will support methods for allowing a user to export their own data, so that users can truly take advantage of all the innovation we are starting to see in this space.

But today, the #DIYPS is just a prototype system, and has only been development tested with a single user.  In order to determine whether it can truly fulfill the unmet need described above for a wider population, additional development and testing are required.  At this point, we are looking to begin collaborating with anyone who has the skills and interest required to move the project forward.  We are not yet ready to begin allowing (and helping) a large number of users to begin testing the system themselves, but hope to open it up (and hopefully integrate it with the efforts underway at Tidepool.org, among others) in a manner that will allow as many people as possible to begin safely using the system as soon as possible.  However, this is currently only a 2-person side project, so if you think it’s worthwhile to move forward more quickly, I would encourage you to get involved (if you have technical or other skills that would be helpful) or pass along the word to others who might be interested and able to help.  (And if you haven’t already, you might also want to read How I Became a DIY Artificial Pancreas System Builder.)

P.S. A special thank you to John Costik, who graciously provided the drivers we’ve been using to pull data off the Dexcom CGM to a Windows PC, and who is working on perfecting the Android (Moto G) drivers as well.  We wouldn’t have done any of this if it wasn’t possible to get data off the Dexcom CGM in real time.  We’d also like to thank members of the diabetes online community for their support, encouragement, and feedback: we know y’all will be pivotal in driving this forward. 

–Scott Leibrand and Dana Lewis

*Update in December 2014: We wrote this post in February, when #DIYPS was truly a “human in the loop” decision-assist system. However, as of December 2014, we closed the loop with #DIYPS, and also have a version of #DIYPS that is essentially a closed loop artificial pancreas. You can read more about our work on the closed loop artificial pancreas version of #DIYPS here.

As some of you may know, I’m a Network and Systems Engineering type who works in the Internet industry.  I have a B.S. in Cell and Molecular Biology, but I have been working in computer networking and for Internet companies for the last decade and a half (since I was still in school), and have had the privilege of working on a number of different types of systems in that time.  Nine months ago, I met a wonderful person who happens to have Type 1 Diabetes, and began learning just what was involved in managing the condition.

On the one hand, I was impressed at the level of data available from her CGM (glucose readings every 5 minutes), and the visibility that provided into what was going on.  But I was also very surprised to find that the state-of-the-art medical technology she uses is, in many ways, stuck in the last century.  Data is not shared between devices; her pump looks and acts like it came straight out of the early 1990s.  (But at least it’s purple!) 😉

Most importantly, it is completely up to the person with diabetes (PWD) to collect all the relevant data on their current state, do a bunch of math in their head, and decide what to do based on the data, their experience, and how they’re feeling.  And as a PWD, that’s not something you just do once in a while: it is a constant thing, every time you eat anything, every time your blood sugars go high or low (for dozens of reasons), every time you want to go exercise, etc. etc. etc.  As a PWD, you’re dealing with life-and-death decisions multiple times a day: Too much insulin will put you into hypoglycemia and can kill you.  No insulin for too long will put you into diabetic ketoacidosis and kill you.  Overreacting to a high or low blood glucose (BG) situation and correcting too far in the other direction can completely incapacitate you.  So as a PWD, you never get a break.

So, very early on, the obvious question was: why can’t we integrate all this data and make this easier?  There were promising signs that it could be done: Artificial Pancreas Systems developed by various research groups and medical device companies are in clinical trials, and are showing promising results.  In fact, she had just signed up for such a trial, and I was able to go with her to the clinic and watch just how a fully automated APS system works.  But despite these brief hints of a better future, day to day we were (usually, she was) stuck doing everything manually.

It was very clear from the beginning that the primary bottleneck to doing something better was the ability to get blood glucose (BG) data off her CGM (a Dexcom G4) in real time.  We knew it was possible to do so manually using Dexcom Studio, a Windows-only software package that is actually quite good for analyzing historical BG data, spotting trends, etc.  But unless we were going to create a Windows macro to open up Dexcom Studio, import the data, export it to CSV, and then close Studio every 5 minutes, we couldn’t get our hands on the data to do anything in real time.

Then, we discovered that John Costik had figured out how to use the Dexcom USB driver’s API to get data off his young son’s CGM in real time, and display it remotely, even on his Pebble watch, and even when his son was at day care or kindergarten.  This was the breakthrough we needed, so we contacted John, and he was able to provide us a copy of his script.

So now we were off to the races.  I cobbled together a system using Dropbox for uploading the BG data from a Windows laptop (that ended up in a bedside drawer), and once I was able to get the data onto a VM server, got working on putting together something that would actually make a difference in her quality of life.

So we worked through a number of ideas.  First off, we needed something that could wake her up at night when her BGs got too high or too low.  The Dexcom G4 is supposed to do this, but even its loudest alerts are still not loud enough to wake a sound sleeper.  People recommend sticking the G4 in a glass or in a pile of change, but that doesn’t work, either. She also had used Medtronic’s CGM system; but even its loudest, escalating “siren” alert isn’t all that loud, even if awake.  We also needed something that would allow me to see the alerts remotely, and see whether she is awake and responding to them.  This is a big deal for many people, like her, who live alone.  So, I started coding, and she started testing.  We ended up using my cracked-screen iPad (it got into a fight with Roomba) as a bedside display (it’s much easier to glance over and see BG values than to have to find the CGM and punch a button), and also to receive much louder push notification alerts.

After we got basic notifications working, we also started trying to enhance them.  We ended up with a prototype system that allows the user to enter when they bolus insulin or take carbs, and snoozes alarms appropriately.  (This enables a de facto remote alert monitoring system, especially handy for individuals living alone.)  With that info, we are also able to do the calculations required to determine how much insulin is needed at any given point (the same calculations PWDs do in their head every time they eat or have to correct for high BG), and by doing those every 5 minutes as new data came in, we were able to provide early warning if she was likely to go low, or if she needed more insulin to correct a persistent high.  Since getting all of that working, we have also developed a meal bolus feature, which takes advantage of an (as far as I know unique) ability to track how fast carbs are likely to be absorbed into the bloodstream to give the user better estimates of how much insulin is required now, vs. how much will likely be required later as the rest of the meal is digested.

Since she started began actively testing the #DIYPS (Do It Yourself Pancreas System), we have seen a decrease in average BG levels, with less time spent low, and less time spent high. It also enables her to execute “soft landings” after a high BG and prevents “rebounds” after a low BG.  We are very encouraged by what we’ve seen so far, and are continuing to iterate and add additional alerts and features.  But today, the #DIYPS is just a prototype system, and has only been development tested with a single user, so lots of additional development and testing are required.  We are not yet ready to begin allowing (and helping) a large number of users to begin testing the system themselves, but hope to open it up in a manner that will allow as many people as possible to begin safely using the system as soon as possible.  However, this is currently only a 2-person side project.  We are looking to begin collaborating with anyone who has the skills and interest required to move the project forward, so if you think it’s worthwhile to move forward more quickly, I would encourage you to get involved (if you have technical or other skills that would be helpful) or pass along the word to others who might be interested and able to help.

If you’re interested, you can also read more about why a #DIYPS is needed now (and how it compares to true closed-loop automated Artificial Pancreas Device Systems).  We’ll also be publishing a full description of how the #DIYPS prototype works shortly, so stay tuned.

–Scott Leibrand