With the recent news that Tandem have suspended their Control-IQ trial due to issues with it’s behaviour under certain CGM circumstances that may cause a hypo (and those in the DIY world can probably read between the lines and guess what that means), I though it was worthwhile digging further into what we know about Control-IQ, based on the articles and documents out there, how it compares to existing solutions and what this means for users and healthcare professionals alike.
To start with, I want to make clear that this is based on feedback on Control-IQ from an article by A Sweet Life and also an article in SixUntilMe, but reading through both of these, it suggests there are some implications of the decisions that have been made.
Firstly, the highlights from A Sweet Life:
- It’s the TypeZero Algorithm, with very little adjustment, running in the Tandem T:Slim X2 pump;
- The algorithm is layered on top of the user’s pump settings;
- Target range is 112.5-160 mg/dl (6.25-8.9 mmol/l) in normal use;
- Sleep setting with range 112.5-120 mg/dl (6.25-6.7 mmol/l);
- Exercise setting (range not given);
- Adjustments to basal rates and automated bolus only occur when the predicted glucose level is expected to be higher than 160 mg/dl (8.9 mmol/l);
- Delivers insulin in a two phase model, first adjusting basal rate every five minutes and then giving a single correction bolus per hour of 40% less than what the pump settings call for (but no bolus functionality when in sleep mode);
- Basic learning capability using total daily dose compared to current settings and adjusting based on this information.
Currently Approved systems
Before we look at DIY systems, a number of things stand out compared to the Medtronic 670G.
- The 670G uses a single target value instead of a range.
- There is no overnight function in the 670G.
- The 670G learns about a user prior to it being enabled in auto mode (two weeks is the normally specified time) and continues to do so following Auto mode start up.
- There is no correction bolus in the 670G.
It’s fairly tricky to make a comparison with what we know about the Diabeloop DBLG1 algorithm as there is little information in the public domain. From conversations with the team there and presentations they have given, we know that Diabeloop has:
- User adjustable targets and “aggressiveness”.
- Zen mode, when you don’t get interrupted by the system.
- Uses a short term loop algorithm with machine learning, and fails back to an expert system when that doesn’t work.
- You only enter Total Daily Dose, meals and weight. The system then adapts to your physiology.
So on the face of it, there are some distinct differences between Control-IQ and the other systems out there. The biggest of these is the self-learning component, and then secondly, the use of a range in the decision making process.
Let’s look at all of these individually.
Reliance on Pump Settings
For me, this is the big one with Control-IQ. As we’ve learned with the DIY systems, it’s the one thing that most people don’t have right when they start using DIY systems, so why, with a commercial system, would we expect their settings to be any better?
Both Diabeloop and Medtronic have recognised that there are issues with this and introduced self-learning into their setups to address when users haven’t got a good set-up in their pumps, so either the TypeZero algorithm manages much better with bad pump settings than any of the other systems, or there’s something else at play.
As documented by Katie DiSimone, we may see users reverting to learnings from the DIY community to get better results from Control-IQ, or there will be a significant overhead on healthcare professionals to work with the users in order to set pump settings correctly, and manage the initial system set-up. This may also be the reason for the next point…
A target range instead of value
As the “A Sweet Life” article mentions, the algorithm in Control-IQ only delivers insulin if it predicts that glucose levels will be higher than the higher end of the target range. The 670G has a set target of 120mg/dl and Diabeloop allows a user defined target, but I’d expect there will be safety limits around exactly what that is.
This means that Control-IQ will only deliver insulin in what a fair few users would consider to be a case where glucose levels were already heading high. What isn’t clear is whether it delivers insulin with the intention of getting you below the high mark, down to 112.5 or a midpoint between the two. Given the output from the Adolescent skiing trial, which resulted in average glucose levels of 7.8mmol/l (140mg/dl), it would suggest that the system is targeting the range midpoint when correcting.
Given the reliance on the pump settings in place when starting to use the automated algorithm, it’s no surprise that it won’t deliver insulin until a higher level is predicted, but with that limitation in place, it would also suggest that the value it targets may not be the lower bound of its range.
I’d expect that when the prediction is below 112.5, it will suspend or reduce basal delivery in order to bring the level back up above that.
If you look at how this is viewed in the DIY world, where a temporary basal rate is generally set for a minimum of 15 or 30 mins and corrections are done towards the midpoint of the range, wider targets are not considered to be a good thing as they result in the need for a significantly higher temp basal when crossing over into marginally “high” at a higher glucose level target than at a lower one. How the range works and the decision making relating to this is clearly important, as is whether those temporary basal rates on Control-IQ last for 5 minutes, or longer.
The other aspect that a greater range provides for is significantly more variation around the median of that range that you’d see with a single target value, which would lead to higher GVI and PGS values, indicating greater variance.
Both Medtronic and Diabeloop have put a lot of effort into the ability of their systems to learn about the user. This comes about from the idea that many users don’t necessarily have optimised pump settings and that this needs adjusting. Even when you look at people undertaking pump starts, the period of time that is usually suggested to get settings right is measured in weeks, if not months.
As a result, they implemented “Machine Learning” techniques to allow the system to discover more about the user. Control-IQ doesn’t appear to do this in the same league, and the description in the article suggests that on a daily basis, it will calculate the TDD delivered versus the TDD required and adjust what it believes the background basal rates, and use a formula for Insulin Sensitvity Factor and Carb Ratio off the back of this. I’d hope that it does more than that and adjusts both the latter based on observed data and calculation, but that’s not what is suggested in the conversation.
In this respect it would sit somewhere very different from the other commercial offerings, and OpenAPS, which Autotunes nightly (all be it on a single ISF/CR). Overall then, it suggests that Control-IQ may not be quite so adaptable as as both other commercial offerings and DIY systems.
User feedback on the system
As ever, there are a few people that have or are participating in trials and have broken cover. Many are very happy with the outcomes, but as ever with the trials, you have to look at the user base on which the trials apply. The SixUntilMe trialist was very happy, and other trial participants seem to be too.
Feedback from DIY users has been less positive, mainly due to the rollercoastering effect they’ve seen by having the wide target range and limited automatic tuning.
In addition, there will always be the feedback from those who run a target range or value that is outside (usually lower) the systems limits. Once again, this is a system targeted at the 72% with Hba1C levels higher than 7.5%, rather than those who maintain a lower value.
While this is an “Over the Air” upgrade, it is expected to come at a price. It’s not yet known what that price will be, however, estimates have been bandied around in the $100 to $200 mark, but there has been no announcement as yet. Given that this is a download on top of existing pump settings, I’d expect that there will be a fair few clinics that would be concerned about users directly downloading this, and that there will be a requirement for HCP approval before starting/paying.
While there is still very little information in the public domain about how TypeZero inControl and Control-IQ work, there is enough for us to make some inferences.
The stand out question that arises from the various articles and interviews relate to getting a user started. Within a trial, you expect there to be a period at the beginning where euglycaemia is achieved effectively, by adjusting various settings to go into the trial. This is where trials and real life often conflict.
From our experience within the DIY world, we know that even experienced pump users often have misaligned settings when moving on to a closed loop system, and as a result, OpenAPS has Autotune and AndroidAPS guides you through a set of objectives to get you started with closed looping, while Katie DiSimone has written the definitive guide to adjusting settings to ensure effective looping.
Given all of this, will new users of Control-IQ be expected to go along to a series of clinic sessions to confirm that their pump settings are good enough to instantiate Control-IQ?
If that’s the case, the workload for healthcare professionals is likely to become the limiting factor for uptake of the system, as we know that this process can take a bit of time and requires user buy-in to be successful. This would also mean that uptake in healthcare systems that already have resource constraints around pump roll outs would only be more constrained. Definitely one to wait and see on.
Secondly, the target range instead of value. Assuming a midpoint correction target (safety reasons) and no action taking place until the daytime 160 mg/dl prediction occurs, it’s possible to imagine average glucose levels on Control-IQ ending up around the 135-140 mg/dl (7.5-7.8 mmol/l) level. The adolescent skiing trial paper seems to back this up, and while this may seem high to many people, it’s around the level you’d expect with an Hba1C value of 6.5%/48mmol/mol, which is what I expect their target was. Like the 670G, this will be too high for some.
Finally, the sleep functionality with the tighter range appears to have a variable time setting on it. One wonders whether that could be extended to most of the day if a user so chooses? Even if it disabled the “one bolus per hour” function, that tighter range that would offer would be attractive to many.
Footnote: The trial suspension issues – some speculation
While we haven’t been given any information relating to the “bug” that’s caused the ControlIQ trial to be suspended, given the information stated relating to hypoglycaemia risk and CGM and that it will be fixed bya software patch, I’m guessing that the issue is caused when the CGM readings jump around with a noisy sensor, and a sudden high value or series of high values greater than the 160mg/dl threshold result in the bolus functionality kicking in when perhaps it shouldn’t have. Regardless, Tandem have promised a quick fix, which is the important thing as it will get the trials back on track.
I think that due to variability in other things which closed loop systems rely on (e.g. CGM sensor input, and pump site efficacy) the only system that will be workable at scale is one which continually self-tunes. E.g. any system which ignores the fact that a pump site on day 3 is unlikely to work anywhere near as well as a pump site on day one is always going to get bad numbers on day 3. I’d love to see C-L systems actually warn about pump site efficacy based on how hard they are having to work.
I can’t soeak for the commercial systems, however, Autosens in AAPS and OpenAPS is effectively doing this already.
Very good points. 100% agree