I’m looking forward to Part 2 of this very insightful series! I would suggest another element of the complex system is the user’s own behavior. I have been using AID systems since the 530G and my A1c’s have minimally improved (I’m generally in the high 6’s with very few lows) but the improved responsiveness, as well as ultra-rapid insulins, have enabled me to vary my eating and exercise habits more while maintaining similar control. I exercise more and on a more irregular schedule than I used to (I used to avoid physical activity in the evening because of fears of spiking or going low before bedtime), and I am able to indulge a bit more.
Would I have better TIR and lower A1c if I ate like I did in 2013? Probably. But my quality of life would also be lower.
For me, while much less measurable, better quality of life and lower anxiety have been some other great benefits of better AID technology.
"the user's own behavior" is not only discussed in this article, it's strongly emphasized that it's the overwhelmingly largest factor in the whole equation.
As for your outcomes being improved by insulin absorption kinetics (and other factors), that may feel apropos to you, but in the end, it's really just a variable that--again--is compensated for by the user's behaviors. Those behaviors are set by expectations, which take time to develop. It's sort of like driving a rental car that has a different kind of responsiveness to the gas pedal and brakes from YOUR car. Initially, it feels off, but after a few days, you get used to it and calibrate your feet accordingly.
T1Ds just learn to adapt to whatever they have available to them. Remember, the data in the article showed that A1c and TIR levels for most adults were nearly the same before automation was ever introduced, and those numbers still haven't changed. And these numbers date back decades.
The tech is getting either credit or blame for whether a user succeeds or fails, when the whole time, it's always been the user's behaviors to govern outcomes.
As for your stress levels and quality of life, there's a lot of research in this area as well that I cite in my "paradox" article, where the claim that automation relieves burden has not borne out in randomized clinical trials. Most people don't like automation, but those who do stick with it. The result is that most people who use AID systems love them, but this bias is not representative of the larger population.
This isn't to suggest there's anything nefarious going on. It's just that people use whatever they like (or, at least, should have that option), so long as their health is well-managed. And that's where metrics beyond A1c and TIR are really important factors. Evidence shows that most AID users gain weight and lose the autonomy necessary for self-engagement, which brings us full circle to your initial point: the user's own behavior.
I take your points, but “The result is that most people who use AID systems love them, but this bias is not representative of the larger population.” can’t really remain true as a larger and larger share of T1Ds are using such systems.
that's a good point that I should have addressed. Many T1Ds are put on AID systems without necessarily understanding what they are, their options, or the training to make an informed decision. In my article on "standards of care", I highlight the fact that the number of endos is at an all time low, and dropping fast. The overwhelming majority of T1Ds are managed by primary care physicians who self-report that they are not trained sufficiently in T1D management to give proper guidance to their patients. They put their patients on AID systems because it's the easiest and safest choice among all other bad choices. It's essentially throwing a life preserver in the water so the patient doesn't drown. Yes, it's good and effective at preventing immediate harm, but this is not a long-term sustainable solution. The metaphorical sharks are swirling around the patient in the form of obesity and reduced metabolic health because the algorithms sort of lock them into a behavior pattern of a sedentary lifestyle. An informed and engaged patient may learn to navigate that, but therein lies the problem: being informed and engaged is not on anyone's agenda. It's "relieve the burden and put diabetes in the background." That messaging is the problem.
Dan, excellent article. Thank you. I wonder if your thesis that pump vs no pump being about equal is biased by the population contained in the T1D Exchange data set. They are among the most prestigious diabetes clinics in the US. Half of all people with T1D don’t see an endo. Another published stat suggests only 1 in 4 ever achieve an A1c of 7 or lower…. If the effect of pumps is a closer to the reduction of a point, then we have research that shows 40% fewer complications and it starts to get meaningful at a population level. This doesn’t take away from your point about the performance ceiling….
Len -- As you've probably discovered, I have a lot of articles that cover all these topics. But one thing that bears particular emphasis. You said, "If the effect of pumps is a closer to the reduction of a point, then we have research that shows 40% fewer complications and it starts to get meaningful at a population level"
The 40% fewer complications is highly deceptive and misleading, largely because of the mixture of populations. It's not just THAT someone's A1c improved by a point, but HOW. Those on automated systems achieve it by getting progressively too much insulin and live a largely sedentary lifestyle. They're avoiding some of the risks at the microvascular level -- kidney, eyes, etc. -- but are amplifying their risk at the macrovascular level. They're gaining weight, have higher lipid levels, among many other problems.
By contrast, those on MDI who achieve lower A1c levels tend not to have any of these problems. When you combine populations, you get the misleading impression that lower A1c unequivocally equates to better health outcomes.
The points you make about the systemic problems--that most T1Ds don't see an Endo--is even made worse because the PCP's they do see self-report that they are not equipped to treat T1Ds because they never had sufficient training. For them, it's easier to put them on an AID system because there's no one whos going to teach the patient how to manage themselves. I cover this in my article: https://danheller.substack.com/p/standard-of-care-primer
The problem I see with your argument is that AID systems are not about perfect predictability, but reasonable control. While it may not be possible to predict chaotic systems long-term, they absolutely can be controlled with fast enough feedback loops. This especially holds true with the development of faster acting insulins, since the needed prediction horizon shrinks.
You can't look at this as simply "insulin goes faster = algorithm improves". The rate of absorption just affects all the other arms in the multi-arm pendulum -- the faster absorption (even if hypothetically possible) only means that the kinetics of *other* hormones now act differently than they were when the insulin was slower. It doesn't make it better or more predictable at all -- it just means that chaotic noise will just have a different kind of chaos.
For me the only way AID systems can improve Is by adding multiple input datas into the equation. Like algorithms that have access to datas like Heart rate, ketones, hormones and stress level, sleep datas, etc. and not only glucose level. Some companies are already working on it.
As the article pointed out, you can have access to each and every variable that goes into the system and still not control it because of its category: Chaos. The parts interact with one another in ways that make them impossible to predict, even if you can measure them. If you can't predict their future states, you cannot control it.
You can't engineer around a mathematically impossible problem.
The easiest thing to do is simply keep the person in the loop. That’s it. Nothing more complicated than that. Doesn’t even have to be much. The person knows In advance when they’re going to do things. Just announce it. There more they fit that, the better the control and the healthier the outcomes. This constant pursuit of total hands-free operation is not only impossible it’s what’s killing diabetics
I love the chaos theory. But there are different levels of chaos. One is the big pharma convincing doctors that a movable target with many variables such as type 1 DM can be handled by only one hormone that we supposedly can master or manufacture depending on how you want to look at it. The other level is convincing people that blood sugars in TIR is only achievable by more and more insulin. I would argue and have many anecdotal experiences that our total insulin use per patient has gone up since AID because de facto all providers and patients are doing is increasing the dose of insulin now to tighter and tighter TIR
What if there haven’t been updates to AID because the economics of the market don’t incentivize it?
Could there be an attribution error in your analysis? Instead of the impossibility of FCL AID attributed to complex systems theory, perhaps the lack of innovation is a classic market incentive problem?
Here’s the logical challenge I have with the reasoning you’re offering: if humans can make the decisions, then it’s not an impossible problem to solve. Following your logic, then MDIers wouldn’t see better outcomes either.
there's a lot to this, so let's tease out a few things separately. First, as the article indicated, the physics are such that it's chaos theory -- you can't predict the state of the parts, therefore, you can't control the system. It doesn't matter how much money you throw at the problem, it's unsolvable.
And the researchers working on this know that -- so instead, they're trying to isolate edge cases: safety from hypoglycemia, or maybe post-meal analysis to cut down on higher excursions. Fine, but again (as the article indicated), there's not a lot of room to maneuver in the area, so eking out a few percentage points may be possible, but comes with risk. This is why the looper community tends to do better---they can personally control their own risk tolerance, and take on slight changes to that risk profile that may be suitable to their own personal physiologies, but are not generalizable to the broader population.
And this is no different than how T1Ds have always managed themselves---they learn through empirical experience what works for them. We've always had the freedom to choose how much to dose at any given time. For this reason, the DIY/looper community is and should always be available to anyone willing to engage in that level of algorithm refinement.
And that brings us to your "logical challenge", and you're exactly right: if humans can make the decisions, then it’s not an impossible problem to solve. True, but now we're not talking about technology, we're talking about humans making decisions that are rooted in both empirical experience (which cannot be reduced to math, therefore, an algorithm), but also --most importantly -- AHEAD of time. If they know they're going to exercise, they turn off basal dosing an hour ahead of time. If they know they're going to eat, they pre-bolus. The more they know about their life experience, the more they can draw on those personal heuristics to achieve "better" glycemic control.
And that brings us to your point about MDI users: as the article points out, *engaged* MDI users achieve roughly the same TIR as *engaged* HCL users, demonstrating the outsized role that human decision-making plays.
The point is, you cannot take the human out of the loop. Everything--especially algorithms--work better when the human is in the loop. The more informed and engaged they are, the better *their* decisions are when it comes to dosing, whether it's MDI, or aiding the AID.
It would be interesting to see by how much exactly does the engaged looper have less burden than the MDI user. If you constantly have to tell your pump what to do, doesn't it defeat the whole automation purpose? I mean, by the time I press those buttons and wait for the piston to push in the insulin, I'd already have working insulin after manual injection. And without any painful, unreliable mediators like catheters, cannulas and all the caveats of insulin being pumped to the same spot for days...
What would be the point for anyone to use or develop this then? Surely there are T1Ds in the DIY Loop community developing this for their own benefit, not just commercial "for thee but not for me" type of thing.
It seems like it stems from the trend of digitalizing everything in our lives, even when digitalization doesn't fit at all or is within serious limits, so people just like deluding themselves with a sense that their pump is in control with all those "settings" while in reality it's more of a illusion, because every little setting is constantly changing and you have no way in measuring anything except one - blood glucose reading.
A bit beyond the topic, but reminded me of a line from Epstein interview lol, where he was talking about how math is old-fashioned, not the end of science and how you can't put a number on everything. https://youtu.be/dOSIZzcV2ks?t=5167
It's a bit "all there" at first glance, but also interesting that I use more intuition than math when managing T1D, the only math I use is choosing the number of the dose on a pen and a CGM reading, it seems to work better than anything.
You're right on several counts. This is called The Christmas Tree Effect. (I wrote about this a year or two ago.)
The core idea comes from Leidy Klotz's research: humans are hardwired to solve problems by adding things, and consistently overlook better solutions that involve taking things away.
AID systems are a Christmas tree. CGM, pump, algorithm, Bluetooth, phone app, cloud uploads, sharing features, auto-corrections, activity modes, sleep modes — each one added to "solve" something, layered on until the system is so complex it introduces new failure modes that didn't exist when you were just taking shots. My Performance Paradox article is basically a case study in this.
The sneakier point: people enjoy decorating the tree. The complexity isn't just tolerated, it's the attraction. Tinkering with settings, watching the algorithm work, checking the graphs, feeling like you're piloting a spacecraft — that's not a cost, it's a feature. Which means the normal corrective ("just subtract") runs into the problem that the ornaments are the dopamine. You're not just fighting a cognitive bias toward addition, you're fighting the fact that the additions are genuinely entertaining.
I’m looking forward to Part 2 of this very insightful series! I would suggest another element of the complex system is the user’s own behavior. I have been using AID systems since the 530G and my A1c’s have minimally improved (I’m generally in the high 6’s with very few lows) but the improved responsiveness, as well as ultra-rapid insulins, have enabled me to vary my eating and exercise habits more while maintaining similar control. I exercise more and on a more irregular schedule than I used to (I used to avoid physical activity in the evening because of fears of spiking or going low before bedtime), and I am able to indulge a bit more.
Would I have better TIR and lower A1c if I ate like I did in 2013? Probably. But my quality of life would also be lower.
For me, while much less measurable, better quality of life and lower anxiety have been some other great benefits of better AID technology.
"the user's own behavior" is not only discussed in this article, it's strongly emphasized that it's the overwhelmingly largest factor in the whole equation.
As for your outcomes being improved by insulin absorption kinetics (and other factors), that may feel apropos to you, but in the end, it's really just a variable that--again--is compensated for by the user's behaviors. Those behaviors are set by expectations, which take time to develop. It's sort of like driving a rental car that has a different kind of responsiveness to the gas pedal and brakes from YOUR car. Initially, it feels off, but after a few days, you get used to it and calibrate your feet accordingly.
T1Ds just learn to adapt to whatever they have available to them. Remember, the data in the article showed that A1c and TIR levels for most adults were nearly the same before automation was ever introduced, and those numbers still haven't changed. And these numbers date back decades.
The tech is getting either credit or blame for whether a user succeeds or fails, when the whole time, it's always been the user's behaviors to govern outcomes.
As for your stress levels and quality of life, there's a lot of research in this area as well that I cite in my "paradox" article, where the claim that automation relieves burden has not borne out in randomized clinical trials. Most people don't like automation, but those who do stick with it. The result is that most people who use AID systems love them, but this bias is not representative of the larger population.
This isn't to suggest there's anything nefarious going on. It's just that people use whatever they like (or, at least, should have that option), so long as their health is well-managed. And that's where metrics beyond A1c and TIR are really important factors. Evidence shows that most AID users gain weight and lose the autonomy necessary for self-engagement, which brings us full circle to your initial point: the user's own behavior.
I take your points, but “The result is that most people who use AID systems love them, but this bias is not representative of the larger population.” can’t really remain true as a larger and larger share of T1Ds are using such systems.
that's a good point that I should have addressed. Many T1Ds are put on AID systems without necessarily understanding what they are, their options, or the training to make an informed decision. In my article on "standards of care", I highlight the fact that the number of endos is at an all time low, and dropping fast. The overwhelming majority of T1Ds are managed by primary care physicians who self-report that they are not trained sufficiently in T1D management to give proper guidance to their patients. They put their patients on AID systems because it's the easiest and safest choice among all other bad choices. It's essentially throwing a life preserver in the water so the patient doesn't drown. Yes, it's good and effective at preventing immediate harm, but this is not a long-term sustainable solution. The metaphorical sharks are swirling around the patient in the form of obesity and reduced metabolic health because the algorithms sort of lock them into a behavior pattern of a sedentary lifestyle. An informed and engaged patient may learn to navigate that, but therein lies the problem: being informed and engaged is not on anyone's agenda. It's "relieve the burden and put diabetes in the background." That messaging is the problem.
https://danheller.substack.com/p/standard-of-care-primer
Dan, excellent article. Thank you. I wonder if your thesis that pump vs no pump being about equal is biased by the population contained in the T1D Exchange data set. They are among the most prestigious diabetes clinics in the US. Half of all people with T1D don’t see an endo. Another published stat suggests only 1 in 4 ever achieve an A1c of 7 or lower…. If the effect of pumps is a closer to the reduction of a point, then we have research that shows 40% fewer complications and it starts to get meaningful at a population level. This doesn’t take away from your point about the performance ceiling….
Len -- As you've probably discovered, I have a lot of articles that cover all these topics. But one thing that bears particular emphasis. You said, "If the effect of pumps is a closer to the reduction of a point, then we have research that shows 40% fewer complications and it starts to get meaningful at a population level"
The 40% fewer complications is highly deceptive and misleading, largely because of the mixture of populations. It's not just THAT someone's A1c improved by a point, but HOW. Those on automated systems achieve it by getting progressively too much insulin and live a largely sedentary lifestyle. They're avoiding some of the risks at the microvascular level -- kidney, eyes, etc. -- but are amplifying their risk at the macrovascular level. They're gaining weight, have higher lipid levels, among many other problems.
By contrast, those on MDI who achieve lower A1c levels tend not to have any of these problems. When you combine populations, you get the misleading impression that lower A1c unequivocally equates to better health outcomes.
See my article: https://danheller.substack.com/p/performance-paradox-of-automated-insulin-delivery
The points you make about the systemic problems--that most T1Ds don't see an Endo--is even made worse because the PCP's they do see self-report that they are not equipped to treat T1Ds because they never had sufficient training. For them, it's easier to put them on an AID system because there's no one whos going to teach the patient how to manage themselves. I cover this in my article: https://danheller.substack.com/p/standard-of-care-primer
The problem I see with your argument is that AID systems are not about perfect predictability, but reasonable control. While it may not be possible to predict chaotic systems long-term, they absolutely can be controlled with fast enough feedback loops. This especially holds true with the development of faster acting insulins, since the needed prediction horizon shrinks.
You can't look at this as simply "insulin goes faster = algorithm improves". The rate of absorption just affects all the other arms in the multi-arm pendulum -- the faster absorption (even if hypothetically possible) only means that the kinetics of *other* hormones now act differently than they were when the insulin was slower. It doesn't make it better or more predictable at all -- it just means that chaotic noise will just have a different kind of chaos.
Thanks for the article!
For me the only way AID systems can improve Is by adding multiple input datas into the equation. Like algorithms that have access to datas like Heart rate, ketones, hormones and stress level, sleep datas, etc. and not only glucose level. Some companies are already working on it.
As the article pointed out, you can have access to each and every variable that goes into the system and still not control it because of its category: Chaos. The parts interact with one another in ways that make them impossible to predict, even if you can measure them. If you can't predict their future states, you cannot control it.
You can't engineer around a mathematically impossible problem.
Yes, it will never be controled. But it maybe can improve a bit the sharpness of the systems. At least thats what I hope :)
The easiest thing to do is simply keep the person in the loop. That’s it. Nothing more complicated than that. Doesn’t even have to be much. The person knows In advance when they’re going to do things. Just announce it. There more they fit that, the better the control and the healthier the outcomes. This constant pursuit of total hands-free operation is not only impossible it’s what’s killing diabetics
I love the chaos theory. But there are different levels of chaos. One is the big pharma convincing doctors that a movable target with many variables such as type 1 DM can be handled by only one hormone that we supposedly can master or manufacture depending on how you want to look at it. The other level is convincing people that blood sugars in TIR is only achievable by more and more insulin. I would argue and have many anecdotal experiences that our total insulin use per patient has gone up since AID because de facto all providers and patients are doing is increasing the dose of insulin now to tighter and tighter TIR
What if there haven’t been updates to AID because the economics of the market don’t incentivize it?
Could there be an attribution error in your analysis? Instead of the impossibility of FCL AID attributed to complex systems theory, perhaps the lack of innovation is a classic market incentive problem?
Here’s the logical challenge I have with the reasoning you’re offering: if humans can make the decisions, then it’s not an impossible problem to solve. Following your logic, then MDIers wouldn’t see better outcomes either.
From your perspective, what am I missing?
there's a lot to this, so let's tease out a few things separately. First, as the article indicated, the physics are such that it's chaos theory -- you can't predict the state of the parts, therefore, you can't control the system. It doesn't matter how much money you throw at the problem, it's unsolvable.
And the researchers working on this know that -- so instead, they're trying to isolate edge cases: safety from hypoglycemia, or maybe post-meal analysis to cut down on higher excursions. Fine, but again (as the article indicated), there's not a lot of room to maneuver in the area, so eking out a few percentage points may be possible, but comes with risk. This is why the looper community tends to do better---they can personally control their own risk tolerance, and take on slight changes to that risk profile that may be suitable to their own personal physiologies, but are not generalizable to the broader population.
And this is no different than how T1Ds have always managed themselves---they learn through empirical experience what works for them. We've always had the freedom to choose how much to dose at any given time. For this reason, the DIY/looper community is and should always be available to anyone willing to engage in that level of algorithm refinement.
And that brings us to your "logical challenge", and you're exactly right: if humans can make the decisions, then it’s not an impossible problem to solve. True, but now we're not talking about technology, we're talking about humans making decisions that are rooted in both empirical experience (which cannot be reduced to math, therefore, an algorithm), but also --most importantly -- AHEAD of time. If they know they're going to exercise, they turn off basal dosing an hour ahead of time. If they know they're going to eat, they pre-bolus. The more they know about their life experience, the more they can draw on those personal heuristics to achieve "better" glycemic control.
And that brings us to your point about MDI users: as the article points out, *engaged* MDI users achieve roughly the same TIR as *engaged* HCL users, demonstrating the outsized role that human decision-making plays.
The point is, you cannot take the human out of the loop. Everything--especially algorithms--work better when the human is in the loop. The more informed and engaged they are, the better *their* decisions are when it comes to dosing, whether it's MDI, or aiding the AID.
It would be interesting to see by how much exactly does the engaged looper have less burden than the MDI user. If you constantly have to tell your pump what to do, doesn't it defeat the whole automation purpose? I mean, by the time I press those buttons and wait for the piston to push in the insulin, I'd already have working insulin after manual injection. And without any painful, unreliable mediators like catheters, cannulas and all the caveats of insulin being pumped to the same spot for days...
What would be the point for anyone to use or develop this then? Surely there are T1Ds in the DIY Loop community developing this for their own benefit, not just commercial "for thee but not for me" type of thing.
It seems like it stems from the trend of digitalizing everything in our lives, even when digitalization doesn't fit at all or is within serious limits, so people just like deluding themselves with a sense that their pump is in control with all those "settings" while in reality it's more of a illusion, because every little setting is constantly changing and you have no way in measuring anything except one - blood glucose reading.
A bit beyond the topic, but reminded me of a line from Epstein interview lol, where he was talking about how math is old-fashioned, not the end of science and how you can't put a number on everything. https://youtu.be/dOSIZzcV2ks?t=5167
It's a bit "all there" at first glance, but also interesting that I use more intuition than math when managing T1D, the only math I use is choosing the number of the dose on a pen and a CGM reading, it seems to work better than anything.
You're right on several counts. This is called The Christmas Tree Effect. (I wrote about this a year or two ago.)
The core idea comes from Leidy Klotz's research: humans are hardwired to solve problems by adding things, and consistently overlook better solutions that involve taking things away.
AID systems are a Christmas tree. CGM, pump, algorithm, Bluetooth, phone app, cloud uploads, sharing features, auto-corrections, activity modes, sleep modes — each one added to "solve" something, layered on until the system is so complex it introduces new failure modes that didn't exist when you were just taking shots. My Performance Paradox article is basically a case study in this.
The sneakier point: people enjoy decorating the tree. The complexity isn't just tolerated, it's the attraction. Tinkering with settings, watching the algorithm work, checking the graphs, feeling like you're piloting a spacecraft — that's not a cost, it's a feature. Which means the normal corrective ("just subtract") runs into the problem that the ornaments are the dopamine. You're not just fighting a cognitive bias toward addition, you're fighting the fact that the additions are genuinely entertaining.
"Feet on the floor" is the hurricane that lifts my car into the next town. Love the imagery.