Support of FIAsp insulin curve #388
Comments
|
I'm happy to attempt to add this (new to all of it actually) but cannot find the curve calculation in Xcode . |
|
There is a suggestion on Gitter to just shorten the DIA instead of touching the code. |
|
Class that computes is evidently called InsulinMath |
|
@bashem It's early days so happy to shorten the DIA for the moment, but it may be better to generate new code for the curve as FIAsp has a very quick initial absorption rate and is currently a bit of an unknown on high doses. You're correct regarding InsulinMath, which is based upon Glucodyn code written by Ken Stack https://github.com/kenstack/GlucoDyn. I've messaged him on Gitter to see if he can help but otherwise we need someone who can convert 4th-order polynomials to algebra. If we're able to generate new algebra more suited to FIAsp I foresee it initially being used / tested by end users editing the code, with the potential of adding it as a secondary option to Master as FIAsp becomes more available worldwide as @CrushingT1D suggested. |
|
I've spent a couple of days playing with this. While I agree with the comments that shortening the DIA is a good initial approach, the seemingly much higher insulin availability of Fiasp in the first 30 minutes does mean that a new algorithm would probably be a good idea. I'll try to synthesize what I've done and upload it - it's just the math part (not that fancy - just using LINEST() in excel to match the data in the graphs of the PK from the studies), not the code. It is clear that the current approach in Loop is a bit ugly - 3 different fourth order polynomials for DIA = 3,4,5 and then the InsulinMath does a linear interpolation of the IOB depending on the DIA input by the user. It's relatively simple to take the output IOB from those functions, manually generate a new IOB curve by taking the "treatment ratios" (AUC(t) for fiasp/AUC(t) for novolog) from the Fiasp studies, and do a new fit, which I've done. But if we're going to do this I wonder if while we're at it we should look at [https://github.com/LoopKit/LoopKit/issues/90] at the same time. @scottleibrand explains here [https://groups.google.com/forum/#!msg/openaps-dev/w_YNAmaguoc/0R00sksCU4wJ] how the oref0 approach works - I want to try that for Fiasp, using the earlier peak and generating an IOB curve using with that approach and see what treatment ratios that produces. Useful references: [http://onlinelibrary.wiley.com/doi/10.1111/dom.12803/full] [https://www.hindawi.com/journals/cmmm/2015/281589/#B18] |
|
Heh, that is an old thread. But as surprised as I would have been at the time to learn it, we're still using that approach in OpenAPS, and rather successfully. The curves we use have peak activity at 75m for a 3h DIA, and all other DIAs are simply stretched from that. So 4h peaks at 100m, and 2h peaks at 50m. Because of that peak time, @tim2000s has been using the OpenAPS 2h-DIA curve with Fiasp, and rather successfully thus far as I understand it. |
|
Looking at the code in LoopKit/InsulinMath, there is no DIA for 2hrs, so the 3h model would be used |
|
Well here goes nothing. I tried to create new curves for both Loop and OpenAPS. The work is in enclosed spreadsheet. I tried to document everything clearly. Eager for any feedback anyone has. My conclusion is that assuming we believe that Fiasp starts hitting a lot faster, as Novo Nordisk claims and as people have reported, shortening the DIA as a way to approximate the curve shape is better than nothing (and as mentioned seems to be working fine for some). But that approach would appear to involve compromises between matching activity in the first 45 minutes versus matching the tail. And since the whole point of Fiasp is to avoid post-prandial excursions, it seems worthwhile to consider adjusted curve shapes. a^4 2.58154E-10 I am doing some more work on the OpenAPS curve to see if we can get a "purer" curve that is produced from user inputs that are intuitive and testable empirically, and does a good job of matching - in line with what @scottleibrand said about the original thinking behind the OpenAPS curves. I'll post that shortly. |
|
I'd be careful on this one at the moment. I started off with DIA of two hours being very effective, and effectively almost no tail, but I've since seen some changes to my body's absorption characteristics with fiasp and until I'm sure what's going on, I'd be loathe to say that the Novo curves are 100% effective every time. By all means do the work but be wary of this factor, as I'm not sure I'm seeing the same level of front loading on the absorption. |
|
Thanks Tim - yeah for sure this is obviously to be treated with great caution. Mostly I wanted to establish the framework - but it needs to be validated with experience clearly. And as you point out, it's all based on the published "Treatment Ratios" showing ratio of AUC between fiasp and novolog at different points in time. If those are wrong, then my curves will be wrong. And if the starting curves are wrong before getting transformed, then the new curves will also be wrong. |
|
@jeremybarnum Thanks so much for the FiaspCurve excel file - I am aiming to trial it. Am I correct in editing the 4hr line to the following: return 2.58153954561464E-10 * pow(minutes, 4) -5.56237306749063E-08 * pow(minutes, 3) -8.06365857215448E-07 * pow(minutes, 2) 0.00427843253655648 * minutes + 0.98763037314288 |
|
@upsetter looks like you are missing a negative sign in the last term, so the linear term before the intercept? should be -.004278...etc. |
|
Thanks Jeremy, I see the missing negative sign in the last term. I will change the three DIAs at once as suggested. |
|
Here is the custom DIA template I mentioned for the personalized curve shape test. Should be pretty self explanatory. Any views/feedback welcomed. |
|
Based on Novo data shown in Figure 2 of this document, I fitted exp curves suggested by @tepidjuice. A general form of the curve is: (scale)*(t^n)exp(-nt/tau), where tau = time of peak insulin activity, and n is a parameter that may be used to adjust DIA. Taking the underlying system dynamics into account, n is expected to be 1 or 2. For Novolog, tau = 75 minutes, while n=1.5 gives a better fit to the data than 1 or 2. For FIASP, tau = 55 minutes, and n = 1. The results shown below include 1. a comparison of the data and the curve fits, 2. a comparison of insulin activity curves normalized to AUC=1, and 3. a comparison of the corresponding IOB curves. |
|
That is really cool @dm61 -- seems like it achieves all the goals at the same time: simple, intuitive functional form with limited number of variables; good fit to published data; and (presumably) ability to easily fit individualized empirical experience as well. Consider me persuaded! If this is the way forward it would also be worth comparing to the current loop Walsh curves so that people know what to expect. Notably, the published data suggests activity all the way out to 6 hours, whereas I believe most users are using DIA between 3 and 4. So perhaps the most appropriate comparison is to compare the output of the Walsh curve for DIA = 4 to a curve that tries to fit that output with appropriate values for n and tau. I'll have a try tonight if I get a second. |
|
Yes, I need to see a comparison to the Walsh curves, like I generated for the simple exponential version of this, preferably normalized by AUC, as you've done in the two top graphs. Another remaining issue is truncating the tail of the decay curve. It's a bit sharp to just lop it off, but maybe that's ok. Might make small kinks in the forecast graphs. Finally, depending on how we truncate the tail, I'm thinking we'll want to have actual insulin activity out to 5-6 hours, which is probably a reasonable choice for those who had 4 hour DIA set before. The tail is smaller in effect intensity, and should go out a bit longer. |
|
Do you have the data to play with somewhere? The curve in the Novolog
curve doesn't very well cover the first 10 - 15 minutes of non-activity.
You might want to play with shifting the zero point by this time.
…On Thu, Jul 6, 2017 at 8:30 PM, Pete Schwamb ***@***.***> wrote:
Yes, I need to see a comparison to the Walsh curves, like I generated for
the simple exponential version of this, preferably normalized by AUC, as
you've done in the two top graphs.
The other remaining thing is truncating the tail of the decay curve. It's
a bit sharp to just lop it off, but maybe that's not an issue. Might make
small kinks in the forecast graphs.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#388 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AAE2arBrYVGIPntPla5-Jwwnk2F387C6ks5sLSfdgaJpZM4MPcQq>
.
--
Jan Dittmer <jdi@l4x.org>, http://l4x.org
|
|
The charts actually show activity by 10 minutes, and I do observed effects at that point for us as well. |
|
I meant for Novorapid. Talking about this corner of the graph.
…On Thu, Jul 6, 2017 at 10:10 PM, Pete Schwamb ***@***.***> wrote:
The charts actually show activity by 10 minutes, and I do observed effects
at that point for us as well.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#388 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AAE2al87uFy2fMsucNxuknjpffUxRZoDks5sLT9FgaJpZM4MPcQq>
.
--
Jan Dittmer <jdi@l4x.org>, http://l4x.org
|
|
Yes, the novorapid graph in the cited paper shows activity by 10 minutes. That's what we use, and I do see activity at that point usually. It may need to be moved over 5 minutes from where it is now. But you can see that even though the activity seems to shoot up fast in the activity curve, it also looks like were still at > 99% IOB at 10 minutes in the COB decay curve. |
|
Here is a comparison of the exp curves to the Walsh DIA=4h curves. |
|
One idea for dealing with the tail is to shift the entire activity curve downward by the amount of the remaining activity at DIA, and then normalize the resulting curve to AUC=1. That would also end up shifting the early activity slightly right (by subtracting a constant, you end up truncating to activity=0 the first few minutes when activity is lower than at DIA), which might make the early part of the curve more accurate. |
|
Here is a zoom-in plot of the above IOB curves, to see how much error in AUC results from just chopping off the curves at a certain point in time. |
|
So if we lop at 6 hours, we'll be chopping off a little more than 1%. Thanks. |
|
Looks promising. I'm travelling at the moment and don't have access to my computer but I have some questions/comments about this. I'd like to try my own fit when I get time. Would someone be able to post the FIAsp and novorapid data? |
|
For reference, attached are excel files with insulin activity data for Novolog and fiasp (digitized by @elnjensen). |
|
Ok, @dm61, can you give me the integral (IOB decay) of your curve? :) That is ultimately what Loop uses. |
|
Thought you'd ask for these at some point :) Novolog: FIASP: These functions return fraction of IOB remaining after t minutes, just as the Walsh polynomials currently do in InsulinMath. Fortunately, looks like swift Foundation includes erf function. |
|
So I've been running the Loop insulin model testing branch #539, with exponential curves, td=360 and tp=75 (Novolog) for couple of days now. This morning I assumed the combined roles of a lab technician and a lab mouse, and performed a semi-controlled experiment: took couple of glucose tablets to bring bg to around 150, waited a bit to get to steady state, and then dialed a 1U correction bolus, while keeping Loop open, and with no food or exercise for the next 5 hours. Below are screenshots illustrating the results. BG started to drop noticeably at around 30min, which seems to be well predicted by the model. The most important graph to look at is Glucose Change (which shows insulin counteraction effect). Ideally, if the modeled insulin absorption curve were perfect, that graph would show zero at all times after the correction bolus. As shown below, not exactly zero, but the values are relatively small and bounce around zero (which is to be expected given bg sensor noise and quantization). A slight negative bias may imply that my ISF or basal rate are a bit off, but that's a different topic. |
|
@dm61 That graph is beautiful! :) Our insulin counteraction effect chart seems to have cleared up a lot since running the new curves as well. |
|
@dm61 Nice experiment! |
|
@tepidjuice, @francesc0-cgm insulin clearance in children under clamp conditions has been shown to be more or less the same as that in adults, and that is >5 hours in both cases. That's why My concern comes about modelling the activity and iob and the relationship between them. There's an additional component that we need to determine given the data we have for half life and activity. |
|
To add to my last comment, (sorry on phone with no edit function) - the physiological mechanism for insulin clearance is a combination of insulin activity within the tissues that have insulin receptors (so typically GLUT4 in T1 in peripheral muscle) and also kidney activity. There's a small amount of hepatic clearance but that pretty low in T1. I think our IOB based on activity looks solely at the effects based off insulin activity and ignores the renal clearance, hence why the IOB curve has a slower rate of decay than the published data. |
|
Hello, Thank you very much！ But now I have a question about insulin time. I want to ask you. My insulin works faster in the morning, and it can be injected once a day, but the onset time at noon and evening is much slower. It can only be injected first, and the other can only wait an hour after it can be injected. This leads to a closed loop pump system. It predicts high blood sugar. And then press it on the basis of the foundation. The results are easy to cause hypoglycemia. How to solve such a problem. What is the best way to fill in the time of insulin action |
|
Are you using Loop or OpenAPS or something else? Probably best to move this over to Gitter. |
What physical value does the activity curve represent? Is it blood insulin concentration in [mU/L], etc? |
|
Insulin activity is measured in units of mg/dL/minute or equivalent. It corresponds most closely to the glucose infusion rate in clamp studies, which is in units of mg/kg/minute. |
Thanks! And what is the significance of the exponential model curves all having an integral equal to one? |
|
The area under the insulin curve represents the total activity of the insulin, which is calculated using ISF. Having the AUC equal to 1 makes it easy to properly scale it by the ISF. |
- https://tidepool.atlassian.net/browse/LOOP-3375 - Update to latest CGMManager, PumpManager, and Service protocols - Remove special case of pump manager onboarding - OnboardingManager handles created, but not yet onboarded, view controllers
|
Hello, |
|
Pharmacokinetics is how insulin is processed by the system/body, so it is the way to know how much insulin is active and how much is not over time. Is my understanding that this is how the insulin activity is calculated.
Insulin activity does not depend on how much glucose there is in the system, it depends on how much insulin has been given at a given time.
Both are important for sure, but tell different things.
May be a clinical pharmacist can explain in better…
… On May 13, 2022, at 7:00 AM, martinlcz ***@***.***> wrote:
Hello,
I would like to ask why pharmacokinetic curves are used to calculate IOB (Serial mean serum free insulin concentration) and pharmacodynamic curves (Serial mean serum glucose collected) are not used. I think the pharmacodynamics better describe the effect of insulin on postprandial blood glucose levels.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.
|
I understand that pharmacokinetics describe the amount of insulin in the blood, but pharmacodynamics describes how insulin lowers blood glucose, which is of more interest to diabetics. |
|
Do you have a good data source showing how the PD curve differs from the PK curve in a way that would be meaningful for IOB calculations? We used the published curves available at the time, and empirically determined that they worked well, and far better than the crude estimations we (and many insulin pumps) were using previously. There may be opportunities to improve the curve shape further, but I would need to see data indicating how the shape should differ from the exponential approximation we use today. |
|
There are few pharmacodynamic studies. So far, I have found that the peak depends on the size of the bolus (the larger the bolus, the later the peak). Most studies of insulin Fiasp and Lyumjev in diabetics are for 15U. In healthy people 7U, 15U and 30U. The "nichi" developer has designed a dynamic model of insulin activity, but it is quite different from the AAPS model, so I wonder if the pharmacokinetic or pharmacodynamic model is correct. nichi - Lyumjev My version - Lyumjev |
|
Here is a comparison for different bolus sizes. |
|
Thanks for sharing this interesting research @martinlcz , I've found myself many times "intuitively" lowering my bolus past a certain large amount of carbs (I rarely bolus more than 6UI / meal in my everyday) because even though my carb ratio is fine tuned, for a big meal with 50/50 high IG and low IG carbs, I would most certainly fall in hypo 3 or 4 hours after the meal+injection, and it's independent of whether I do an extended bolus or not. I thought I was crazy and wondering if there was any possibility that my carb ratio was nicely tuned for a certain range (say until ~70gr of carbs), and that the more I would eat carbs, the higher it would gradually switch. But I felt it had to do with the insulin activity rather than the CR. |
|
Hi! Sorry to revive this thread, but I'm implementing IOB in my own loop and was curious about something: Why does treatment.insulin reflect the total insulin dose given, and not a curve (eg some step function)? Ideally you would convolve the infusion curve (in U/min) with the impulse response (these exponential curves) to get the overall system response, but it seems like all insulin treatments are assumed to be given instantaneously at a moment in time. Of course, you could take the total dose from an infusion curve and use that as your treatment.insulin, but that seems like it would work less well when your infusion is given over a longer period of time (more than 5 minutes or so). Anyone have thoughts on this? |
|
@ijustlovemath Yes, Loop's models are not physiologically accurate in this respect, and you may improve the very short term modeling of glucose a small amount by considering an infusion curve, but since Loop is mostly looking out several hours for dosing calculations, this will have a very small effect on dosing. |
|
Am I correct in thinking you simply integrate your infusion curves to get the "treatment.insulin" amount? |
|
The treatment curves approximate solutions to a second order linear system
of DEs.
From memory these are normalised so the integral is always 1 over a fixed
time. Allowing the dose to be a scalar multiple of a fixed function.
This isn't really physiological. Absorption is dose dependent. But it's
close enough.
The initial condition is a delta function with magnitude of the bolus.
Given the time constant(s) of insulin absorption are quite large you won't
see much of a difference if you look at the response to a pulse with a
duration of 5 minutes.
For pulses longer than about 10 minutes there'll be a bigger difference but
generally we call that basal.
In summary sure the insulin model isn't perfect but there are much larger
modelling issues in diabetes and it what we have is simple and works well
enough.
…On Tue, 19 Dec 2023, 16:21 ijustlovemath, ***@***.***> wrote:
Am I correct in thinking you simply integrate your infusion curves to get
the "treatment.insulin" amount?
—
Reply to this email directly, view it on GitHub
<#388 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEZQODTN7R3YO6NQD6QNGCTYKGWHHAVCNFSM4DB5YQVKU5DIOJSWCZC7NNSXTN2JONZXKZKDN5WW2ZLOOQ5TCOBWGI4TOMBWGAYA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
|
hello! the question's off subject, but may be someone know how to convert 1 U of insulin to mU/min |
|
Hello! This link is broken. Could you kindly send somehow a working link or article itself. I am trying to understand underlying maths and can't find appropriate articles. Maybe this article can help! Thanks :) |
and others
Add setting to select type of insulin
A) Novo/Humalog
B) FIAsp
If A selected use current curve calculation
If B selected use maths to match curve as described on page 12 of
http://www.ema.europa.eu/docs/en_GB/document_library/EPAR_-_Product_Information/human/004046/WC500220890.pdf
The text was updated successfully, but these errors were encountered: