Support of FIAsp insulin curve

[CrushingT1D]

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

[peterlynton]

I'm happy to attempt to add this (new to all of it actually) but cannot find the curve calculation in Xcode .

[bashem]

There is a suggestion on Gitter to just shorten the DIA instead of touching the code.

[bashem]

Class that computes is evidently called InsulinMath

[peterlynton]

@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.

[jeremybarnum]

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]

[scottleibrand]

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.

[bashem]

Looking at the code in LoopKit/InsulinMath, there is no DIA for 2hrs, so the 3h model would be used

switch nearestModeledDuration {
            case TimeInterval(hours: 3):
                return -3.2030e-9 * pow(minutes, 4) + 1.354e-6 * pow(minutes, 3) - 1.759e-4 * pow(minutes, 2) + 9.255e-4 * minutes + 0.99951
            case TimeInterval(hours: 4):
                return -3.310e-10 * pow(minutes, 4) + 2.530e-7 * pow(minutes, 3) - 5.510e-5 * pow(minutes, 2) - 9.086e-4 * minutes + 0.99950
            case TimeInterval(hours: 5):
                return -2.950e-10 * pow(minutes, 4) + 2.320e-7 * pow(minutes, 3) - 5.550e-5 * pow(minutes, 2) + 4.490e-4 * minutes + 0.99300
            case TimeInterval(hours: 6):
                return -1.493e-10 * pow(minutes, 4) + 1.413e-7 * pow(minutes, 3) - 4.095e-5 * pow(minutes, 2) + 6.365e-4 * minutes + 0.99700
            default:
                assertionFailure()
                return 0
            }

[jeremybarnum]

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.
Example for loop DIA = 4, all the other DIAs are in the spreadsheet.

a^4 2.58154E-10
b^3 -5.56237E-08
c^2 -8.06366E-07
d -0.004278433
intercept 0.987630373

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.

FiaspCurve_part1.xlsx

[tim2000s]

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.

[jeremybarnum]

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.
I want to experiment with some parameterization to allow the user to create custom shapes based on DIA test that a person can do for themselves - essentially adding some other "dials" to drive curve shape beyond just the DIA.

[peterlynton]

@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

[jeremybarnum]

@upsetter looks like you are missing a negative sign in the last term, so the linear term before the intercept? should be -.004278...etc.
Also a few additional things:
-please be careful as I haven't been able to test it. It's just math, trivial change, should be fine, but...
-I assume you were just asking about DIA=4 as an example. If you're going to go through the trouble, you might as well change all 3 DIAs, so that if you change your DIA to something other than 4, it produces reasonable answer.
-if you're willing to do the experiment and are changing the code anyway, I would argue for doing a personalized curve shape test. Recall that everything I did presumes that the existing Novolog curve shapes are right, and that the academic data is right. It's so easy to generate a new polynomial based on your actual experience with the insulin, and update the code accordingly, it's probably worth it...of course the experiment is a pain, and arguably you would want to do it under a few different conditions, and average the result, and you should definitely read @tim2000s blog if you haven't already, but anyway. I was working on a template to facilitate the generation of the polynomial coefficients based on an experiment, I'll try to finish that and upload it. I'm hoping we can do the experiment in my house too, I'm just waiting for the right conditions.

[peterlynton]

Thanks Jeremy, I see the missing negative sign in the last term. I will change the three DIAs at once as suggested.
This is a trial and I will be careful. I appreciate that the curves have not been tested in the wild but I'm a willing guinea pig! I will look into a personalised curve shape test, I'm not sure what's involved yet.

[jeremybarnum]

CustomDIA_v0.1.xlsx

Here is the custom DIA template I mentioned for the personalized curve shape test. Should be pretty self explanatory. Any views/feedback welcomed.

[dm61]

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.