January 26, 2023

Expert diabetes DIY’er Chris Wilson presents “A New Perspective on Insulin Sensitivity: Insights from a Hybrid Neural Network” at the Fall 2021 DiabetesMine D-Data ExChange.

Good morning everyone and welcome to my presentation a new perspective on insulin sensitivity insights from a hybrid neural network before we get started i do have a couple of disclosures i have filed preliminary us patent applications on some of the stuff uh covered here uh but i am offering it royalty free license for personal non-commercial use so everybody in

The we are not waiting community go for it do with this what you will this is really the uh crux of the research that i did basically that people who live with diabetes know what dose they need in given situations just through years and years of experience and this actually provides some insights that we can use to inform our healthcare professionals a little

Bit about who i am my name is chris wilson i’m on the admin team for multiple uh type one diabetes facebook groups uh primarily device focused i’ve had 24 years of experience living with type 1 diabetes and i’ve been working professionally in information technology for about 30 years uh most recently focusing on databases and architect and cloud systems uh but

I’ve also worked on developing open standards for exchange of recreational water quality data as part of an international team with regulatory agencies and ngos i also uh built the interactive music system and lighting controls for aol’s home of the 21st century back in the early 2000s uh and i am a frequent clinical trial participant uh including having been in

The dexcom g6 acetaminophen challenge trial and the phase three trial for the xeris g voc one of the things that is always fun to do is new research not just confirming things that we already know and so that’s what i wanted to do here and focusing on insulin sensitivity we know that it’s not static we know that as your blood glucose goes up uh insulin is less

Effective so i wanted to look at this and to do that i asked the diabetes online community how do you dose when you’re doing corrections specifically i asked a question given a target blood sugar how much of a correction dose would you give based on all of your years of experience and knowledge living with type 1 diabetes and at various blood glucose levels and then

I calculated back what the effective insulin sensitivity factor was that people were using to uh arrive at these doses even if they don’t know it in their minds themselves and really it’s uh about the vibes we joke about dosing based on vibes nailing a bolus for something that we don’t know how much the carb count is but those vibes actually have validity uh what

They are is the outputs of our the neural networks in our brains that we’ve spent years training whether we like it or not so uh this is the data that i collected uh displayed just in two dimensions uh total daily dose versus insulin sensitivity factor uh and i’ve overlaid a line on there with the 1800 rule that’s currently used for setting insulin sensitivity

Factors in insulin pumps at least initially uh this is from a group of about 50 people uh the age range is 2 to 75 total daily doses between 8 and 125 units a day the big thing that stands out in the sample at least from the demographic data that i did collect is that it is 75 female which probably introduces some extra noise into what we’re seeing here but when

We stretch this data out and we look at it in three dimensions uh it’s clear that uh there is actually a pattern underlying what looks like noise when it’s only looked at in two dimensions a lot of that stuff that looks like noisy data really isn’t there’s uh structure underneath it so i took all of this and i fed it into a simple neural network trained at over

A thousand iterations across this data and then fed known inputs into the neural network and looked at the outputs to determine the underlying mathematical function here is the result of this uh it’s an approximation but it appears to work very well at least in the testing that’s been done so far this is not the actual current insulin sensitivity because that

Changes minute to minute but it’s the average over the duration of the insulin action and it’s actually the fact that our insulins are so slow that allows this to work but it also explains the dosing differences that we see with insulins that are super fast like a fresa hailed insulin where the insulin has its effect before the body actually has a time to adjust

Insulin sensitivity really i know dune’s hot right now but this holds true when we’re thinking about diabetes you can’t understand something just by looking at it as a snapshot in time uh the it needs to be constantly reevaluated so that our understanding can move with the flow of the process and when we look at the individual total daily dose curves if we sort of

Flatten them all together what we see is that it’s actually the inflection point of the of the graph that shifts and the whole line shifts so the slopes of all of these lines are all different and this is probably what has made uh determining overall insulin sensitivity so hard in the past is that the rule changes depending on the individual where you’re doing

The experimentation uh interactions with our current dosing regimens looking at this uh the function that i presented earlier uh intersects with the 1800 rule line at a blood glucose of about 154 milligrams per deciliter that is about where we tell people to test so that actually makes sense uh many uh users of control iq crank up that insulin sensitivity factor

Uh and only correcting at about sixty percent of the correction dose uh which is what we know the algorithm does actually uh brings it down to close to the dose required if someone has set their insulin sensitivity factor more aggressively and we can look at this comparing the the recommended doses given uh based on the 1800 rule that’s the straight black line here

It actually adheres very very closely to uh the line given by this function but then as blood glucose gets higher and higher and higher and people need more and more corrections there’s actually very significant divergence which explains why people get stuck at higher blood glucose levels and get frustrated and wind up eventually rage bolusing and crashing we’ve

Actually been testing this tim street has been uh very kind in uh assisting with uh looking at implementation he shared some initial results from his preliminary testing back in july of this year uh there’s a link to the blog post it’s towards the end uh that he discusses this uh but he saw about ninety percent time in range in uh some of his initial testing and

Uh he wasn’t easy on the algorithm by any stretch of the imagination but this is also no user intervention so the fact that he is very good at his blood glucose management shouldn’t really have been a factor in here so as i said uh implementation in open eps and it’s sort of descendant uh systems uh tim’s been helping with testing and refinement he’s using the

Total daily dose input calculated as a weighted seven day average uh that’s actually not what i’m doing i’m just using yesterday’s uh total daily insulin use but both systems seem to work reasonably well tim has roughly 30 people walking around with closed-loop systems using this right now it has a key benefit in that you don’t need to adjust insulin sensitivity

In the pump profile as the insulin needs change and he’s reporting that people are seeing time in range of about 85 although there has not yet been any formal data collection on this uh these are my results using it sort of on top of control iq adjusting the various modes that the algorithm can operate in to get it to behave more closely to what this new math

Says that it should this is a 90 day uh includes lots of big insulin sensitivity shifts due to uh experimental meds that i’ve been on uh but this 94 time in range uh is actually a improvement from the baseline which at the beginning of this year was about 88 so we’ve actually cut highs in half and almost no readings below 70 at all to do this what did i use most

Of the stuff is free it didn’t really cost a whole lot of money it just took a bunch of my time i did all the data collection with google forms uh organizing and collating the data in google sheets jasp which is free statistical analysis software for the mac uh tensorflow is free software i did uh rent an amazon machine learning instance to run the analysis

On for a total cost of 11 cents uh and there’s a couple of other tools that i use that are either free or very very low cost as well but cost is not a major issue to doing this kind of research and really uh when we think about what this means for the we are not waiting community at large uh it means that we can gain insights from tactic from tapping the hive

Mind that we have in the diabetes online community we’ve got lots of individuals that are out there running daily experiments on themselves refining their doses uh figuring out what works and what doesn’t but those vibes the outputs of the biological neural nets we can then combine into a meta-analysis of their output and use that to discern useful insights even

If the data are noisy we can sort of overcome that noise with sufficient quantity to identify underlying patterns one of the big things i think that this presents an opportunity for is focusing on unknown knowns things that we know but may not even realize that we know and contrasting those to things that we know that we know the known knowns the known unknowns

The things that we know that we don’t know and then even farther out the unknown unknowns the things that we don’t know that we don’t know but really it’s this fourth category that presents an opportunity to gain new insights as to uh how diabetes interacts with our bodies let me zoom in a little bit more uh if this relationship between blood glucose and insulin

Sensitivity holds as it appears to at least so far uh and i will note that yeah we’ve got the the numerator uh of 277700 uh 277 thousand seven hundred that’s what the neural net determined is probably not exact it is very close to e times 10 to the fifth which shows up in a lot of functions used to describe biological processes but we’ve still got open questions

Uh does the carb factor scale like the correction factor does and do basals affect things uh does this numerator change slightly if someone has a different basal bolus ratio than roughly 50 50 that seems to be the standard assumption but regardless of that it does give us at least a mathematical framework with which we can evaluate glycemic impacts basically

Flipping the traditional insulin clamp where we’re infusing glucose to match the insulin on its head where we’re now infusing insulin to match the glucose but mathematically assuming we have accurate tracking of the insulin on board and blood glucose over time we can calculate and evaluate uh the effects of that insulin and therefore the effects of the glycemic

Impacts of whatever it’s countering to enable fully remote studies if we can capture all this data we already have great cgm data it’s getting better i have yet to find an easy way to record constant iob in five minute increments although i’m sure somebody’s got this data somewhere if not it can probably be recalculated from the insulin delivery data and then

Monitored over time but thinking about this what if we could actually run a study that looked at how to successfully dose for pizza what is the most successful strategy similarly what about beer uh things that we would not necessarily study in a clinical setting uh but that people do encounter in the real world and do make real differences in people’s control

Of their diabetes uh obviously we need ways to organize collate and normalize the data binary or probability based outcomes work better neural nets really like percentages if things can be reduced to percentages that makes the analysis that much easier but the ultimate question is is there an underlying pattern underneath whatever it is that we’re looking at

That a neural network can find so at prerequisites to sort of make this all work uh obviously we need the accurate real-time blood glucose data we also need accurate insulin-on-board calculations and curves uh i’ve been using eight hours with uh standard lispro insulin uh tim street has been using a seven hour curve with uh loom jeff the uh even faster lisp bro

Uh obviously we need volunteers who are willing to provide data and then the data analysis and collection capacity um we also probably need to think about what are the problems that we need to solve we can obviously identify priorities through surveys but uh if it’s not something we can measure uh then there’s not a whole lot we can do there uh but if we can

Identify problems where there are underlying patterns and that will make a difference in the lives of people who live with diabetes uh this is some place where we can actually have a real impact it would be nice to have support from industry and academia either advisory boards to help sort of guide the research and let us know where we’re doing things right and

Wrong underwriting of analysis tools and software or providing uh irb review such that this the results of anything that we do can actually be published in peer-reviewed journals and then also any neural networks or machine learning capacity that they have to help us tease out the patterns because the data from this will obviously continue to be noisy but the

Upshot here is that yes when you live with type 1 diabetes every day of your life is in fact a science experiment uh we just need to record what we’re doing and what the results are and then from there figure out a way to quantify it analyze the data in mass and actually extract meaningful insights from it thank you uh thanks for paying attention uh i want to

Send a big thank out thank you to uh amy and the old teammate diabetes mine uh obviously tim street uh who has helped with uh testing it and uh implementation of this i’m sure he’ll be sharing details shortly uh numerous clinicians endocrinologists that i’ve bounced ideas off of and talked about this stuff with over a number of years and obviously the greater

Diabetes online community for uh providing the data uh giving me access to the outputs of their neural networks so that we can do this analysis and now it’s time for q a thanks everyone

Transcribed from video
"A New Perspective on Insulin Sensitivity" – #WeAreNotWaiting Diabetes Chris Wilson #Data Fall 2021 By Amy Tenderich