Welcome And Guest Introduction

David Klonoff 0:03

Welcome to Diabetes Technology Report Podcast. I'm Dr. David Klonoff. I'm an endocrinologist at Mills Peninsula Medical Center, part of Sutter Health in San Mateo, California. I'm here with my co-host and with an entrepreneur who has a very interesting idea. I will introduce my co-host who will get things started.

David Kerr 0:35

Thanks, David. Hello, everyone. Once again, David Kerr speaking to you as usual from Santa Barbara, California. Today we have Amir Hayeri from BioConscious. Welcome, Amir, to Diabetes Technology Reports Starts. We're very keen to hear what you've been up to. But before we go there, I always like to find a little bit about the person behind the idea. So,

From Treating Disease To Preventing It

David Kerr 1:02

how on earth have you ended up being interested in diabetes technology?

Amir Hayeri 1:08

Thank you, David. Thank you, David, for um inviting me over. Well, let me take a step back. So um I was uh I was raised by two very loving doctors, and my family's been involved with healthcare for generations, and they really, really wanted me to follow in the family footstep and become a doctor. And unfortunately, to their uh surprise and much disappointment, I uh I didn't become a doctor. My passion was with the rest of my generations. And for some reason, we like machines too much. I don't really know why. So when I was much younger, I was uh I would ask my parents a question, and this would deeply unsettle them and send them off into a rage rant, and sometimes they would rap um about it. And the question was kind of like, I don't understand why you're spending all your time treating when you can be preventing this. And this was deeply unsettling to them, and I would get you know different iterations and answers. And so um, you know, after I was done with my education, that question actually became the uh sort of the motto for bioconscious. We um strive to deliver um, you know, prevention as a service. We started very small in something that we thought was achievable in terms of learning people's glucose fluctuations, modeling them using different types of machine learning, and then using those models to predict what is going to happen ahead of time. And so we thought that would be a very interesting place to start. And then obviously the idea has morphed and evolved over the years, but um we thought that would be the best place to start. Um, and you know, I had a good friend of mine who was type one, he was patient zero that we tested the you know the original algorithms on. And type two also runs pretty rampant in my family because of our ethnicity and genetic background. And so, and the rest is history. Right.

David Kerr 2:58

Okay. So, what is the fundamental problem you're trying to solve at BioConscious

Solving For Biology Not Workflow

David Kerr 3:04

when it comes to diabetes?

Amir Hayeri 3:06

That's a great question. So that there are a lot of there are a lot of like, you know, uh diabetes vendors around, they're remote patient monitoring vendors outside, um, and they're all trying to solve for uh staffing, for remote patient monitoring, for workflows, which are crucial. The difference between what they're doing and what we're doing is that we're solving for the biology. So um early on when um we got started, the sort of bet was, and this is again 2017, 2018, none of the sort of like stuff that exists today that we take for granted existed back then, including AIs, CGMs were not as popular, not as robust as they were. So, what we started off was, you know, we wanted to start very small and show that there is a proof of concept. So in 2017, we started at BC Children's Hospital. What we wanted to show there was that a machine learning algorithm can learn and model uh the user's glucose behavior without knowing exactly what is affecting its behavior to a point where the model can then be used to predict 60 minutes ahead of time in terms of what's going to happen. And um we're able to show that, you know, we were successful more than 90% of the time in terms of predicting future glucose values. And the way we assessed success back then was that we would make a projection or a prediction in terms of where the user's glucose is going to be, wait an hour, wait for the CGM to report the value and compare the two values. If the projected value was less than 10% apart from the real value that was being reported by the CGM, we would count that as a hit. If it was more than 10% apart, we would count that as a miss. And we showed that we were, uh, if memory serves well, 94% accurate. And we did it for 10 teens and preteens, and it's extremely hard to get those algorithms to work for teens and preteens, uh, as as you know. And that was the start of it. So we then launched this um as a user-facing app. So um we it was first it was called Tocto, then it was we renamed it to Diabits. Today we refer to this as the EndObits Companion app. It's one of the earliest iterations of our algorithms. And the way it works is that basically has a bunch of general models. It has seen a few user data points in terms of how glucose fluctuates. And when the new user downloads the app and uses the application to see their future glucose values, we have a learning period where we take one of the general models, we re-optimize,

Proving 60 Minute Glucose Prediction

Amir Hayeri 5:35

and then once the learning period is done, you would see the tail of the graph growing. So we can now make predictions in terms of where the user is going to be. And every time there's a new CGM reading, it gets recalculated. But what we learned when we launched that on the market was that, well, yes, predictions were possible. They're getting better and better, but without a clinical workflow around it, it didn't really have a whole lot of impact. Yes, it was useful for people to better manage their glucose fluctuations, if you will, but it required that clinical workflow around it. So we built a clinical component around that that integrates CGM information, brings in EMR EHR information as much as it is mineable. Uh, there's a lot of stuff there that we can talk about. But assuming the data that we're getting is good quality, what we do then, which is different about us, is that we put the two together. And instead of trying to help the clinic for workflow or treatment, we try to turn that information into foresight. So for instance, if you have 5,000 patients, they're all wearing CGMs. Um, you know, your user base is a mix of type ones, type twos, and type twos are different types of treatments, some or medication, some are taking insulin. Um, from a clinical vantage point, we really don't know who needs help and what do I need help with. So we get all that information. Um, we create these clusters based on user data coming in. Um, and these clusters are highly indicative of like, okay, from this group of 5,000, which one of them needs your help today, right now, and what do they need help with? And that process is extremely tedious. It takes a lot of time and effort for the clinic to get all this data in one place, try to manage it. And so we get that all in one place. It takes us a fraction of a second to do that, and we figure out who needs help, what do they need help with, and that information is passed to the clinic. So you know ahead of time who's gonna be your unscheduled ER visit, who's gonna be your unscheduled sort of patient that's gonna give you the call. So that's kind of like what we solve for. I don't know if that answered your question or not. That was a very verbose way of answering that question.

David Kerr 7:42

Yeah, it that certainly answered my question. So I'm just trying to visualize what this looks like. Um you can take all coma, all types of diabetes, all ages, genders, ethnicities, everything. The more you're quite comfortable with the model in terms of these predictions. What are the clinicians

Turning CGM Data Into Clinical Foresight

David Kerr 8:05

saying to you about it? Are they going, waha, this is the best thing I've ever seen? Or thank you very much. We'll take a look at it. What's what's the general feedback you've been getting?

Amir Hayeri 8:13

Well, a mix of all the above. And so the first thing that comes up is liability, is like, well, now I have to react to all this information that's coming in. And there's already deluge of information coming in from CGMs, right? How do how do we deal with that? And so that's that's the first thing. It's like, listen, it's the same information comes to you, but it takes you much longer to go through it. This is, think of this as a risk mitigation system. Things that need your attention bubble up to the top real fast, right? And in terms of how you want to look at it, you want to look at it every week, every other week, or whatever, you will get to your high-risk patient that's going to create most of liability much faster. So that would be the very first thing. The second thing is everybody is always afraid of, oh, you're trying to replace a doctor with an algorithm. You'll be successful. My parents were very, very like adamant about this. You will never get to replace a doctor with an algorithm. But the point we're trying to make here is that our goal is not to replace a human with an algorithm or vice versa. The point is to put the two together to create a third entity that is far more capable, far more reliable than each of its parts. So, for instance, what do we if I were to use this example, we want to give the staff members that we have superpowers? What if you as a human could do the same thing? What if you could triage your entire patient population that's wearing continuous glucose monitoring devices in a fraction of a second? Wouldn't that help? I mean, I'm sure it would be, right? But that that would be the first thing. So the first thing is once we get past liability, is like, how do we interact with it? How do we trust it? And as we deploy more and more of these machine learning and AI sort of solutions into our everyday stuff, I believe explainability is a huge thing. So we design the user interface. So when you get recommendations as to this is your high priority patient list, why are they high priority, right? And the system provides explanations as to why it has come up with those recommendations, right? And the goal here is to get the clinical staff to interact with machine-generated labels at the clinic level and then assess whether they find those labels, those clusters or groups of patients useful in treatment in faster triage or not. And so far, um, it takes about, like when we deployed at clinical levels, it takes about two to three weeks for people to really get familiar, get a hang of it. But beyond that three-week period, like people would click and read the explanations. Beyond that period,

Liability Fears Trust And Explainability

Amir Hayeri 10:44

what we see is that people end up actually trusting the algorithm. But it takes a while. And then obviously we have to explain to them like, listen, we're not here in the business of replacing you. We are here in the business of making it easier, right? That's kind of the assessment.

David Klonoff 10:59

Uh, have you published your results? It sounds like you have good results.

Amir Hayeri 11:04

Oh, we have published some, yeah. So every year prior to COVID, we would be at the ADA. That's how Jerry and I, uh, our chief medical officer, Dr. Jared Fisher and I met um initially. We're presenting the poster at the after the work was done at BC Childness Hospital. Post-COVID, it turned for a couple of years into this hybrid weird format. Um, we didn't really do it, but there are third-party assessments of our work out there. Um and recently we've turned into just like um white papers. Just publish the results on the website. And if anybody wants to take a deeper dive, reach out to us. We're happy to share uh some of our data.

David Klonoff 11:39

Do you have plans to publish in peer-reviewed uh journals?

Amir Hayeri 11:43

Yes, we have a couple of interesting in studies in terms of moving the science of uh what I refer to as dysglycemia further. Dr. Kerr is aware of one of those studies. We're working um with him on that, especially around gestational diabetes. And then another one, which we call the MUNS Plus, and then there's another study that we're working on for patients that are showing up, what we refer to them as HBA1C negative. So if you do an A1C, they might be slightly elevated. But in fact, they have different types of dysglycemia that we believe it is invisible to the naked eye, but visible to the algorithm, to our clustering algorithms that we can do. So, yes, absolutely. Moving forward, we're looking at it.

David Klonoff 12:23

If someone doesn't have diabetes, they choose to use a CGM and use your service. What types of clusters are you identifying? Do you have names for each one?

Amir Hayeri 12:34

Yes, there are things I can share for sure, which is which is interesting. And there's some stuff that we're working on and I can't I can't share. For instance, what we were focused on up until this point was that we wanted to show that this methodology works. And um, the assessment here is that we are trying to teach the algorithm, not really, oh, here's type one diabetes, here's type two diabetes. That's how we got started. But as as you're aware, there are instances where you know late-stage type two and type one are exactly similar if you look at their CGM charts. So the software got easily confused. And when we switched over to using different types of neural nets, we stopped imposing our own understanding on the algorithm and instead waited for the algorithm to extract hyperparameters that we thought were kind of useful. And so when what it's trying to hone in on is different types of glucose

Publishing Plans And New Studies

Amir Hayeri 13:22

curves, right? Or what we refer to as like glucose curves that are representative of different types of dysglycemia. And we use different types of machine learning techniques to figure out clusters that represents this, represent the same biology, but from patient to patient, they might be compressed, they might be longer, there might take much longer to happen. So we use like techniques such as dynamic time warping and other things. So sometimes when you look at curves, they all represent the non phenomena or they're nocturnal hypoglycemic. But from patient to patient, there are variations. And the software is, you know, being trained to get really, really good at detecting these different types of variations that are all representative of the same biology in the background. And we are clustering them together. And in order to check that work, we've started in two different separate ways. So we use machine learning to cluster patients together. We focus on diabetic patients only, so type one and type two originally. And we were able to show with a third party recently that in diabetes, you can easily figure out who is going to be what we refer to as like high touch patients. So these are roughly 20%-ish of your patient population that would require, require 80, 90% of your attention, right? They need a lot of help. And so we can do that. Then we can figure out, okay, who are who are going to be your laggards, who are going to be the ones that are going to have, you know, certain type of events happening more than others. And so we can really tinker around with it. But what we're trying to do is to, again, as I said, really hone in on the dysglycemic part of it or dysglycemia and build a map, what we refer to internally as a glucose atlas. So, you know, when somebody, as you mentioned, goes to the drugstore, buys a stello

Clustering Dysglycemia And A Glucose Atlas

Amir Hayeri 15:10

or a Libre Lingo, if you will, and wears it, um, yes, that's really good. Uh, you know, it tells them to eat their vegetables and go for a walk after they have a meal, but there's not really a need that's being satisfied there. And I think in in a few short years, we would be able to assess that um reading, be it for 10, 14 days, figure out what level of dysglycemia is present and what is the window of horizon that the patient has. And this could be, you know, you're you're becoming, you know, insulin sensitive or not sensitive enough, or you're moving towards diabetes. But what we have seen is that when we put a CGM on everybody, there is a level of dysglycemia. The lap tests that we have, such as A1C, are not sensitive enough to pick it up. And when we see even within people who are non-diabetic, I was one of them, we still suffer from dysglycemia, but it is episodic. It's not continuous, right? So there are in at certain hours of the day, be it because of the meal or sitting down for too long and whatnot, we would have dysglycemia and different levels. And over time, it would just, you know, take us to, you know, closer and closer to diabetes. But diabetes in and of itself, we view it as like a gateway, but it's not real, it's not a destination, right? It's just a stop to another chronic condition, so to speak.

David Kerr 16:30

Thank you. I mean, just to wind things up here a little bit, um this is very glucosentric, and there's nothing wrong with that. Um but in this world of the quantified self and so on and so forth, what's your what's your advice to people about wearing a CGM plus? Is it do they just only need a step counter or a sleep device or a combination of them? And going forward, what would be the ideal set of devices?

Amir Hayeri 17:06

No, that's that's a great question. So, right now, what we have are different types of algorithms, right? So there's one core algorithm, if only CGM data comes in, we assess that and we make predictions off of it. But if you have other types of information, so for instance, if your carbs, insulin, or exercise information is available, we can take that into account and there's a different model that runs for that one. Um, if your EMR EHR data is available, there are different models that will run uh as well. What we have seen though, right now, and this is a limitation on our pre-processing and also our ability to assess different types of data coming in and assessing if it's helping the algorithm converge or diverge. So it was very easy for us to get to a point where we have a core algorithm, it's outperforming everything based on CGM data only. Then we start when we start reintroducing other variables complementary, such as EMR EHR information, such as step count, such as other devices, we often see the algorithm gets

CGM Plus Wearables What Helps Or Hurts

Amir Hayeri 18:06

confused. Um, and so performance drops, and there are different areas, and I can get into, I can talk about this until the sun burns out. We believe the limitation right now is in twofold. One is sometimes the accuracy of the data that's coming in from your cell phone, your aura ring, or other things is not highly accurate. You know, again, CGM is a medical device. You know, your cell phone, your Aura ring, these are not medical devices. So the data that we're reading is not as high quality as those are, so it requires more pre-processing before that information goes in. That's one. And then two is figuring out, you know, what we have learned over the years is that when we have larger models, and the model in and of itself is keeping track of a bigger number of variables, oftentimes performance drops. So the way we're trying to solve that is by having, you know, specialized models, you know, one model that's really, really good at processing CGM information, processes that information, generates metadata, concatenating that model with a different model where the inputs and outputs uh sort of work together to provide it. But in terms of like a clear answer and a silver bullet right now, it's a work in progress. I can't really that's good.

David Kerr 19:19

That means that means you're gonna be gainfully employed for the near future anyway. Thanks very much.

David Klonoff 19:25

Amir, this was an interesting interview. Uh, I hadn't known about your company, but I'm gonna keep an eye on it. I think BioConscious has a lot to offer, both uh clinicians

Closing Thoughts And Where To Listen

David Klonoff 19:36

and patients. So uh I want to thank you, Amir, for being uh a guest on our Diabetes Technology Report special section called Diabetes Technology Starts, and also thank my co-host, Dr. David Kerr. Um, thank the audience. This podcast will be posted on the Diabetes Technology Society website. It's also available through the Apple Store and Spotify. And until our next podcast, everybody be well and uh take care of your diabetes. Bye, everybody. Thanks for having me. Take care. Cheers