Thank you so much again. I think you learned the basics of the AI. It was a wonderful presentation so I don't want to be you know repeating the most of the things and then my talk is about how are we using the AI in the diabetes field especially integrating into the diabetes technologies and how the AI is helping us to give a better care in the especially in typhoid diabetes management and my confidence hasn't changed in the last two hours and is mentioned in the AI I think that's there is a confusion in the terminologies in the AI is the machine learning deep learning and the neural networks and other things I think it was a great you know explanation in the from the first talk that AI is an umbrella so the AI is kind of you know includes a lot of different things that we use and then is a machine learning is a subset of AI that uses statistical methods and to you know analyze and shows this data and I will just focus on a little bit more into the deep learning which that we use the like the human neurons think about like a human neurons and then this the neural networks are just making a lot of decisions in in the automation in the AP systems so do we need AI in clinical decision support you know and then as mentions also you know we have short appointment times we have a complex healthcare system and we have different guidelines you know different times and high patient volume new treatment and we have to follow patients for the adherence and we have a limited time space and you know intelligence for all of these things in the healthcare system so we need some sort of an help and I think AI can be a great help and for this all goals so how to be how to integrate this all of this to do clinical diabetes care and AI intelligence you know artificial intelligence driven diabetes care so we have a lot of opportunities first of all we have opportunities and then to create an internet of medical things like think about like a Wikipedia or a Google of you know everything is a medical related and then you are just using AI and for risk assessment and certification and screening classifications communicating with your patients and replying their messages in a faster way is explained you know in the earlier talk also like you know you just send a message to the boats from the Amazon and other things you are just asking your ticket you want to change that and your patients are asking a lot of questions can I take my cholesterol medication can I increase my Lantus this has happened that happened and a lot of things can be done and at least can be some basics that you can just monitor and save a lot of time on that so do we have some challenges absolutely right I think this the first challenge is the performance so how can we know that the ensuring accuracy and general liabilities of the models so you know they are tested in certain ethnic backgrounds and certain countries and other things you know does it work for everything else and safety so a lot of people will be worried about the safety and then managing trust and accountability and responsibility and enhanced and seeing user and stakeholder acceptance and ease of use you know and then you can just see that do you trust AI or not and then how much you'll be trusting and is AI replacing our jobs or not I think a lot of things that people are when we talk about AI people have just have different things they are thinking about and infrastructure not every computer system or anything that is going to be you know available to do that so how we use the AI in the diabetes care is explained also you know multi-sources input data and we get so many data for the from the charts electronic health data and from retrospectively and everything else that you name it so there's a data analysis and there is some clinical relevant actionable outputs that we use and it includes some alerts predictions reports automation and some you know suggestions based on the lifestyle advice some educations and self-management prompts because diabetes care takes a lot of time you know to think about like a type 1 diabetes think about type 2 diabetes somebody's on nutrition you know using some devices CGM's AI the systems it takes a lot of time to teach all of these things and people just have problems and then you know you have to repeat the same things again and again and sometimes then AI can help all of these things you know when you do that so how do we use the AI in the AP systems and the CGM so if we talk about the basics first what is an AP system which I don't like this term by the way but the FDA first approach is an artificial pancreas was the first time it was used in the FDA in 2017 it's actually what we use an AP is an automated insulin delivery system but we like to do call a little bit better a ID and an hybrid close-up system in an HCI so there is a closed-loop control with an artificial pancreas and there's a measuring glucose the CGM and sent to the sensor glucose to the pump and there's a control algorithm embedded in that is the insulin pump and automatically the termites insulin dose and deliver insulin to the patient and so we have different systems that control algorithms that are in the FDA approved pumps or in the you know other European Union pumps right now we have different algorithms that if work in a very different way and we have PID algorithms for the Medtronic in your early generation pumps we have MPC algorithms for most of the currently available pumps in the Europe and the US and we have PID and MD logic for the metering 780 G and some new ones that the combination for the dual hormone studies that we use for this systems so if you look at a little bit briefly so and I will combine that how we use with this things with the AI and this you know we have proportionally integrate derivative algorithms like a PID and we use in the first generations of the Medtronic think about like a temperature control in this room you know if you put the temperature to 23 Celsius so if this temperature is more than 23 and your AC starts and then when it decreases the 23 it stops so think about the same way as the PID algorithms are working the algorithm only you know includes the the difference between the CGM and or your target and then try to use the rate of change and proportional component and also incremental components so it's a basic system that works that way so the other pumps that are using the model predictive control which is a measuring the CGM and then the glucose values and predicts that in a way that either in a target range or in a target value and then gives the insulin amount or decreasing and increasing based on that and with this way is using the adaptive multivariable and nonlinear models for better blood glucose predictions for the accuracy and most of the pumps are using the MPC control and some of them like using a range like the tandem ones and some of them using a target like the OP5 and some of them are trying to combine some way of the you know the basic ones like the PID with an MD logic system which is a fuzzy logic which is the new metronic pump which is a 780 G so how this is working is as I explained that PID just think about like is it cold or hot and it says yes or no and there's nothing in between but if it's something is it cold and then the fuzzy logic system says that you know 90% cold and it's you know 75% hot so there are some in between it's like think about like a rainbow there is nothing like only two colors here so and in this fuzzy logic system is like think about like a PID as I explained is like a timed green or red light in the you know when you are driving is in the fuzzy logic if it's a rush hour it just you have more you know green lights if it's after midnight is nobody's nothing is going on there is no red light if there is one car is coming it turns to red light it's not timed so it's a little bit smarter way of doing the PID when you add on these things. The problem is the systems are that we use there is a big misconception that oh my instrument pump is already I'm using an AID because I am using you know sorry AI because it's an AID and then it's because there is an AI right the word is there so the problem is the current AI the systems we use in diabetes care that lack of multiple users data so it learns from you but you are the only single user and then it's so you are not feeding the system with the you know the millions of people's you know use there is a system it's an algorithm it just gives you more or less based on what you do and lack of computation power and that's a problem just think about like if you want to go from here to United States you need a bigger plane because that's bigger plane has a bigger engine so if you want to do a lot of computations you need a bigger computer so and if you are we all try to get a smaller computers smaller iPads smaller iPhones but we are forgetting that we then we want them to get a great picture and doing things very fast but you need a space and a month and computation to do that so to do that and then this AI can help us to decrease the computation because only can APS pump for example is using the phone as a computation which is the algorithm is embedded on the phone all other AI the systems are using the computation on the pump which is significantly smaller in terms of the space and all the computation so how the AI we can integrate into our AI the systems are is just think about this neurons and the neural networks is the feature of the AI the system that we will be using so in neurons think about like a biological neuron there is an input and output and you have dendrites then there's an axon and there's an axon terminal you give an input and it turns to an output so we do the same thing with the simple neural network then there is an input there is a hidden layer there is an output layer so you can input anything so if we talk about the you know here in the AID management you can put the CGM values CGM trend you know insulin on board and there are so many things you can integrate in here and there is in the hidden network and there is a lot of computation which you know decreases the computational demands and there are output then you can say that okay just give me the insulin dose that you know this person should get or the alerts from the system that can use that so neural networks will be the feature of the you know the next AI the system that's a great study which is a game-changer and so what they did in the University of Virginia in the Kovachev group is they randomized the University of Virginia MPC algorithm and with the neural network artificial pancreas which is a fully closed loop and they did a 220-hour intervention in 15 patients with typhoid diabetes again one is the neural network other one is the one of the most advanced MPC algorithms it requires still a user input and you can see the results are very similar so there is no difference in terms of the time and range in terms of the you know time in tight range variability mean CGM glucose so it's a small but it's a big step and more of this studies results will be available in the ATTD in March in in the 2025 meeting so and then you can see that the neural networks can make decisions you know for the for the AI the systems so if you look at that the meal detection is also possible and then and basically you are just teaching the system that how to recognize based on certain things so decreasing the you know CGM or the increasing the CGM values based on the trend so there are different shapes you are teaching like decreasing shape and steady decrease you know accelerating and decelerating decrease or increase or a constant. And then you're just using this model and in a fuzzy logic system and then other than using a zero or one, you are just using in between things. And in this small study with 117 meals and snacks, retrospectively, they look at 11 patients with typhoid diabetes. So the system was able to detect the meals 93.5% of the time with the snacks 68% of the time. So it gives that amount of the carbs and you can see that in the graph, that actually the real amount of consumed carbs and the system is with the machine learning and is detecting that is actually it's a food and other than something else. And the time of the food is also important. So these are all the prototypes maybe, but this is I think the Pandora's box is opened. And so the AI systems will be using the AI integrated and they will be using it in a different level. So another feature is that Medtronic is working is the hands-free closed loop. So what they did is you are wearing a smartwatch based Medtronic clue application. And then whenever you are eating with a spoon or a fork, and based on that how many times you are eating and your spoon or fork, and the system is detecting and guessing that your carbon hydrate and other meal amount and how much you eat and giving you the amount of the insulin without any other input at all. And if you see that it's five days each, in the 17 people with typhoid diabetes, it's a small study. And they compared the advanced hybrid closed loop, which is a 780G in current form that with the carb counting. And they compared with the clue app with using the advanced hybrid closed loop without any other input. And you can see that the results are very similar. So and then the systems are detecting if you are eating with that smart app and watch. Obviously you have to use that hand if you are eating from the other hand, the system is not going to detect it. And then I think things are getting smarter. And then we are using in these things and these things can help for our patients a lot. And then just think about the quality of life that this person is going to have in terms of that not doing any carb counting and just getting the exactly the same outcomes. And sometimes even if you're not increasing the time and range or the hemoglobin A1c, if you are decreasing a huge burden in diabetes care, I think it's a gain. And I always think about that way. So we are always focusing on too much for the numbers and other things. And I was like, oh, this the time and range is the same or time and range is, and then a little bit less, but these are not different results and you are not doing anything basically, just to think about it. And how we integrate this to our AP systems and it's a fully multivariable, you know, the AI, the systems, and these are not in the hybrid close-up anymore, they are fully closed loop. Then if you see that there's a study they compared the glucose only, the AI, the system to the multivariable fully closed loop system. So basically in the glucose only AP system, we only use CGM, right? So the only there's input is coming from the CGM and the value and it's going up or down and based on the trend. But if you use some other variables, like your heart rate, you know, other like your temperature and your menstruation, and then all kind of so many things, then you can just predict the hypoglycemia and you know, hyperglycemia and exercise detection and exercise a very individual thing for the people are always complaining about the AI, the systems, it works or not. I think that's another way of doing that. If you look at this time and range and other things, and there is actually a higher time and range, even though this mean CGM glucose was a little bit higher in this study, but it's a completely fully closed loop, but unfortunately you have to wear other things to get some other inputs. But obviously, you know, glucose is not based on only what we eat and there are so many things, the hormones, stress level, there are more than 40, 50 different things are changing our glucose level in a second. So we are trying to implement this, so many things into this model to see that what we can do to decrease the burden of the people with especially with type one diabetes. So using the AI models for the full artificial pancreas, and then you know, there are different models you can use. And you have the short term glucose forecasting and then you have the, you know, multi output fully connected neural networks and you have the random forest. And I get this slide from the NIH artificial pancreas workshop and then it was last year. Then there are also discussion about using the AI in the AI systems in the type one diabetes care. So there are different models that when you include into the, you know, MPC algorithms that, you know, you can just get so many different results. Another way of thinking is that is kind of a new terminology is the digital twins. So what is a digital twins is the digital twin basically, the difference is the digital twin is, is a virtual representation of you and like your patient is kind of you are getting your exact one person and making a digital twin. And you are working all your models only for one person, you know, in the AIs and machine learning and using the mass data and using all of this to other people. In digital twins things about that is more like an individualizing things for the certain things. And right now there are some only small studies are focusing on the very individual things in type one diabetes management. For example, the exercise. Exercise is a very individual thing. So as probably you all know that your patients are doing exactly the same exercise, exactly the same thing. Some people are having more hypoglycemia than the other ones. So everybody's response is different. So in two different studies in the one is in the University of Padua from Italy, other one they did in Ohio State University. And then they, you know, compare that in the Ohio, they did this, you know, aerobic exercise and resistance exercise. And they created a digital twin and then they used in this in silico model. And they compare the time and range and comparison to the recommendations based on the consensus guidelines using the, you know, you take that amount of carb, you wait, decrease your insulin and take another carb and take this, take that. Versus the no intervention, this person did everything else. And you can see the digital twin learned in the system so much better and achieved a higher time and range in the systems. And in the University of Padua model also they compared in this, you know, CGM glucose versus in the, by the time and they see that when the digital twins can adding a rescue carb to, you know, prevent hypoglycemia. So maybe not in the, you know, all the models, but for the, you can integrate certain things into make things more individualized. So do we use AI systems in CGMs? You know, there is always a prediction system like you'll be getting a 20 minutes later and other things. But to you learning and using the more neural networks and the one of the companies are coming up, it's in the Europe first. And I think it's gonna take some time to come to US a little bit more. And this is the Roche CGM. And there is a two hour glucose prediction system, 30 minutes hypoglycemia prediction, and an overnight hypoglycemia prediction. So they did a study more than 300 patients and they just published with type one and type two diabetes. And there's a low prediction, Ariander curve ROC was 0.95 and nighttime prediction was 0.85. So as you can see here in this model that it shows that what can be your highest and lowest sensor glucose ran for the overnight and for the system. This is also using the neural networks and it's using more than the CGM trend and the CGM, you know, glucose and try to learn from the system that with some other input. So do we use the AI system to detect and monitor the complications? We already do. So there are FDA approved already and a couple FDA approved diabetic retinopathy screening systems that uses AI. So how it works is, is again, is from the deep learning. You have, for example, in this example, you have 500 pictures of the retina and then in a thousand cycle, you are just with the deep learning model. You are just showing to do, you know, the system. And for example, in this prediction in the first example, it says no diabetic retinopathy, but actually it's a moderate diabetic retinopathy. You are just sending back and to the, you know, to the system and saying, then update the weights. And then you are just going to the next one. This, you just do the same thing for the other 500. And then for example, in this 500 example was, is predicting mild DR and, but actually proliferative DR, then you are sending back. And then this way it's just a learning that what to do, you know, in doing that. So in retinopathy screening, for example, it just saves a lot of time for us in the primary care clinics, you know, and then all the other endocrinology clinics in five minutes, people are getting a picture and this is getting a diabetic retinopathic screening and the insurance is mostly covering on that. And it saves, and it can access to more, a lot of people. And if it's a gradable, you know, it gives the grade. If it's not gradable, it just suggests that there's a human should do that. So how they do that, they just, you know, train the models in different data sets. And in here, for example, in model one and model two, three, four, and using the different data sets. And then we are getting this oldest models and using the same data set to get it trained for the screening for the diabetic retinopathy. And you can see that there is a different are under curves for the same, you know, the data set. And because they were trained in a different model and then, you know, different data set, the different models. And when you use from the same data set and then you can see who has the highest, you know, AUC and then you can just, you know, get a sense on doing that. And we use the, also integrate the machine learning with the CGM data to predict the retinopathy. And for example, there's a very recent study and just published and there's a very prototype in the 30 patients using only the CGM metrics to detect the diabetic retinopathy. And this model was successful is with the 0.92 for the machine learning model. And using the different things in the time and range, I mean, tight range and CV and other things and getting a sense of that, which can be a little bit more predictive in this model. And we use the same thing for nephropathy detection with the AI. And then you see that in this one was 76% in predicting nephropathy external validation. And then, you know, the majority of them, the data were used to train the system. And then, you know, the similar amount was just for the testing. And you can see the duration of diabetes, you know, and then postprandial glucose, serum creatinine, you know, blood pressure and other things and the different weights. And in terms of the, what can be, you know, used in the detection of the diabetic nephropathy. So we can use so many different things to do that. Do we use AI in clinical decision making? We already use. So there are companies, you know, like for example, one of them is EndoDigital. And then, you know, for it's a great tool to use in the primary care or it's the busy endocrinology practices to making decisions for the insulin dosing. Then you just upload your blood glucose meter results, your CGM results, or your pump results. And then the system is just giving you a recommendation. Okay, increase the glargine dose from 22 to 23 or change the, you know, carb ratio from 10 to five or from, you know, your correction factor from 50 to 60. And there is a couple of studies showing that these are the, you know, very similar recommendations based on the endocrinologists are expert in the field. So to summarize, the AI-driven diabetes solutions are thriving from complication detection to advanced AI, the systems AI is used. Challenges exist to the use of AI in diabetes care. Cost effectiveness, equity, and acceptance are some of the challenges. So I just would like to thank you, the Barbara Davis Center and my lab members and for their help and their great work. Thank you very much.

artificial intelligence

diabetes care

continuous glucose monitors

automated insulin delivery

clinical decision-making

complication monitoring