Adding AI Technology to Artificial Pancreas Improves Device Efficiency Compared With Automated Insulin Delivery Algorithm
In this video, Boris Kovatchev, PhD, describes his team's first-of-its-kind study, which found that adding artificial intelligence (AI) technology to an artificial pancreas improves the device's efficiency and has a comparable safety profile versus an automated insulin delivery algorithm. Dr Kovatchev details what promoted his research, whether it fills a gap in our knowledge, how and why AI technology is useful for these devices, and more.
Additional resource: Kovatchev B, Castillo A, Pryor E, et al. Neural-net artificial pancreas: a randomized crossover trial of a first-in-class automated insulin delivery algorithm. Diabetes Technol Ther. 2024;26(6):375-382. doi:10.1089/dia.2023.0469
TRANSCRIPTION
Boris Kovatchev, PhD: I'm Boris Kovatchev at the University of Virginia, affiliated with the School of Medicine and the School of Data Science, and I'm also Director of the UVA Center for Diabetes Technology.
Consultant360: What prompted this research?
Dr Kovatchev: So the artificial pancreas field has been in development for approximately 15 years now. The first systems out there serve people with type 1 diabetes and the control algorithms that have been used in this system, in these systems, have been usually based on models of the human metabolic system that are sometimes difficult to track in real time. And we decided that a good medium for a control algorithm, artificial pancreas control algorithm to reside would be artificial intelligence and new approaches such as neural networks and machine learning. So that idea of switching the base of a control algorithm from the traditions to new technologies that are becoming increasingly available and more and more powerful daily prompted this study.
Consultant360: How does this study fill a gap in our knowledge?
Dr Kovatchev: I wouldn't say it fills a gap because there was really no gap. It just introduces a new paradigm. This is first-in-class system, an artificial pancreas that uses neural net instead of traditional calculations. And the main field that it opens is the process of learning. And I will explain a bit about that. Traditional algorithms have difficulties learning about the person and learning over time.
What are the specifics of these individuals so they can practice precision medicine and precision insulin delivery? On the other hand, AI and machine learning, they have learning in the name, basically, the major capabilities of machine learning are to learn over time as they go. So once you initialize it, you let it go to a person and it learns and adapts. So that's what this opens.
It doesn't do it yet. In the first study. We did not practice learning. We had to just try to see whether a neural net can handle the task, but the future is open towards learning systems that adapt to every person over time and use modern AI technology to do so.
Consultant360: What were the results of your study and were there any surprises in those results?
Dr Kovatchev: So the results, let me explain first what the study was. We had a top-of-the-line traditional algorithm. And we designed a neural net that learned from the data of this algorithm, how that algorithm behaves.
And then we compare the two. We compare the traditional algorithm, which was again top of the line algorithm, with what the neural net has learned. And the basic premise, it was randomized crossover study.
The basic premise was that they will behave the same way, that the neural net will learn exactly what that algorithm was doing, and it did, and the results were identical between the two algorithms which was the premise of the study that they will be identical. The surprise was that the neural net did it six times faster in terms of computing time. So, it was much faster than the traditional heavy algorithm. And that was a new element that we found out very useful for future use in smaller devices.
Consultant360: Why is this new technology useful for smaller devices?
Dr Kovatchev: It's in the math. That the algorithm has an approximation of parameters and numerical calculations and so forth, while the neural net is basically linear algebra and multiplication of matrices, so it's expected to be faster, but we didn't know for sure.
Consultant360: Looking ahead, what’s next for this type of research?
Next, is indeed learning and to run future studies that will use machine learning methods to have this base algorithm, base neural net algorithm, learn from the data of a particular person as it goes and adapt to it and anticipate changes.
So it has applications, what is known out there is, they call it the Holy Grail of artificial pancreas, but basically it's a fully automated control that doesn't require mill announcements or physical activity on anything like that. So we are running a study now. My colleague, Mark Breton, is running a study in Sil Brown, a fully automated version of this system that tries to anticipate these disturbances, the meals, and exercise, and react automatically, as opposed to be taught to do so.
Consultant360: What is kind of like the main takeaway of this study?
Dr Kovatchev: I guess the main takeaway is that for the first time around AI controls the delivery of a drug insulin, which is potentially dangerous in people with diabetes and it does the job well and It will be getting better.