AI-powered diabetes technology still has a long way to go.under both America and England Medical device regulations, commercial automated insulin dosing systems without AI fall into the highest risk class. AI-driven systems are in the early stages of development, so the debate about how they should be regulated is just beginning.
Emerson’s experiment was entirely virtual. Testing AI-assisted insulin administration to humans raised a number of safety concerns. In a life-or-death situation like administering insulin, giving control to a machine can be dangerous. “The nature of learning makes it absolutely possible to take a step in the wrong direction,” says Mark Breton, a professor at the University of Virginia Center for Diabetes Technology who was not involved in the project. “A slight deviation from the previous rule can make a big difference in the output. That’s the beauty, but it’s also the danger.”
Emerson focused on Reinforcement Learning (RL), a machine learning technique based on trial and error. In this case, the algorithm “rewarded” good behavior (meeting blood sugar targets) and “punished” bad behavior (blood sugar too high or too low). Since the team was unable to test on real patients, they used offline reinforcement learning, which utilizes previously collected data, rather than learning in situ.
Thirty virtual patients (10 children, 10 adolescents, and 10 adults) were UVA/Padova Type 1 Diabetes Simulator, a Food and Drug Administration-approved animal preclinical alternative. After offline training on seven months’ worth of data, we let the RL take over insulin dosing for a virtual patient.
They mimic device failures (missing data, inaccurate readings) and human errors (miscalculating carbs, irregular meal times) to see how they deal with real mistakes. We conducted a series of tests designed to It’s a test that most researchers without diabetes would never think of. Run. “Most systems only consider two or three of the factors: current blood sugar, previously administered insulin, and carbohydrates,” says Emerson.
Offline RL handled all these difficult edge cases well in the simulator, outperforming current state-of-the-art controllers. The biggest improvement came in situations where some data was missing or inaccurate, simulating situations like someone moving too far from the monitor or accidentally squashing her CGM.
In addition to reducing training time by 90% compared to other RL algorithms, the system kept virtual patients within their target blood glucose range for 1 hour longer per day than commercial controllers. Next, Emerson plans to test offline RL against previously collected data. real Patience. “Most people with diabetes [in the US and UK] We have them record data continuously,” he says. “We have a great opportunity to capitalize on that.”
However, translating academic research into commercial devices requires overcoming significant regulatory and corporate barriers. Bretton says his study results show promise, but that they were obtained from hypothetical patients and their relatively small group. “That simulator, however great it is, represents only a fraction of our understanding of human metabolism,” he says. The gap between simulation research and real-world applications is “not unfillable, but large and necessary,” Bretton continues.
Medical device development pipelines can feel terribly bogged down, especially for people with diabetes. Safety testing is a time-consuming process, and even after new devices hit the market, the lack of code transparency, data access, and cross-manufacturer interoperability leaves users with little flexibility. . There are only five pairs of compatible CGM pumps on the US market, and their high prices limit access and usability for many. “In an ideal world, systems would be plentiful and people would be able to choose the pumps, CGMs, and algorithms that are right for them,” said Dana Lewis, founder of the open-source Artificial Pancreas Systems movement. (OpenAPS). “You can go through life without thinking about diabetes.”
