Illinois Tech Project Using Machine Learning to Fully Automate Insulin Delivery Receives NIH Funding

CHICAGO—December 7, 2022—A project led by Ali Cinar, an Illinois professor of technology in chemical engineering, that aims to help ease the mental burden of people with type 1 diabetes, has received $1.2 million from the National Institutes of Health over the next four years to develop a machine learning system that can be integrated with its artificial pancreas system to improve artificial pancreas accuracy.

The typical person with type 1 diabetes has to make between 100 and 200 decisions every day just to keep their blood sugar levels stable.

“Part of the function of their pancreas is entrusted to their brain,” says Cinar, who also holds the Hyosung SR Cho Endowed Chair in Engineering.

If they misjudge or forget to deliver the correct insulin dose, they may experience weakness, dizziness, fainting, or more severe symptoms when their glucose levels get too low. People whose glucose levels are often outside the target range or are too high may experience long-term complications ranging from cardiovascular disease and kidney failure to retinopathy.

Cinar has been on the cutting edge of this technology for many years. His research group was the first to integrate data on physical activity received by the sensors of wearable systems such as a sports bracelet into the control system of the artificial insulin-delivering pancreas.

This project goes beyond that, analyzing a person’s past behavior in greater depth using machine learning and customizing the device’s decision-making algorithm to improve its ability to determine if someone engages or will soon engage in behaviors that may impact glucose levels.

The predictive ability is important because there is a delay between when insulin is given and when it starts working.

“If someone is eating lunch every day at noon and the meal is typically 20-30 grams of carbs, then if their current blood sugar level is not very low, at 11:45 a.m. we could say, ‘Everything indicates that the This day is a typical weekday for that person, so let’s give them not the full dose of insulin, but a small amount to lessen the effect of the meal on glucose,” Cinar explains.

Machine learning and artificial intelligence algorithms developed in collaboration with Associate Professor of Computer Science Mustafa Bilgic will match the current day’s pattern to the behavior patterns of the specific individual.

The system would assign a probability to the likelihood of the person having breakfast soon based on the person’s behavior on the current day and deliver an insulin dose accordingly. Then he would continue to monitor the glucose level and if, as expected, it started to rise because the person was eating, additional insulin would be given.

Current automated insulin delivery systems on the market require the user to calculate the carbohydrates in their meals and manually report them to the system. They also expect the user to make manual adjustments while exercising. It takes time and effort, and it leaves this critical medical function open to human error.

Certain groups such as children or forgetful people are at a disproportionate risk of not entering their calorie information or entering it incorrectly.

Current monitoring systems also lack great complexity that can impact glucose levels. Beyond food and exercise, stress, sleep, and other factors can raise or lower glucose levels.

A person who manages their insulin manually may consider these factors when deciding the insulin dose, but Cinar aims to design the artificial pancreas to detect and incorporate the presence of these factors into automated decision-making.

If a person is stressed, a system that infers physical activity based on heart rate information may assume that they are exercising because their heart rate is high. But stress and exercise impact glucose levels in opposite directions, so the system can lower insulin and make things worse, raising their blood sugar even further.

Moreover, several factors can occur simultaneously. During a run, a distance runner’s glucose level could be affected by the combined effect of exercise, stress, and any foods they eat while running.

“That’s why we really changed our focus from just detecting the exercise to detecting the state of the person,” says Cinar. “And it’s getting more and more interesting and challenging.”

With enough historical data, Cinar says the machine learning system could even learn to predict the behavior of someone with seemingly irregular habits.

“The benefit of powerful machine learning tools is being able to unravel the secondary relationships that exist. No matter how erratic people claim their behavior is, there are always certain patterns that can be picked up,” says Cinar. “It could be five patterns for someone who is very routine and 15 patterns for someone less routine. The system can look at how the day has gone, then look at the pattern dictionary to say, “Oh, that’s similar to pattern number 17, so let’s assume the rest of the day will unfold accordingly.”

In short, Cinar wants to give people with diabetes the chance to live their lives without constantly evaluating whether to log what they’re doing in their insulin delivery system.

“Someone may be running to catch the bus because the bus arrives too early or they left their house late. It’s not something they want to stop halfway to adjust the dosage of l insulin. That’s why we would like to create a fully automated system,” explains Cinar.

Cinar collaborators at Illinois Tech include Bilgic and Assistant Research Professor Mudassir Rashid.

Disclaimer: “Research reported in this publication was supported by the National Institutes of Health under award number 1R01DK135116-01. This content is the sole responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Ali Cinar, “Integration of AI and Systems Engineering for Glucose Regulation in Diabetes,” National Institutes of Health; Award number 1R01DK135116-01