Systems biology informed neural networks for glucose prediction and control in diabetes

Type 1 diabetes (T1D) affects nearly 2 million people in the United States, including more than 300,000 children and teenagers. In T1D, the body cannot produce enough insulin, the hormone that regulates blood sugar. Without proper control, blood sugar can rise or fall rapidly, leading to dangerous complications such as heart problems, confusion, coma, and even death.
Modern wearable devices—like continuous glucose monitors and insulin pumps—have brought us closer to building an automated “artificial pancreas” that adjusts insulin delivery in real time. However, current systems still face major limitations. The data they rely on can be noisy or incomplete, and many existing algorithms only use one type of information, such as glucose readings alone. This prevents them from fully capturing the complex and constantly changing biology that drives blood sugar levels.
Our project aims to address these challenges by developing systems biology–informed neural networks—advanced computer models that imitate how each person’s glucose and insulin levels change throughout the day. These models combine two key strengths: (1) biological knowledge based on well-established physiological equations, and (2) modern machine learning that can learn patterns from large, real-world datasets. They also consider real-life factors such as meals, exercise, stress, and insulin dosing.
We have three main goals. First, we will create machine learning models that use multiple types of data to improve glucose prediction accuracy. Second, we will build mathematically consistent models that blend deep learning with biological equations to better represent glucose-insulin dynamics. Third, we will design safer and more reliable insulin-control strategies using advanced reinforcement learning methods that learn effectively from past patient data.
If successful, this research will enable more accurate glucose predictions, individualized insulin recommendations, and safer automated glucose control. These advances will bring us closer to creating effective, personalized tools that improve the daily lives and long-term health of people living with type 1 diabetes.