GLIMMER: Advanced Blood Glucose Prediction Model for Type 1 Diabetes

Type 1 diabetes (T1D) is an autoimmune disease where the body attacks and destroys insulin-producing beta cells in the pancreas and affects nearly 10 million people globally. Because glucose needs to be continuously monitored for precise insulin dosing, managing T1D is challenging. While automated insulin delivery (AID) systems and continuous glucose monitors (CGMs) attempt to prevent dysglycemia, there are still limitations in their prediction algorithms.

 

Researchers at Arizona State University have developed a novel light-weight machine learning model, GLIMMER, designed to improve blood glucose level predictions, enhancing safety and management for Type 1 Diabetes patients. GLIMMER leverages a advanced machine learning algorithms with a custom loss function optimized by a genetic algorithm to accurately predict blood glucose levels, especially in critical dysglycemic regions. By utilizing continuous glucose monitoring (CGM) data, insulin, and carbohydrate intake, GLIMMER outperforms existing models in predicting hypoglycemic and hyperglycemic events, supporting improved patient safety and diabetes management. Even though GLIMMER has only 10,000 parameters, its performance is comparable with existing glucose forecasting models that use 18 million parameters.

 

 

This approach uniquely combines tailored feature engineering, extensive hyperparameter tuning and a novel custom loss function which prioritizes accuracy in dysglycemic regions to support earlier corrective actions in real-time.