Glycemic-aware and architecture-agnostic training framework for blood glucose forecasting in type 1 diabetes

Managing Type 1 Diabetes (T1D) demands constant vigilance as individuals strive to regulate their blood glucose levels to avert the dangers of dysglycemia (i.e., hyperglycemia and hypoglycemia). Despite the advent of sophisticated technologies such as automated insulin delivery (AID) systems, achieving optimal glycemic control remains a formidable task. AID systems integrate data from wearable devices including continuous subcutaneous insulin infusion (CSII) pumps and continuous glucose monitors (CGMs), offering promise in reducing variability and improving time-in-range. However, these systems often fail to prevent dysglycemia, partly due to limitations in prediction algorithms that lack the precision to anticipate impending glycemic excursions. This gap highlights the need for more advanced blood glucose forecasting methods. We address this need with GLIMMER ( G lucose L evel I ndicator M odel with M odified E rror R ate), a modular and architecture-agnostic training framework for glucose forecasting. GLIMMER combines structured preprocessing, region-aware loss formulation, and weight optimization using a genetic algorithm to emphasize prediction accuracy in dysglycemic regions. We evaluate GLIMMER on two datasets: the publicly available OhioT1DM dataset and a new dataset (AZT1D) constructed by collecting data from 25 individuals with T1D. Our extensive analyses show that GLIMMER consistently improves glucose forecasting performance over baseline architectures, enhancing RMSE (Root-Mean-Square Error) and MAE (Mean-Absolute-Error) by up to 24.6% and 29.6%, respectively. Additionally, GLIMMER achieves a recall of 98.4% and an F1-score of 86.8% in predicting dysglycemic events, demonstrating its effectiveness in high-risk regions. Compared to state-of-the-art models with millions of parameters—such as TimesNet (18·7 M), BG-BERT (2·1 M), and Gluformer (11·2 M)—GLIMMER achieves comparable accuracy while using only 10K parameters, demonstrating its efficiency as a lightweight, architecture-agnostic solution for glycemic forecasting.