Improving Long-Term Glucose Prediction Accuracy with Uncertainty-Estimated ProbSparse-Transformer

Abstract

Accurate prediction of blood glucose (BG) with precise data recorded by continuous glucose monitoring (CGM) is essential to improve the safety of closed-loop insulin delivery systems for diabetic patients. However, predicting BG trends under long-term prediction horizons is challenging due to the dynamic complexity of glucose changes. In this work, a ProbSparse-Transformer model, which alleviates the long-term error spreading effect seen in traditional autoregressive models, is developed. This model incorporates a trustworthy uncertainty-estimation approach to reduce output variance, further improving predictive accuracy. Additionally, an open-source benchmark is established using four public datasets and five evaluation metrics to comprehensively assess model performance. This model shows significant improvements in both short-term (15–30 min) and long-term (45–60 min) BG predictions. In the 60 min task, it achieves root mean square error values of 10.86, 15.33, 20.46, and 13.74 mg dL⁻¹ across four datasets, representing a 20%–39.4% improvement over previous methods. Finally, the model on edge devices is compressed and deployed, demonstrating its potential for practical application in real CGM systems.