Predicting glucose level with an adapted branch predictor.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Diabetes Technology

Background:

• Diabetes mellitus is characterized by high blood glucose levels, damaging organs over time.
• Closed-loop artificial pancreas systems use continuous glucose monitors and insulin pumps for glucose control.
• Accurate short-term glucose prediction is crucial for efficient artificial pancreas function, but complex methods may have drawbacks.

Purpose of the Study:

• To develop a novel, low-complexity, and explainable blood glucose prediction method.
• To compare the performance of this new method against a state-of-the-art deep learning approach.

Main Methods:

• The proposed method is derived from the Intel P6 branch predictor algorithm.
• Meta-Differential Evolution was used to optimize predictor parameters on benchmark datasets.
• The new algorithm was compared to a deep learning method using the Blood Glucose Level Prediction Challenge dataset.

Main Results:

• The new algorithm achieved comparable predictive accuracy to deep learning.
• It extended the prediction horizon by 5 minutes with 95.8% of predictions within a 30% relative error.
• The method operates solely on the glucose-level signal, demonstrating significantly less computational complexity.

Conclusions:

• The novel algorithm provides comparable predictive accuracy to deep learning for blood glucose levels.
• Its low complexity and interpretability offer advantages for artificial pancreas systems.
• This approach enhances safety and feasibility for low-power devices and new patients.