Multivariate Glucose Forecasting Using Deep Multihead Attention Layers Inside Neural Basis Expansion Networks.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Diabetes Technology

Background:

• Continuous glucose monitoring (CGM) is vital for closed-loop diabetes systems but faces physiological delays.
• Deep learning models offer improved glucose forecasting but have limitations like non-interpretability and high computational costs.
• Accurate glucose forecasting is essential to prevent hyperglycemia and hypoglycemia by informing insulin dosage.

Purpose of the Study:

• To develop and validate a novel neural network architecture for improved glucose forecasting.
• To address challenges in accuracy, data requirements, and personalization in deep learning-based glucose prediction.
• To create a partially interpretable model with high forecasting accuracy for diabetes management.

Main Methods:

• Proposed a novel network architecture combining multi-head attention, dense, and theta layers within neural basis expansion network layers.
• Validated the proposed network on the OhioT1DM database.
• Evaluated forecasting performance for prediction horizons of 30 and 60 minutes.

Main Results:

• Achieved an average root mean squared error (RMSE) of 16.57 ± 2.56 mg/dL and mean absolute relative difference (MARD) of 6.81 ± 1.39% for a 30-minute prediction horizon.
• Obtained improved results for a 60-minute prediction horizon with a mean RMSE of 29.25 ± 6.02 mg/dL and MARD of 12.15 ± 3.15%.
• Demonstrated superior performance compared to previous works in glucose forecasting accuracy.

Conclusions:

• The novel network architecture offers high glucose forecasting accuracy and potentially enhanced interpretability.
• The proposed model addresses key limitations of existing deep learning methods in diabetes management.
• This approach shows promise for more personalized and effective closed-loop diabetes control systems.