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

Glucose forecasting is a crucial feature in a closed-loop diabetes management system relying on minimally invasive continuous glucose monitoring (CGM) sensors. Forecasting is required to prevent hyperglycaemia or hypoglycaemia due to delayed or incorrect dosage of insulin while using CGM sensors, which suffer from physiological delay due to diffusion kinetics between blood and interstitial fluid. Recent works have demonstrated better performance with deep learning methods in this application. However, deep learning methods suffer from model non-interpretability, time lag, prediction accuracy, large training data requirements, and high-end computational resource requirements. In this paper, we seek to address challenges of accuracy, data requirements, and personalisation problems by proposing and validating a novel network architecture where multihead attention layers followed by a combination of fully connected dense and theta layers embedded deep inside neural basis expansion network layers. This network leads towards possible partially interpretable models with high forecasting accuracy. In comparison to previous works, the proposed network leads to an average root mean squared error (RMSE) of 16.57 $\pm$ 2.56 mg/dl and mean absolute relative difference (MARD) of 6.81 $\pm$ 1.39%, which is better in comparison to previous work, when the network was validated on OhioT1DM database for a prediction horizon (PH) of 30 mins. Better results compared to previous work were also obtained for a PH = 60 mins where mean RMSE was 29.25 $\pm$ 6.02 mg/dl and MARD was 12.15 $\pm$ 3.15%.