Data-Driven Solutions for Blood Glucose Management

Abstract

Type 1 diabetes (T1D) affects millions of people worldwide. People with T1D must regularly monitor their blood glucose (BG) level and administer insulin to manage it. This process is burdensome, necessitating frequent measurements, meal size estimation, and bolus calculations. Automated management solutions have been proposed, but still require patients to accurately estimate meal sizes and manually update user parameters. We aim to discern and address challenges in the utilization of data-driven approaches for BG management. First, estimating meal size is challenging, resulting in noisy carbohydrate counts, which hamper BG management. We address noise in patient-reported carbohydrate estimates by developing a novel training method for denoising autoencoders. Our approach leverages the relationship between carbohydrates and the BG signal. Second, while data-driven approaches could obviate the need for manual patient updates via their capacity for online learning, current approaches fall short of the level of accuracy required for safe automated BG management. More specifically, learning the impacts of carbohydrates and insulin boluses on future BG values is challenging due to the relative sparsity of these variables and the correlation between them. We propose a forecasting approach that accounts for the relative sparsity of bolus and carbohydrate values by isolating and constraining their effects to align with domain knowledge. In addition, we address bolus and carbohydrate entanglement with an approach that leverages correction bolus values to disentangle individual variable effects. Combined, these contributions represent a clear step towards data-driven BG management, offering the potential to the reduce patient burden associated with T1D.