Modeling Efforts for Improved Meal Prediction with Application to Blood Glucose Control

摘要

Type 1 Diabetes Mellitus is a disease characterized by the loss of insulin production from beta cells in the pancreas, which results in unregulated blood glucose (BG). The condition is permanent, exacting a high toll on an individual in terms of both health outcomes and treatment burden. Acute risk of low BG includes coma, seizure, and death, while longer term risk of high BG includes damage to the circulatory and nervous systems. To mitigate the risk of both ends, individuals must constantly stay vigilant, regulating BG levels with insulin injections and constantly monitoring BG levels.The artificial pancreas aims to reduce both health and treatment burdens by automatic regulation of BG levels. It consists of an insulin pump that injects insulin, a continuous glucose monitor that provides BG measurements, and a control algorithm that calculates dosing decisions. The next step in the development of artificial pancreas systems is fully closed loop control around meals. Unannounced meals present a challenge for the control algorithm because of the uncertainty surrounding the presence and content of the meal in addition to the slow and irreversible action of insulin paired with the acute risk of low BG.We propose a meal model, developed on gold standard triple tracer data, that considers meal size and shape and explicitly estimates uncertainty within meals. To quantify the quality of prediction in the context of BG control, we propose a new metric that considers asymmetry of prediction error and assesses prediction distributions in addition to single point predictions. Using the proposed metric, we tune an extended version of the model on a large data-set of free-living patient data and compare to previous work.The proposed model is first compared against three simpler models using triple tracer meal data. Prediction root mean square error is improved by 11% relative to the next best non-linear model. A simple implementation of control for all models suggests improved control capability for the proposed model: the proposed model is the fastest to compensate for a meal in 4 out of 6 cases and overcompensates the least (8% excess) in the worst case compared to other models (25% excess). The model is also validated on a large data-set by evaluating prediction capability and control performance. The proposed model improves predictions by 37% relative to previous work in terms of the proposed metric. In a retrospective simulation of control, the proposed model reduces clinical risk by 12% over previous work.An open source artificial pancreas system currently in use by many is Loop. Part of the presented work is an effort to introduce Loop into the control literature. We formulate the Loop control algorithm as a "coincidence point" model predictive control strategy paired with a linear state space model. We evaluate Loop in silico using error-prone scenarios and suggest improvements that can be made to the meal announcement functionality.