Learning of Input Penalty Function for Zone Model Predictive Blood Glucose Control

Abstract

Maintaining appropriate blood glucose levels within an acceptable range is essential for patients with type 1 diabetes. Zone model predictive control seeks to maintain its output within a given range and is often employed to keep blood glucose within the acceptable range. However, the use of a linear time-invariant model for blood glucose control results in poor accuracy for long-period predictions. Hence, previous studies implemented an input penalty function in order to correct errors in the calculated insulin dose caused by inaccurate model prediction. Parameters of the input penalty function were optimized to improve control performance for specific patient populations under one-meal scenarios, and it might not keep performance for each patient or other scenarios. In contrast, we construct a learning method of an input penalty function under the control of each patient. The learned input penalty function improves the control performance of each patient on various meals.