Kernel-Based Learning With Adaptive Physiological Constraints for Personalized Postprandial Glucose Prediction

Abstract

Accurate and physiologically interpretable postprandial glucose prediction is of importance in diabetes self-management. In this work, the problem of personalized glucose prediction is considered, and an interpretable postprandial glucose trajectory prediction framework is proposed based on kernel-based system identification under physiological constraints. Considering treatment requirements in inpatient scenarios, an online prediction update mechanism is developed to deal with intrasubject variability. Through incorporating physiological constraints abstracted from linearized compartmental models, a posterior performance assessment and adaptation mechanism is designed to guarantee the interpretability of the predicted glucose responses. The proposed method is evaluated through clinical data from type 1 diabetes mellitus (T1DM) subjects, and the results indicate that the proposed method can achieve physiologically interpretable postprandial glucose trajectory prediction with satisfactory performance.