Prior-knowledge-embedded model predictive control for blood glucose regulation: Towards efficient and safe artificial pancreas

Abstract

Objective:The artificial pancreas (AP) system based on model predictive control (MPC) has the potential to provide effective and reliable regulation of blood glucose concentration (BGC) for people with type 1 diabetes by utilizing glucose prediction models and appropriate safety constraints. Methods:In this work, a MPC strategy with novel model and safety constraints is proposed for BGC regularization. A prior-knowledge-embedded glucose prediction model based on kernel-regularized latent variables (LV) regression method is developed where the prediction power of the model is improved by integrating the prior knowledge of glycemic dynamics and balancing model complexity and flexibility by tuning the hyper-parameters of the kernel. Based on the prior-knowledge-embedded model, MPC strategy is formulated and BGC constraints that are not dependent on the announcement of meals and physical activity are proposed and incorporated into the MPC to reduce the risk of hypoglycemia. The benefits of the proposed model and MPC approach were evaluated through in-silico studies. Results:The proposed model achieves comparable or improved prediction accuracy with a root mean squared error of 14.08 mg/dL, 18.47 mg/dL, and 21.23 mg/dL for 30-min, 60-min, and 120-min-ahead prediction, respectively. And simulation results showed significant improvement in the time in the safe range (70–180 mg/dL) from 58.04% to 71.27% and from 78.81% to 82.31% without causing hypoglycemia. Conclusions:The proposed MPC strategy can provide safe and effective regulation of BGC without requiring manual announcements of meals and physical activity. Significance:This work is a step forward towards a fully-automated AP system.