Personalized Prediction and Control of Blood Glucose Levels using Dynamic System Modeling

Abstract

Patients with diabetes in the ICU typically rely on two methods to monitor their blood glucose levels: blood tests analyzed in the laboratory and fingerstick tests. While fingerstick tests offer a cost-effective and convenient approach, they sacrifice accuracy, providing only approximate measurements. In this paper, we propose a dynamic system model combining deep learning models and Kalman filtering to accurately predict blood glucose trajectory and assist with personalized glucose control over time. The resulting model, called a deep Kalman filter (Deep-KF) model, can be applied in the ICU and other healthcare settings. We validate our model using randomized real-world glucose and insulin clinical data. We demonstrate how our model can deliver precise blood glucose forecasts and insulin delivery recommendations and be used as a decision-support tool for healthcare providers, hence making glucose monitoring more efficient and patient-centric.