Optimal Blood Glucose Prediction based on Intermittent Data from Wearable Glucose Monitoring Sensors

Blood glucose prediction is to predict the glucose trend over time based on historical glucose data, and it plays a crucial role in the closed-loop control of artificial pancreas, which can reduce the risk of complications by regulating insulin dose and injection time. This paper proposes a Kalman-filter-based glucose prediction method through minimizing the mean square prediction error, which assumes that the data is sampled every 15 min from a wearable flash glucose monitoring sensor. This method calculates glucose estimates every 5 min and provides glucose predictions for the next 30 min. The method is evaluated on in-silico data generated from the 10-adult cohort of the US FDA-accepted UVA/Padova T1DM simulator. The predicted results are compared with CGM data with 5-min sample-period through multiple metrics, including the mean square prediction error and the mean absolute relative deviation. The results show that the performance of the proposed approach with slow-rate glucose data (15 min) is close to that obtained based on fast-rate data (5 min).