Continuous performance monitoring of a closed-loop insulin delivery system using Bayesian Surprise

Abstract

The development of an artificial pancreas for the treatment of insulindependent diabetes is a big challenge for control theory. Many closed-loop algorithms have been widely evaluated for their ability to recreate, as closely as possible, glucose and insulin profiles observed in healthy individuals. Nevertheless, an artificial pancreas system also involves a critical necessity for supervision of the control loop functioning and the early detection of anomalies as well as performance deterioration. This work presents a performance monitoring approach using Bayesian surprise to fast detect functional degradation and guarantee adequate control of blood glucose levels. Bayesian surprise is significantly affected by any deviation from desired operation in a controlled system, which allows its use for continuous glucose monitoring.