Design of Insulin Pump Control by Reinforcement Learning

Abstract

  Applying the minimal amount of insulin to effectively control plasma glucose concentrations is important to decrease potential insulin resistance and allow for the effective treatment of diabetes patient for entire lifespan. We are interested in the design of insulin pump infusion controller that is represented by a system of ordinary differential equations reflecting the interaction between glucose and insulin within the human body. The main objective of insulin infusion profile control over time is to maintain plasma glucose concentrations within a healthy range. We use Reinforcement Learning (RL) and investigate the structure of the weighted reward function, to evaluate the efficacy of RL to find an optimal solution that meets predetermined requirements. The RL insulin infusion profile will be compared to those produced from a Proportional-Integral-Derivative (PID) controller or by best practice by analyzing the area under curve for each profile to determine if RL is appropriate for this situation.