Reinforcement Learning Models and Algorithms for Diabetes Management

Abstract

With the advancements in reinforcement learning (RL), new variants of this artificial intelligence approach have been introduced in the literature. This has led to increased interest in using RL to address complex issues in diabetes management. Using RL, a decision maker (or agent) observes decision-making factors (or state) from the dynamic operating environment, selects actions based on its optimal policy (i.e., a mapping between states and optimal actions), and subsequently receives delayed rewards. The agent adapts its policy to changes in the operating environment to maximize its cumulative reward as time goes by, thereby improving system performance. This paper presents how various variants of RL have been used to improve diabetes management, such as a higher time in range during which the blood glucose level is within the normal range and a higher similarity between RL and physician’s policies. Key highlights focus on the application of RL in diabetes management, including a taxonomy of the attributes of RL (e.g., roles and advantages), essential elements for training (e.g., data and simulators), representations of diabetes attributes in RL models, and variants of RL algorithms. In addition, this paper discusses open issues and potential future developments in the use of RL in diabetes management.