Type 1 Diabetes Blood Glucose Management Based on Fuzzy Logic and Reinforcement Learning

Developing a personalized insulin regimen for the treatment of Type 1 Diabetes (T1D) is a healthcare task with very high safety requirements. Compared to traditional control algorithms, current reinforcement learning-based methods for blood glucose control can increase the duration of blood glucose within the target range while reducing risks. However, they do not account for the uncertainties of factors affecting blood glucose levels, such as diet, age, weight, etc. These factors do not have a simple linear relationship with blood glucose changes. For example, consuming a certain amount of carbohydrates does not always immediately result in the same degree of blood glucose increase, and this variability is precisely the key to personalized insulin regimens. To address this issue, this work proposes a personalized insulin regimen based on fuzzy logic reasoning and reinforcement learning, namely the Fuzzy Rule-Based TD3-BC model (FRBTD3-BC). This regimen learns from real-world treatment data of T1D patients, uses membership functions to obtain the initial conditions and outcome categories of the data, and then calculates the weight of each rule based on confidence levels. Subsequently, a fuzzy rule base is obtained and embedded into the Actor module of the TD3-BC algorithm in the form of a fuzzy reasoning logic layer. By interacting with the FDA-approved UVA/Padova glucose kinetics simulator, the utility of FRBTD3BC in blood glucose control was studied in 30 virtual patients. The results show that the FRBTD3-BC model outperforms other comparable methods, with the TIR indicator increasing by 5.53 ± 0.64% compared to the original TD3-BC.