Reinforcement learning for long term adaptation of an artificial pancreas

CECILIA SERAFINI; NICOLÁS ROSALES; FABRICIO GARELLI

Barcelona

Conferencia; 15th International Conference on Advanced Technology & Treatments for Diabetes.; 2022

Background and AimsPrevious clinical trials and in-silico work show that insulin sensitivity variation in diabetic patients can compromise the performance of closed loop glycemic control. Generating adaptive controllers is key to overcome this issue.MethodsIn this work, a a Q-learning based long-term adaptation technique for the previously introduced Automatic Regulation of Glucose (ARG) algorithm [1] is presented. The presented configuration modifies only one parameter in the controller (the insulin-on- board -IOB- limit) instead of replacing it entirely, avoiding the "black box" issue associated with Reinforcement Learning.The resulting policy is evaluated in-silico using the UVA simulator and compared against a manual rule-based strategy.ResultsResult shows that the RL agent manages to limit insulin infusion when the patients sensitivity increases, thus avoiding hypoglycemia without significantly increasing time in hyperglycemia, while the manual scheme does not achieve similar results. It is also worth noting that insulin infusion is reduced showing that the RL strategy improves insulin infusion profile, as well as overall glycemic excursion.ConclusionsReinforcement learning for long term adaptation shows great promise as it reduces insulin infusion and avoids hypoglycemia when patient's sensitivity increases.