Artificial Pancreas Control for Diabetes using TD3 Deep Reinforcement Learning

Diabetes Mellitus is a chronic condition that affects approximately 6.5% of the population in Ireland. As well as being a burden on those who suffer from it, it is a huge burden to the state and accounts for approximately 10% of total global health spend. Diabetes cannot be managed from a clinical setting so there is a requirement for self-management with a constant need to understand what current blood glucose values are and responding by treatment with an appropriate dose of insulin. Fortunately, diabetes technology has improved dramatically in the last number of years with the invention of the continuous glucose monitor (CGM) that can report a blood glucose reading as frequently as every five minutes and insulin pumps that infuse insulin in frequent small doses mimicking endogenous insulin. Currently humans are still required to manage these devices, but it is every patient's (and clinicians) wish to close the loop and automate control. This study looks at control algorithms and asks if deep reinforcement learning (DRL) can be used as a potential solution for devising patient specific policies for control. A Twin Delayed Deep Deterministic Policy Gradient (TD3) model is implemented in a simulated environment and tested on three in-silico patients. The results show promise in controlling blood glucose profiles for the patients but in a limited setting. It concludes that while DRL is capable of learning to control blood glucose further research is required before it could be considered for human use.