Evaluating Deep Q-Learning Techniques for Controlling Type 1 Diabetes

Abstract

Patients with type 1 diabetes (T1D) must continually decide how much insulin to inject before each meal to maintain an acceptable level of blood glucose. Recent research has worked on a solution for this burden: the artificial pancreas (AP), which is a closed-loop system combining a continuous glucose monitor (CGM) and an insulin pump with a decision-making algorithm.

The goal of this thesis is to implement and evaluate several hybrid closed-loop deep Q-learning (DQL) algorithms for the task of regulating blood glucose in T1D patients. Firstly, we will review the diabetes disease, its burdens and challenges, and existing treatment models. Secondly, we will study the foundations of reinforcement learning (RL) and deep reinforcement learning (DRL), with the emphasis on DQL techniques. Then we will merge the theories and implement DQL algorithms with the application of regulating blood glucose for T1D in-silico patients. Finally, we will test these algorithms on a T1D glucoregulatory simulator.