Taiyu Zhu, United Kingdom
Imperial College London Electrical and electronic engineeringPresenter of 1 Presentation
PERSONALIZED MEAL INSULIN BOLUS FOR TYPE 1 DIABETES USING DEEP REINFORCEMENT LEARNING
Abstract
Background and Aims
Due to the high inter- and intra-population variability, it is often difficult for basal-bolus insulin therapy to effectively prevent postprandial hyperglycaemia and hypoglycaemia. Deep reinforcement learning (DRL) has recently achieved great success in different areas and has the potential to meet this challenge.
Methods
We propose a new DRL algorithm using an actor-critic architecture and fully connected neural networks, following the approach of deep deterministic policy gradient. We employ the UVa/Padova Type 1 Diabetes (T1D) Simulator as a testing platform and let the agent interact with the environment at each incidence of a meal intake. The agent’s states consist of real-time samples from continuous glucose monitoring, carbohydrate estimation, and meal ingestion time. Its action is to estimate the dose of premeal boluses. The percentage time in range (TIR) (70–180mg/dl) is set as the reward. We first trained the agent in a long-duration average scenario through long-term self-exploring to obtain a generalized model. Then, personalized tuning was performed for each T1D subject within a shorter scenario (two months). Finally, we evaluated the model on 10 virtual adult subjects over 12 months.
Results
Compared to baseline methods, i.e. standard bolus calculator, the overall mean TIR increased from 81.9±8.6% to 88.3±5.1% (p<0.005) with a non-significant decrease in hypoglycaemia.
Conclusions
This work presents a new meal insulin bolus recommender for T1D using DRL which has been proven to achieve, in silico, a significant improvement in glycaemic control.