Presenter of 1 Presentation
A REINFORCEMENT LEARNING BOLUS CALCULATOR WITH NO MEAL INFORMATION FOR PATIENTS WITH TYPE 1 DIABETES
Abstract
Background and Aims
In hybrid artificial pancreas systems (HAPs) insulin boluses are usually calculated based on patient estimations of the amount of carbohydrates to be ingested. The aim of this study is to calculate the bolus insulin without knowing the patient’s carbohydrate intake, thus alleviating the patient’s management burden.
Methods
A Q-Learning agent (QLA) was trained to optimize bolus insulin doses for in-silico type 1 diabetic patients. The area under the curve of glucose profile, maximum and minimum glucose values were defined as states. The glucose value before meal was utilized to define the range of bolus values in the action space to restrict the exploration of the QLA in a safe zone.
Results
The algorithm was tested for a cohort of 68 virtual patients and the results were compared to the standard bolus calculator (SBC) in open loop therapy. The results are given as median (interquartile range). A mean glucose value of 153.57 (145.54 - 166.88) vs 154.48 (145.52 - 164.21); p=0.0027, time below range of 0.049 (0.04 - 1.15) vs 1.17 (0.41 - 2.34); p=0.000000642, time in target range of 72.37 (59.99 - 81.94) vs 69.64 (61.56 - 77.40); p=0.0096 and time above range of 1.38 (0.27 - 4.68) vs 1.59 (0.7 - 4.44); p=0.645 were achieved for SBC and QLA respectively.
Conclusions
The reinforcement learning methodology using Q-Learning to compute insulin boluses without information on the amount of carbohydrates in meals showed similar performance as compared to the SBC.