Blood-glucose prediction : Comparing insulin treatment methods

Abstract: Type 1 diabetes requires its patients to inject artificial insulin in their bodies to control their blood-glucose levels. This can to some extent be automated through the use of insulin pumps and continuous blood-glucose monitoring systems, enabling automatic insulin injections and automatic blood-glucose measurements. To inject an appropriate amount of insulin, a prediction of the future blood-glucose values has to be made, the accuracy of which dictates how autonomous such a system can be. In this paper, the performance of a machine learning model is examined, when using data from different insulin treatment methods. The two treatment methods compared are the closed-loop insulin pump system and the traditional insulin pump system. By training a convolutional recurrent neural network separately on the different datasets, the resulting models were compared on four different performance metrics; root-mean-square error, mean average relative difference, Matthews correlation coefficient for hypoglycemia, and Matthews correlation coefficient for hyperglycemia. While the results showed some indication of the closed-loop models being better, the differences were too small to be statistically significant. To get more conclusive results, a study involving more clinical patients would be needed.