Personalized LSTM-based alarm systems for hypoglycemia and hyperglycemia prevention

Abstract

Hypoglycemia and hyperglycemia prevention is the main challenge of an efficient Type 1 Diabetes (T1D) control. Alarm systems that alert the patients when their Blood Glucose (BG) levels are going to be critical can be useful instruments in order to react and avoid upcoming hypoglycemia and hyperglycemia events. These alarm systems can be used with both the conventional basal-bolus therapy or in conjunction with the advanced closed-loop control system, the so-called artificial pancreas. Model-based alarms use patient models to predict future BG levels and then activate alarms, so these models have to be reliable and to ensure good performances. In recent studies, neural network techniques for glucose forecasting obtained promising results, for both population and personalized models. These recent works showed that personalized Long Short-Term Memory (LSTM) models for BG predictions obtained good results on the 100 in silico patients of the most recent version of the UVA/Padova simulator. In this work personalized alarm systems for hypoglycemia and hyperglycemia prediction based on personalized LSTM models are proposed. Promising results have been obtained, detecting correctly the 77% of the hypoglycemia and the 89% of the hyperglycemia events.