Personalized LSTM models for glucose prediction in Type 1 diabetes subjects

Abstract

Blood Glucose (BG) regulation is the main challenge of Type 1 diabetes patients, that need exogenous insulin treatments to keep BG levels under control. The Artificial Pancreas (AP) is a closed-loop system developed to automatically regulate BG, using continuous glucose monitoring data to compute and directly inject the optimal insulin quantity. The AP includes controllers, like the Model Predictive Control, and alarm systems that need patients models to predict BG levels. In recent years, thanks to a growing availability of both in silico and in vivo data, neural network techniques have been investigated for glucose prediction, showing a good capability to understand the non-linearities of the system. In this work, personalized LSTM models are proposed for the 100 patients of the UVA/Padova simulator. BG levels are predicted with a prediction horizon of 40 minutes, considering meals, insulin and past BG values as input. The proposed LSTMs with a simple architecture obtained satisfactory results with respect to the literature (RMSE=7.67, FIT=75.86%).