Development of a weighted ensemble approach for prediction of blood glucose levels

Abstract

Type 1 diabetes (T1D) is a serious global problem. Several differential equation models have been proposed to model the blood glucose levels (BGLs) in the human body. However, a comparison of prediction of BGLs via non-recurrent models (multilayer perceptron or MLP), recurrent models (long short-term memory or LSTM), and statistical models (vector auto-regression or VAR) has received relatively little attention. This research's primary objective is to address this literature gap and propose BGL predictions via non-recurrent, recurrent, and statistical models.
A T1D dataset was compiled from a differential equation-based mathematical diabetes simulator, Automated Insulin Dosage Advisor (AIDA). A non-recurrent model (MLP), a recurrent model (LSTM), a statistical model (VAR), and a weighted ensemble model of the individual models were developed. In all models, the timeseries BGLs were divided into 80% training and 20% test. Results revealed that the recurrent LSTM model outperformed the non-recurrent MLP and statistical VAR models during training and test to account for the BGLs from the AIDA model. Furthermore, the weighted ensemble model performed the best to account for the BGLs from the AIDA model among all models. Overall, machine learning models may provide an alternative to differential equation models for simulating healthcare variables. We illustrate the significance of using ensemble models for blood glucose predictions.