A Comparative Study of Time Series Models for Blood Glucose Prediction

Abstract

AbstractDiabetes is a serious metabolic disorder which itself is a cause of many diseases. The only way to cure diabetes is to control blood glucose levels at regular intervals. Machine learning models play a significant role in the prediction of blood glucose levels. Data collected by blood glucose monitoring systems at regular intervals forms a time series data which requires the sequential layered network to predict the level of blood glucose. Various time series models like ARMA, ARIMA, LSTM and Bi-LSTM have been critically examined in this study. This research work compares the performance of state-of-the-art time series models for the prediction of blood glucose values for T1D patients. The experiment is conducted on the UVA/Padova dataset which consists of fifteen days of data of 30 patients (10 adolescents, 10 adults and 10 children) at four different meal timings. Results show that the Bi-LSTM model outperforms the LSTM, ARMA and ARIMA models. KeywordsBlood glucoseDiabetesPredictionLSTMBi-LSTMARMAARIMA