Analysis and Comparison of Machine Learning Models for Glucose Forecasting

Abstract

Continuous blood glucose monitoring (CGM) is a central aspect of the modern study of diabetes. It is also a way of improving the quality of life of patients. To make appropriate decisions for patients with diabetes, it needs an effective tool to monitor these levels in order regarding insulin administration and food intake to keep blood glucose levels within the range target. Efficient and accurate prediction of future blood sugar levels repeatedly benefits the diabetic patient by helping them to reduce the risk of blood sugar level extremes, including hypoglycemia and hyperglycemia. In this study, we implemented several time-series models, including statistical and machine-learning-based models, using two direct and recursive strategies, to forecast glucose levels in patients. We applied these models to data collected from 171 patients in a clinical study. For the 30-min prediction horizon, the average of mean absolute percentage errors (MAPEs) and root mean squared errors (RMSEs) for each model respectively shows that ARIMA, XGBoost, and TCN can yield more accurate forecasts. We also highlight the difference between statistical and machine-learning-based models, where statistical models perform effectively in predicting CGM levels, although they cannot perceive changes in variation, like neural-network-based models.