Blood Glucose Monitoring Using Non-Invasive Features of Wearable Devices and Machine Learning

Abstract: With the increasing number of people with diabetes and the popularity of wearable devices, it becomes a novel research direction to use sensor data from wearable devices to monitor blood glucose in healthy people. There are many methods used for blood glucose prediction, such as using NIR and PPG for accurate blood glucose prediction, however, these data are usually not easy to acquire, and using data from wearable devices may be a potential direction. In this paper, two methods were proposed using the latest dataset released by Physiconet in 2023, 1-D CNN and tsfresh, for feature extraction of sensor data, and then construct a GRU network to predict blood glucose values, comparing the effectiveness of the two methods. Besides, the effects of different sensor features on the prediction results are explored. The study shows that the features extracted by tsfresh can follow the trend of blood glucose changes well, but it is still far from predicting specific blood glucose value. A larger dataset is needed to make a conclusion on the feasibility of the study.