Joint Empirical Mode Decomposition and Singular Spectrum Analysis Based Pre-processing Method for Wearable Non-invasive Blood Glucose Estimation

Abstract

This paper proposes a joint empirical mode decomposition and singular spectrum analysis based pre-processing method for performing the blood glucose estimation. Here, both the near infrared signals and the electrocardiogram signal are acquired from a wearable non-invasive blood glucose estimation device. For the empirical mode decomposition based pre-processing method, the intrinsic mode functions are selected based on the correlations between the intrinsic mode functions and the acquired near infrared signals. For the singular spectrum analysis based pre-processing method, the singular spectrum analysis components are selected based on their mean frequencies. After that, four statistical features are extracted from each near infrared signal, as well as seven features of the heart rate variability parameters and the pulse transfer time are extracted from each electrocardiogram signal. They are used as the features. Then, the feature matrices and the corresponding reference blood glucose values are fed to the random forest for performing the training. In this study, the Clarke error grid is used to evaluate the estimation performance. The experiments show that the root mean square errors of the estimation are between 0.57mg/dL and 2.50mg/dL, the average of the root mean square errors is 1.44mg/dL and there are 80.35% of the points fall in the zone A of the Clarke error grid.