Time Frequency Analysis-Based Averaging and Fusion of Features for Wearable Non-Invasive Blood Glucose Estimation

This paper proposes a feature averaging method based on the fusion of three classical time frequency analysis techniques, namely the discrete cosine transform (DCT), the singular spectrum analysis (SSA) and the empirical mode decomposition (EMD) based techniques, to reduce the measurement errors. Since different techniques can eliminate different inherent problems of the unreliable features, this paper proposes an improved stacking fusion strategy to fuse these three feature averaging methods together to perform the blood glucose estimation. Finally, the random forest is employed as the regression model. Two datasets are employed for demonstrating the effectiveness and the robustness of out proposed method. The computer numerical simulation results show that our proposed method can yield 90.5882% and 86.8421% of the test data in the first dataset and the second dataset falling in the zone A of the Clark error grid, respectively. Also, it can yield the mean absolute relative difference (MARD) at 11.33% and 9.84% for the data in the first dataset and the second dataset, respectively. Besides, our proposed method outperforms the existing methods for the photoplethysmograms (PPGs) in both datasets.