Wearable Non-invasive Blood Glucose Estimation via Empirical Mode Decomposition Based Hierarchical Multiresolution Analysis and Random Forest

Wearable non-invasive blood glucose estimation plays an important role in the biomedical signal processing community. The common blood glucose estimation method is via the direct random forest algorithm. However, since the useful information of the signal is usually corrupted due to the low SNR, the distorted features inputted for the training algorithm result to a poor estimation performance. This paper proposes to employ an empirical mode decomposition (EMD) based hierarchical multiresolution analysis for performing the pre-processing and the random forest for performing the wearable non-invasive blood glucose estimation. More precisely, two levels of decompositions are employed in the EMD based hierarchical multiresolution analysis and only the first two intrinsic mode functions (IMF) in the second level of decomposition are discarded. Next, the features exacted from the processed near infrared (NIR) signal are trained via the random forest regression algorithm. The computer numerical simulation results show that the proposed method outperforms the classical method without the EMD pre-processing and with conventional EMD based pre-processing in terms of the average estimation accuracy and the distribution error shown on the Clarke error gird.