Abstract

Info

Title:: Dual-stage machine learning framework for non-invasive blood glucose prediction based on photoplethysmography signals

Author(s):: WANG Meilin; WENG Xinkai; LING Yongquan; School of Information Engineering, Guangdong University of Technology, Guangzhou 510006, China

Keywords:: blood glucose photoplethysmography signals Savitzky-Golay filter feature importance ranking dual-stage prediction model random forest extreme gradient boosting model

Abstract:: A dual-stage non-invasive blood glucose prediction method based on photoplethysmography signals (PPG) is proposed in this study. PPG signals collected from 257 subjects are denoised using the Savitzky-Golay filter and corrected for baseline drift using the sparse weighted asymmetric least squares method. Subsequently, multidimensional features covering time domain, frequency domain and morphology are extracted, and feature selection is performed using extreme gradient boosting (XGBoost) and Pearson correlation coefficient (R) incremental threshold strategy. A dual-stage modeling scheme is adopted for the prediction model, with a random forest in the first stage to achieve preliminary prediction and XGBoost in the second stage to correct the prediction residuals. After verification in 40 independent subjects, the prediction Pearson correlation coefficient (R) reaches 0.844 3, and the coverage rate of Clarke error grid (CEG) A+B region is 92.5%. The results demonstrate that the constructed method outperforms existing similar non-invasive blood glucose prediction methods in prediction accuracy and CEG distribution performance.