Multi-View Cross-Fusion Transformer Based on Kinetic Features for Non-Invasive Blood Glucose Measurement Using PPG Signal

• Published in: IEEE journal of biomedical and health informatics Vol. 28; no. 4; pp. 1982 - 1992
• Main Authors: Chen, Shisen, Qin, Fen, Ma, Xuesheng, Wei, Jie, Zhang, Yuan-Ting, Zhang, Yuan, Jovanov, Emil
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.04.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Biomedical monitoring; Blood; Blood glucose; Blood glucose estimation; deep learning; Diabetes mellitus; Error analysis; Feature extraction; Fitting; Glucose; Kinetic theory; Kinetics; non-invasive measurement; Optimization; photoplethysmography; Physiology; Root-mean-square errors; Signal processing; Time-frequency analysis; Transformers
• ISSN: 2168-2194; 2168-2208; 2168-2208
• DOI: 10.1109/JBHI.2024.3351867

Noninvasive blood glucose (BG) measurement could significantly improve the prevention and management of diabetes. In this paper, we present a robust novel paradigm based on analyzing photoplethysmogram (PPG) signals. The method includes signal pre-processing optimization and a multi-view cross-fusion transformer (MvCFT) network for non-invasive BG assessment. Specifically, a multi-size weighted fitting (MSWF) time-domain filtering algorithm is proposed to optimally preserve the most authentic morphological features of the original signals. Meanwhile, the spatial position encoding-based kinetics features are reconstructed and embedded as prior knowledge to discern the implicit physiological patterns. In addition, a cross-view feature fusion (CVFF) module is designed to incorporate pairwise mutual information among different views to adequately capture the potential complementary features in physiological sequences. Finally, the subject-wise 5-fold cross-validation is performed on a clinical dataset of 260 subjects. The root mean square error (RMSE) and mean absolute error (MAE) of BG measurements are 1.129 mmol/L and 0.659 mmol/L, respectively, and the optimal Zone A in the Clark error grid, representing none clinical risk, is 87.89%. The results indicate that the proposed method has great potential for homecare applications.