Estimation of Non-Newtonian Patient-Specific Blood Viscosity Via Photoplethysmography Using a Neural Network with a Physics-Informed Loss Function
12 Pages Posted: 1 Oct 2024
Abstract
Blood viscosity information can be used to monitor and diagnose various circulatory system diseases. Therefore, if blood viscosity can be calculated from wearable PPG, the versability of non-invasive health monitoring system can be broadened. The aim of this study is to extract a patient-specific viscosity equation from PPG data. Shear-viscosity equation derived from viscometer was used as ground-truth data. Signal obtained from wearable device was processed with noise filtering, wandering elimination to gain stable blood pressure wave. 1D convolution network model was used as base AI model. Physics-informed loss function and k-fold cross-validation method was used along with weight factor optimization for training the algorithm. Results have shown that accuracy of final algorithm was 81.1%. The accuracy in the phycological shear range (50~300s-1) was 84.0%, which was higher than other low and high shear regions. Accuracy and contribution of each parameter for Carreau-Yasuda model was also analyzed. The result has shown that contribution of each parameter varies by shear range, providing insight for weight factor optimization. By obtaining viscosity from PPG, the versability and usability of wearable healthcare system might be broadened to target various diseases related to circulatory system.
Note:
Funding Information: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A5A1022977)
Conflict of Interests: No competing interests to declare
Ethical Approval: Yonsei University Health System, Severance Hospital, Institutional Review Board (IRB) - Number: 4-2023-0648
Trial registration: Not relevant
Keywords: Blood viscosity, Photoplethysmography, Wearable healthcare, Artificial intelligence, Physics-informed loss function
Suggested Citation: Suggested Citation