Connecting Blood Rheology to Physiology Through Machine Learning

• Kaitlin Vaccaro, Chemical Engineering, University of Delaware

• Antony Beris, Department of Chemical and Biomolecular Engineering, University of Delaware
• Norman Wagner, Department of Chemical and Biomolecular Engineering, University of Delaware

Blood is a colloidal suspension of cellular components such as red blood cells in aqueous plasma containing proteins. Physically, it has non-Newtonian rheological behavior which can be characterized by its yield stress and shear thinning viscosity. Accurate rheological modeling is essential for understanding, preventing, and treating disease, yet a major challenge is the variability of its rheological characteristics from donor to donor. Prior efforts have linked physiological factors such as hematocrit and fibrinogen to yield stress and viscosity using the Apostolidis Beris (AB) Regression Model and several machine‑learning approaches, with Gaussian Process Regression (GPR) previously providing the best performance on a 17‑sample dataset.

This work improves those correlations by implementing customizable neural networks (NNs) and redefining the objective function used during training. One NN minimizes the relative mean‑square error (RMSE) of yield stress and viscosity independently, while the second NN minimizes the RMSE of Casson‑predicted shear stresses computed over experimentally used shear rates. Both approaches were evaluated against the original AB equations and a refitting of those equations using the Horner dataset by testing their predictions on 11 new donors. RMSE values were computed using experimentally measured shear stresses as truth rather than Casson‑fitted stresses. Results show that the combined loss NN and the refitted AB equations outperform both the independent‑loss NN and the original AB model. Future work will extend this framework to predict the parameters required for modeling transient blood flow using time‑dependent rheological behavior.