Machine learning for predicting microvascular network hemodynamics

We investigate the applicability of the machine learning (ML) techniques for the prediction of time-averaged and time-dependent blood flow rate and red cell distributions in physiologically realistic and large capillary vessel networks. To train and test the ML models, we acquire data from high-fidelity simulations of the flow of deformable red blood cells suspension in the networks. We use two networks that are geometrically quite different, with one used for the model building and the other for prediction. For the prediction of time-averaged blood flow rate, a regression-type ANN model is used, while for the prediction of time-average RBC distribution, a classification-type ANN model is used. With the flow rate and hematocrit specified at an inlet of a vasculature, the models predict time-averaged flow rate and RBC distributions in the entire network. For the prediction of time-dependent quantities, we use LSTM models. With these, we first predict the time-dependent flow rate and hematocrit in each vessel in isolation over a long time, as well as such quantities in isolated vascular bifurcations. These models constitute a model bank which is then used to predict simultaneous spatially and temporally evolving quantities through the vessel hierarchy in the networks.