Recurrent Neural Networks in Prediction of Blood Flow in Hybrid-Digital Model of Cardiovascular System

The goals of the paper were the development of a new method of estimating end-systolic elastance of the left ventricle and the proposition of a new way of estimating pressure and volumetric blood flow occurring in the circulatory system. 6 recurrent neural network models were trained using 2 separate datasets. The purpose of the models was regression of circulatory system parameters. Datasets were collected using a hybrid-digital model of the cardiovascular system. Feature selection was performed based on correlation analysis and literature research. Neural networks’ architecture was designed based on experiments and literature research. Bayesian optimization was applied to select the final version of the architecture. Empirical procedure followed by the Bayesian optimization helped setting the details of the stochastic gradient descent learning algorithm. Final tests showed that models realizing the estimation of end-systolic elastance as well as those performing blood pressure and volumetric flow regression gave satisfactory accuracy on the test datasets.