Comparison of Feed-Forward Neural Network training algorithms for oscillometric blood pressure estimation

Abstract

Feed-Forward Neural Network (FFNN) has recently been utilized to estimate blood pressure (BP) from the oscillometric measurements. However, there has been no study till now that consolidated the role played by the different neural network (NN) training algorithms in affecting the BP estimates. This paper compares the estimation errors in the BP due to ten different training algorithms belonging to three classes: steepest descent (with variable learning rate, with variable learning rate and momentum, resilient backpropagation), quasi-Newton (Broyden-Fletcher-Goldfarb-Shanno, one step secant, Levenberg-Marquardt) and conjugate gradient (Fletcher-Reeves update, Polak-Ribiére update, Powell-Beale restart, scaled conjugate gradient) that are used to train two separate NNs: one to estimate the systolic pressure and the other one to estimate the diastolic pressure. The different training algorithms are compared in terms of estimation error (mean absolute error and standard deviation of error) and training performance (training time and number of training iterations to reach the optimal weights). The NN-based approach is also compared with the conventional maximum amplitude algorithm.