Optimization of an artificial neural network to study accelerations of foetal heart rhythm

Abstract

Given the importance of accelerations of the foetal health rate (FHR) in the monitoring of the foetal wellbeing along the course of the pregnancy, and taking into consideration the contribution of computerized analysis of biosignals as well as the emerging role of artificial intelligence in medicine, this study describes the optimization and use of artificial neural networks (ANNs) as a tool for predicting and investigating FHR accelerations. To this aim, nineteen features have been extracted from 187 FHR signals recorded from healthy women by cardiotocography. Three training methods, including Levenberg-Marquardt (LM), Scaled Conjugate Gradient (SCG), and Bayesian Regularization (BR), have been tested by training ANNs with increasing number of neurons in the hidden layer. The optimal network configuration has been selected by checking at the coefficient of determination (R2) and the Root Mean Square Error (RMSE). Results suggest that a proper ANN configuration not only enables maximizing the predictive capability of the selected model but, mostly, could be helpful in investigating the influence of linear and nonlinear indices of the FHR variability (FHRV) on the total number of accelerations in the foetal heart rhythm.