A study of artificial neural network training algorithms for classification of cardiotocography signals

Abstract

Cardiotocography (CTG) containing of fetal heart rate (FHR) and uterine contraction (UC) signals is a monitoring technique. During the last decades, FHR signals have been classified as normal, suspicious, and pathological using machine learning techniques. As a classifier, artificial neural network (ANN) is notable due to its powerful capabilities. For this reason, behaviors and performances of neural network training algorithms were investigated and compared on classification task of the CTG traces in this study. Training algorithms of neural network were categorized in five group as Gradient Descent, Resilient Backpropagation, Conjugate Gradient, Quasi-Newton, and Levenberg-Marquardt. Two different experimental setups were performed during the training and test stages to achieve more generalized results. Furthermore, several evaluation parameters, such as accuracy (ACC), sensitivity (Se), specificity (Sp), and geometric mean (GM), were taken into account during performance comparison of the algorithms. An open access CTG dataset containing 2126 instances with 21 features and located under UCI Machine Learning Repository was used in this study. According to the results of this study, all training algorithms produced rather satisfactory results. In addition, the best classification performances were obtained with Levenberg-Marquardt backpropagation (LM) and Resilient Backpropagation (RP) algorithms. The GM values of RP and LM were obtained as 89.69% and 86.14%, respectively. Consequently, this study confirms that ANN is a useful machine learning tool to classify FHR recordings.