Classification and Feature Selection Approaches for Cardiotocography by Machine Learning Techniques

Abstract

Cardiotocography (CTG) is the commonly used tool to monitor fetal distress (hypoxia), other fetal risks such as fetal heart rate, and autonomous nervous system maturation. If not rectified in the early stages, these problems may lead to fetal death. Thus, it is important to know which selected features are necessary to predict the risk. The objective of this research is to carry out the classification model and feature selection on the derived dataset with R-based CARET and Python-based Scikit learn packages. Despite different analytical techniques used, it is observed that the nature of the tools may play a role in model classification on the given dataset. The classification accuracies of the dataset are found to be similar when compared with the UCI repository CTG dataset. The similar performance of accuracies has been noticed in the random forest and naive Bayes, and average accuracy with respect to complete features (R-based machine learning techniques). On the other hand, the selected features showed classification accuracies with similar performance in naïve Bayes, bagging and boosting (Pythonbased machine learning techniques). However, the study found that correlated features contributed to the increase of classification accuracy of complete features. The selected features show the accuracies similar to the complete dataset indicating as these features play a role in the prediction of CTG data.