Enhanced Classification Performance of Cardiotocogram Data for Fetal State Anticipation Using Evolutionary Feature Reduction Techniques

Abstract

Role of computers became inevitable in healthcare sector and computers with information and communication technologies are found to be widely used for assessment, patient monitoring, documentation, and telemedicine. Data mining is a field which helps to obtain knowledge from massive amount of data from any industry or organization. Cardiotocography (CTG) is a test that is done during the third trimester of pregnancy to measure the heart rate and movements of fetus and helps to monitor the contractions in the uterus and thereby for monitoring the signs of any distress, before the delivery of baby and during the labour. The physical interpretation of information from CTG is found to be a challenging task, and any contradictory interpretation will lead to erroneous 372diagnosis on fetal condition, which may lead to the extent of fetal death. Feature selection is the process in which an optimal subset of features is selected based on some defined criterion which helps to considerably improve the performance of classification in terms of learning speed, accuracy of prediction, simplicity of rules, etc. Also, the reduction in size of feature subset helps to remove noise and irrelevant features. Several approaches have been introduced for improving the performance of computerized classification of CTG data which leads to an improved diagnosis of fetal status. In this chapter, Filter and Wrapper feature selection techniques are applied to CTG dataset available in UCI machine learning repository. Evolutionary algorithms such as genetic algorithm, firefly algorithm, and a hybrid technique incorporating information gain and opposition-based firefly algorithm have been used to improve the classification performance of CTG dataset. The results of simulations show that the proposed methodologies are highly promising when compared to the other existing methods. To assess the performance of these proposed methodologies, various performance measures namely accuracy, sensitivity (or) recall, specificity, precision (or) positive predictive value, negative predictive value, geometric mean, F-measure, and area under ROC have been used and the hybrid model incorporating information gain and ppposition-based firefly algorithm proves better performance than the other techniques.