Neonatal heart disease screening using an ensemble of decision trees

This paper is concerned with the occurrence of a heart disease specifically for the neonate, as those seriously affected may face an increased risk of death. In this paper, a novel computer-based tool is proposed for a medical centre diagnosis aimed at monitoring neonates who are potential vulnerable to heart disease. In particular, cardiac cycles of phonocardiograms (PCGs) are first pre-processed and then used to train an ensemble of decision trees (DTs). The classifier model consists of 12 trees, with bagging and hold-out methods used for training and testing. Several feature encoding methods have been experimented with to generate the feature space over which the classifier has been tested, including Shannon energy and Wigner bispectrum. On average 93.91% classification accuracy, 96.15% sensitivity and 91.67% specificity have been obtained from the given data, which has been validated with a balanced dataset of 110 PCG signals taken from healthy and unhealthy medical cases.