JoWUA

Abstract

Cardiovascular disease (CVD) is the leading cause of death worldwide. Accurate and early diagnosis of cardiovascular disease (CVD) is essential for its timely treatment and management. However, this is challenging because traditional techniques for detecting heart diseases, such as auscultation, are highly subjective and prone to error. This study addresses this issue by building a novel customised deep learning architecture, SAINet, for automated CVD detection through heart sound analysis. Research is being conducted on the application of artificial intelligence (AI) to analyse phonocardiograms to detect CVD. This study aims to address this challenge by detecting heart disease using a novel customised neural network consisting of transfer learning techniques and convolutional neural networks to analyse heart sounds with increased accuracy, precision and recall and reduced computational complexity compared when compared to others.  Approximately 1000 recordings of heart sounds were used to train and test the model. Data augmentation was performed to increase the size of the training data. Two combinations of datasets were used in the experiments. The first combination consisted of two categories of heart sound recording: normal and abnormal. The second combination consisted of one normal and four different abnormal categories of heart sounds. An accuracy of 99.68% was achieved with the first combination, and 99.58% with the second combination. Both combinations yielded values above 99% for precision, recall, specificity, and the F1-score. The method proposed in this study is suitable for embedding CVDs in real-time devices such as an electronic stethoscope.