Deep learning for early detection of cardiovascular diseases via auscultation sound classification

Abstract

Heart diseases are one of the most prominent causes of death globally, which requires immediate and accurate diagnosis. The auscultation methods used in conventional medical practice, where the doctor listens to the sounds produced by the body without intervention is very ineffective because of the limitations in the actual skills and perception of the doctor. The main goal of this project will be designing a mobile-based system for the early detection of cardiovascular disease (CVD) by utilizing deep learning for auscultation sound classification. The approach involves the use of deep learning structures to classify cardiac sounds into normal and abnormal patterns on its own. Wavelet transformations, time-frequency representations, and Mel frequency cepstral coefficients (MFCC) have been used in feature extraction. The ResNet152V2 model showed high classification performance with area under the receiver operating characteristic curve (AUROC) of 0.9797 and 0.9636 on two datasets. Contrary to that, data augmentation, hyperparameter optimization, attention mechanisms, as well as input-output residual connections, led to better functionality and interpretability. This research seeks to overcome the limitations of traditional stethoscope use through the incorporation of sophisticated algorithms and the availability of mobile technology that could result in early diagnosis and prevention of CVDs, especially in underprivileged areas.