Classifying Heart Sounds for Disease Detection Using Deep Learning Methods

Heart diseases are known to be among the notable causes of mortality in the world. Heart diseases can be of different types, viz., vascular, ischemic, and hypertensive heart disease. The public datasets were been collected from the PhysioNet/CinC 2016 Challenge. In this study, we propose 2 different neural network models, for classifying heart sounds for disease (abnormality) detection using deep learning methods. ‘Normal’ category denotes that a person has normal heart sounds. ‘Abnormal’ category denotes that a person has abnormal heart sounds. Each category expresses a specific type of disease. To achieve better results for our binary classification algorithms, we perform a data pre- processing step that involves converting input audio signals to spectrograms. Experiments were conducted with a convolutional neural network (CNN) which is hinged on ResNet18 model and a neural network that is an extension of ResNeXt50 model. In both of these architectures, a novel component called the Shuffle Attention layer has been introduced. It is used to selectively attend to relevant spatial features and enhance the representation of the feature maps. Shuffle Attention layer is implemented using two main components: the squeeze-excite blocks and the channel shuffle layer. The squeeze-excite blocks help in reducing the spatial dimensions of the feature maps to a single value and the channel shuffle layer is used to increase the connectivity between channels and reduce overfitting. The results showed that ResNet18 model with accuracy of 91.27% is more efficient than the ResNeXt50 model having accuracy 88.72%. The results observed are constructive and encouraging for heart disease dataset. Thus, we were able to demonstrate that neural-network based models could help in classifying an unbalanced heart disease dataset. Our proposed framework can also produce highly accurate models suitable for real-world clinical data and diagnostic use.