Rheumatic Heart Disease Detection Using Deep Learning from Spectro-Temporal Representation of Un-segmented Heart Sounds.

Abstract

Rheumatic Heart Disease (RHD) is an autoimmune response to a bacterial attack which deteriorates the normal functioning of the heart valves. The damage on the valves affects the normal blood flow inside the heart chambers which can be recorded and listened to via a stethoscope as a phonocardiogram. However, the manual method of auscultation is difficult, time consuming and subjective. In this study, a convolutional neural network based deep learning algorithm is used to perform an automatic auscultation and it classifies the heart sound as normal and rheumatic. The classification is done on un-segmented data where the extraction of the first, the second and systolic and diastolic heart sounds are not required. The architecture of the CNN network is formed as an array of layers. Convolutional and batch normalization layers followed by a max pooling layer to down sample the feature maps are used. At the end there is a final max pooling layer which pools the input feature map globally over time and at the end a fully connected layer is included. The network has five convolutional layers. This current work illustrates the use of deep convolutional neural network using a Mel Spectro-temporal representation. For this current study, an RHD heart sound data set is recorded from one hundred seventy subjects from whom one hundred twenty four are confirmed RHD patients. The system has an overall accuracy of 96.1% with 94.0% sensitivity and 98.1% and specificity.