Heart Sound Classification using a Hybrid of CNN and GRU Deep Learning Models

Abstract

Auscultation is a process where a stethoscope is used to listen to the heart sound signal to analyse the heart’s functionality. Due to the stethoscope’s non-invasiveness, convenience, and cost-effectiveness, it is the most common primary screening tool medical fraternities use. However, the scarcity of medical experts and the subjectivity in the analysis hinders the reliability of diagnosis using auscultation. Therefore, computer-aided analysis of heart sound signals will be helpful in this scenario. This paper presents a hybrid deep learning-based method to classify the heart sound signal into five classes. The method begins with the signal pre-processing followed by decomposition using Discrete Wavelet Transform (DWT) up to five levels. The obtained DWT coefficients are used to train the hybrid model, composed of two Convolution neural network (CNN) layers following one Gated Recurrent Unit (GRU) network layer. CNN models are suitable for extracting meaningful features, while the GRU exploits the time-dependent features. This combination helps classify the heart sound signal since they exhibit complex quasi-cyclic features. An overall accuracy of 99.3% is obtained for a publicly available dataset. It shows the proposed method’s efficacy for classifying heart sound signal and superiority over the existing methods. Such a method will be beneficial in reducing the burden of heart valve diseases by early detection of diseases and initiating the proper medication.