Phonocardiography (PCG)-based early detection of heart anomalies using deep learning

Abstract

A simple-to-use and cost-effective Phonocardiography (PCG)-based system is proposed for the early detection of cardiovascular anomalies. A summary is provided related to the main results of a comparative study of various signal processing and deep learning techniques applied to PCG signals to automatically differentiate normal heart sounds from five types of common murmurs. Initial results indicate an average of over 92 % and 71 % heart anomaly detection and classification rates, respectively. This is accomplished using a 5-layer-Deep Neural Network (DNN) with 100 neurons per layer using the Hyperbolic Tangent (tanh) as an activation function. The Discrete Wavelet Transform (DWT) and Heart Sound Envelogram (HSE) were found to be best in respectively denoising and segmenting the PCG signal. Mel Frequency Cepstral Coefficients (MFCC) Features were found to outperform their Frequency and Time Domain counterparts and were used to train and run the DNN. The proposed system should be useful in providing early warnings and initial indications to potentially sick people directing them to cardiologists in order to pursue more accurate diagnostics, making it a handy home-care tool in the context of smart and preventive public health systems.