A New Robust Model for Heart Disease Detection on PCG signals using Entropygrams and a CNN

Abstract

Every year, 17.9 million people die from heart-failure–related conditions, making it the leading cause of mortality worldwide; early diagnosis could prevent many deaths. The most common way to detect cardiac abnormalities is through medical auscultation. Accurate diagnosis often depends on clinicians’ auscultation skills; however, they frequently have to listen in noisy environments. We propose a method for the automatic diagnosis of cardiovascular disease that is robust to noise. We extract entropy spectrograms from phonocardiograms to reliably convert audio signals into images, and then use a two-dimensional convolutional neural network (2D-CNN) to classify patients as healthy or unhealthy. To evaluate the method, we added white noise to the original recordings. The results show that entropy spectrograms are more robust than conventional feature-extraction techniques such as energy spectrograms or mel-frequency spectrograms.