Development of Convolutional Neural Network for Classification of Heart Sounds Utilizing Mel-Frequency Cepstral Coefficients

Abstract

Cardiovascular diseases are the leading cause of death worldwide, accounting for more than 30% of all global deaths. It is estimated that more than 17 million people lose their lives each year due to the heart attacks and strokes. In daily clinical practice, listening to and analyzing a patient's heart sounds via auscultation is essential in detecting any irregularities, such as murmurs, clicks, snaps, etc. This non-invasive initial screening procedure allows doctors to gather information about the heart’s condition without any invasive procedures, while being quick and cost-effective. The high prevalence and mortality rate of heart diseases have led to an increased need for automatic and efficient diagnostic tools that could assist medical professionals in the analysis of heart sounds. This paper introduces a convolutional neural network (CNN) model that is designed to identify and categorize various irregularities in heart sounds from stethoscope audio recordings. In order to train proposed CNN model, publicly available dataset containing hundreds of heartbeat audio recording was utilized. Since sound signals can be quite complex, it was necessary to extract the main characteristics of the sound and format them in a way that is suitable for training of proposed CNN model. Mel-Frequency Cepstral Coefficients (MFCCs) are designed to capture key spectral properties of a sound signal automatically and format them as small sets of numbers that are easily understandable by a deep learning model. Our proposed model achieved promising results, not only in detection of irregular heart sounds, but also in a classification between different types of heart sound abnormalities.