Heart Sound Classification using Deep Learning

Cardiac auscultation is the process of listening to the sounds of the heart with a stethoscope, which can provide important diagnostic information about a patient’s heart function. It is a key component of a physical examination and can help doctors identify potential heart problems such as murmurs, arrhythmias, and valve disorders. Deep learning-based cardiac auscultation is of great interest to medical professionals as it can reduce the burden of manual auscultation by automatically detecting abnormal heartbeats. However, automatic cardiac auscultation is a challenging task due to the need for dependable and precise results, as well as the interference of background noise in heart sounds. In this work, we propose a heart sound classification method that could potentially accelerate automated cardiac auscultation. The method is based on Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNN). We use heart sounds from the dataset Classifying Heart Sounds Challenge and extract input features using Mel Frequency Cepstral Coefficients (MFCCs). Further, we examine the use of a simple data augmentation technique and show that adding noise, stretching, and shifting of heart sound signals significantly improves the classification accuracy of the normal, murmur, and artifact heart sounds.