Multi-Class Classification and Prediction of Heart Sounds Using Stacked LSTM to Detect Heart Sound Abnormalities

The changes in lifestyle, food habits, and working conditions cause various diseases in human lives, cardiovascular diseases are one of those. Not only aged people, middle-aged and young people are also suffering due to this and lead to death in the early ages. So there is a significant need in detecting cardiovascular diseases in beginning itself. Through early detection and persistent treatment, the death rate in the early ages due to cardiovascular diseases can be reduced. However, it is necessary to have an efficient model to detect heart disease at an early stage even without the presence of a trained clinical expert. This paper studies the implementation of deep learning models to classify heartbeat sounds into various classes. We proposed a stacked LSTM model to classify the heartbeat sound into multiple classes based on the features obtained from the audio signals. The implementation can even predict the class of an unlabelled heartbeat sound. The model classifies the heartbeat sounds into 4 classes with accuracies 85% and 87% on training and validation sets respectively. In further the proposed model parameters can be improved to increase the classification and prediction accuracy.