Differentiating S1, S2 Noises from Abnormal Heart Sounds Generated in Closure of Atrioventricular and Semilunar Valves using MFCC and LSTM

Abstract

Heart Sounds are generated by heartbeats in the form of noises. Each beat in a healthy human produces two normal heart sounds, which are sometimes referred to as a lub and a dub (or dup). For example, S1 and S2 heart sounds are produced by the closing of both atrioventricular and semilunar valves. This heart sound dataset has been collected in real-world circumstances and frequently incorporates various types of background noises. Differences in heart sounds that correspond to different cardiac symptoms might be quite subtle and difficult to distinguish. To overcome this type of situation, a deep learning-based classification model named long short-term memory (LSTM) with RNN (Recurrent Neural Network) technology has been proposed in this project. This approach featured with unique additive gradient structure that can differentiate the heart sounds i.e. S1 and S2 from abnormal noises. Along with Mel Frequency Cepstral Coefficient (MFCC) at the pre-processing step, ensures the patients' successful representation of heart sounds present in audio files into their respective frequency domain forms. Although, MFCC can move forward with four basic processes that include Fourier Transform (FT), mapping to Mel Scale, use of triangular windows, and log of each Mel frequency. Moreover, it may be difficult to solve problems like classifying heart sound that requires long-term memory due to having gradient problems present in RNN. Such problems can be overcome by applying both RNN and LSTM techniques as LSTM maintains the information in the memory cell for a longer period. To restrict the amount of data passed through the cell, this proposed model uses three sigmoid gates and one tanh layer. When information enters into the memory, when to output, and it is forgotten, that is controlled by a set of gates that allow for greater control of the gradient flow and maintenance of long-range dependencies. This RNN-LSTM model has been successfully implemented and differentiated heart sounds with a greater test accuracy result of 80 percent.