Extracting the Features of Heart Sounds and Identifying Murmurs Based on Mel Frequency Ceptral Coefficients

This paper examines the use of Mel-frequency Ceptral Coefficients in the identification of PCG (phonocardiography) signal. Various Heart sounds are analyzed to extract their coefficients. These coefficients are reduced using principal component analysis. Multi-layered Perceptron is trained using principal components. The network is trained for MFCC values of normal heart sound. This trained network is then used to classify different input heart hound samples. By training and testing the network on a different number of coefficients, the optimum number of coefficients to include for identifying abnormality with heart functioning is identified. We conclude that using 4 principal components from 13 coefficients gives more accurate classification results.