Classification of Heart Health by LPC and MFCC Coefficients and Statistical Features

This article presents a human heart health status classification based on a public database that contains hospital recordings of various heart sounds using a digital stethoscope. Three different sounds were analyzed in this study; normal heart sounds, heart murmur sounds and extra systolic sounds. Ten general and statistical features and 16 parameters of the classical techniques; Linear Predictive Coding (LPC) and Cepstral Frequency-Mel Coefficients (MFCC) were calculated to create a numerical database. Through the genetic algorithm Galgo, the most significant features of all cardiac audio samples were extracted. These features were analyzed with three different classification methods; Logistic Regression (LN), Neural Networks (NN) and Suport Vector Machine (SVM). The area under the curve (ROC) and the accuracy (ACC) were the metrics used to evaluate the classifiers. The Neural Network classifier with the data normalized gave the best results with a ROC = 0.7558. While the same classifier with the data not normalized gave the worst results with a ROC = 0.6571.