Classification of heart sounds using linear prediction coefficients and mel-frequency cepstral coefficients as acoustic features

This article presents a method that uses Linear Prediction Coefficients (LPC) and Mel-Frequency Cepstral Coefficients (MFCC) as features to classify normal and abnormal cardiac sounds. Three different feature vectors were tested: LPC-only, MFCC-only and LPC + MFCC. Different experiments were made with three classifiers: Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and Random Forests, using 238 samples (150 normal and 133 abnormal). Results show that the best performance was obtained for the combination of LPC + MFCC as features vectors, plus SVM and KNN as classifiers, with an accuracy of 94.6%, a specificity of 98.6% and a sensitivity of 89.4%.