Comparison of Spectral and Sparse Feature Extraction Methods for Heart Sounds Classification

Cardiovascular diseases (CVDs) remain the leading cause of morbidity worldwide. The heart sound signal or phonocardiogram (PCG) is the most simple, low-cost, and effective tool to assist physicians in diagnosing CVDs. Advances in signal processing and machine learning have motivated the design of computer-aided systems for heart illness detection based only on the PCG. The objective of this work is to compare the effects of using spectral and sparse features for a classification scheme to detect the presence/absence of a pathological state in a heart sound signal, more specifically, sparse representations using Matching Pursuit with multiscale Gabor time-frequency dictionaries, linear prediction coding, and Mel-frequency cepstral coefficients. This work compares the performance of PCGs classification applying features as a result of averaging the samples or the features for each PCG sound event when feeding a random forest (RF) classifier. For data balancing, random under-sampling and synthetic minority oversampling (SMOTE) methods were applied. Furthermore, we compare the Correlation Feature Selection (CFS) and Information Gain (IG) for the dimensionality reduction. The findings show a SE=93.17 %, SP=84.32 % and ACC=85.9 % when joining MP+LPC+MFCC features set with an AUC=0.969 showing that these features are promising to be used in heart sounds anomaly detection schemes.