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