Murmur Detection from Phonocardiogram Recordings using Spectral Features

Imad Eddine Toubal, Yang Yang Wang, Filiz Bunyak, Kannappan Palaniappan
University of Missouri


We propose our method for heart murmur detection from Phonocardiogram recordings. Our method puts focus on the spectral features of the recording's audio signal. We make and test the assumption that murmur sound can occur under specific frequencies. First, for each auscultation locations (if exists), we extract N Mel-frequency Cepstrum Coefficients MFCCs. For each MFCC, we take simple statistics (mean, variance, and skew). In addition, we take into consideration, the change/difference in the signal over time (DMFCC). Random Forest classifier is used for predicting the existence of heart murmur given the patient demographic and audio features. The parameters chosen for the random forest are: 10 trees, 100 maximum number of leaf nodes in each tree. The random forest was optimized based on the Gini Index. We perform an ablation study on the training set of the CirCor DigiScope Dataset provided for the PhisioNet Challenge. We compare our results of different configurations that make use of demographic patient features, raw audio signal statistics (mean, variance, and skew), as well as MFCC and DMFCC statistics. Adding MFCC cepstral features provides significant improvement in classification accuracy, compared to temporal only features, that is increased when using a larger number of MFCCs. DMFCC further improves the accuracy and provides a best score of 516.13. Ongoing work focuses on improving the performance using additional features and time-series segmentation algorithms.