Heart Murmur Classification from PCGs Based on Multi-Source Features

Peng Wang
Aerospace Information Research Institute, Chinese Academy of Sciences (AIRCAS)


Phonocardiography (PCG) is the one of non-invasive ways to diagnose different heart abnormalities in the primary care phase. The algorithmic classification of heart murmurs in the PCG signal is the goal of the Physionet/CinC Challenge 2022. In this work, we proposed a method of heart murmur classification from multi-channel PCGs based on multi-source features. Firstly, the S1/S2/systolic/ diastolic of each PCG channel are segmented using a hidden semi-Markov model, and morphological features of them containing the time domain and the frequency domain features are extracted. Secondly, the deep residual neural network (ResNet) architecture, proven useful in PCG classification, was used to learn potential features from original PCG recordings. Then the random forest (RF) classifier was used to select the most important features learned by deep ResNet. Finally, the above morphological features, features selected by RF, and demographic data were concatenated to train a XGboost classifier. The proposed algorithm was evaluated on the public training dataset, and achieved a 5-fold cross-validation score of 821 by using the Challenge metric. Our team, PhysioDreamfly, ranked 15th out of over 60 teams by an official validation score of 736 on the hidden validation set.