A Fusion of Handcrafted Features and Deep Learning Classifiers for Heart Murmur Detection

Abstract

As part of George B. Moody Physionet Challenge 2022, our team Melbourne Kangas, proposed an algorithm for identifying abnormal heart sounds from paediatric phonocardiograms (PCGs). We developed a Deep Learning (DL) approach and a handcrafted feature-based approach. The DL classifier was based on bidirectional long-short-termmemory and Mel-frequency cepstrum coefficients from raw PCG signals. The feature-based approach used nonnegative matrix factorisation to denoise PCG signals and then extracted the features based on the whole and segmented recordings, followed by feature selection. A random under-sampling boosting classifier for murmur classification and robust boosting classifier for outcome classification were given the subset of features. The feature-based performed better than the DL classifiers on the validation set. The feature-based classifier received a weighted accuracy of 0.632 (29th out of 41 teams) and a challenge cost of 11,735 (3rd out of 39 teams) on the test set. Decision fusion of the two approaches decreased 10-fold crossvalidation results.