Enhancement of phonocardiogram segmentation using convolutional neural networks with Fourier transform module.

Abstract

The automated identification of the first and second heart sounds (S1 and S2, respectively) in phonocardiogram (PCG) signals plays a pivotal role in the detection of heart valve diseases based on the known occurrence of heart murmurs between S1-S2 or S2-S1 in valve disorders. Traditional neural network-based methods cannot differentiate between heart sounds and background noise, leading to decreased accuracy in the identification of crucial cardiac events. Therefore, a deep learning-based segmentation on PCG signals that can distinguish S1 and S2 heart sounds with the Convolutional Fourier transform (CF) modules, which are two sequentially connected CF modules, was proposed in this study. Internal datasets, alongside the publicly available PhysioNet 2016 dataset, were used for the training and validation of the CF modules to ensure a robust comparison against existing state-of-the-art models, specifically the logistic regression-Hidden semi-Markov model (LR-HSMM). The efficacy of the CF modules was further evaluated using external datasets, including the PhysioNet 2022 and the Asan Medical Center (AMC) datasets. The CF modules exhibited superior robustness and accuracy in segmenting S1 and S2, achieving an average F1 score of 97.64% for S1 and S2 segmentation, which indicated better performance compared with that of the previous best model, LR-HSMM. The integration of the CF modules ensures the robust performance of PCG segmentation even amidst heart murmurs and background noise, significantly contributing to the advancement of cardiac diagnostics. All code is available at https://github.com/mi2rl/PCG_FTseg.