MCHeart: Multi-Channel-Based Heart Signal Processing Scheme for Heart Noise Detection Using Deep Learning

Figure 5. Spectrogram images of proposed multi-channel features.

Figure 1. Overview of the proposed approach. We show how peak intervals, boundaries, signal envelopes, and log-mel spectrograms are extracted from phonocardiogram (PCG) raw audio data and combined with models to predict murmur (abbreviations: PCG = phonocardiogram; S1 and S2 = the two distinct sounds produced by the heart during each heartbeat; LCNN = light convolutional neural network; FC = fully connected).

Figure 2. Absence vs. presence of heart murmur.

Figure 3. Fundamental heart sounds and heart murmurs.

Figure 4. Our proposed heart feature extraction framework. This framework consists of a total of six stages visually depicting how the original PCG signal transforms through each stage.

Figure 6. Proposed model architecture.

Figure 7. Model blocks. (a) LCNN block architecture. It has a convolutional layer with max-feature-map activation. Next, max pooling, batch normalization, and Swish activation are applied sequentially. (b) Residual block. The input feature map is added to the output feature map as a residual connection.

Figure 8. The auscultation locations: aortic valve (AV), pulmonic valve (PV), tricuspid valve (TV), and mitral valve (MV).

Figure 9. Overview of the automated murmur detection (AMD) system.

Figure 10. We have depicted the performance results for 12 models across six evaluation metrics in barplots. According to the weighted accuracy and cost metric used in the PhysioNet Challenge 2022, the ReLCNN-9 model and the ReLCNN-3 model exhibited the most favorable outcomes. Excluding the cost metric, the ReLCNN-9 model appears to be generally superior.

Figure 11. Extraction of cardiac sound envelopes based on the window size parameter from smoothing.