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

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Accurate detection of abnormal cardiac sounds is crucial for diagnosing heart conditions.
• Traditional methods often rely on subjective interpretation of auscultation data.
• Developing objective, automated methods for cardiac sound analysis is an ongoing challenge.

Purpose of the Study:

• To construct a deep learning model for predicting abnormal cardiac sounds.
• To enhance cardiac sound analysis by integrating novel feature extraction techniques.
• To improve the accuracy of automated cardiac sound classification.

Main Methods:

• Utilized diverse auscultation data from various body positions.
• Transformed cardiac signals into log-mel 2D spectrograms as input for a Convolutional Neural Network (CNN).
• Developed a multi-channel-based heart signal processing (MCHeart) scheme and the ReLCNN model incorporating residual blocks and Multi-Head Attention (MHA) mechanisms.

Main Results:

• The ReLCNN model, incorporating murmur features and a smoothing function, achieved a weighted accuracy of 83.6%.
• This represents a performance improvement of approximately 4% compared to the baseline LCNN model (79.6% accuracy).
• The integration of deep learning with expert-derived cardiac features demonstrated significant potential.

Conclusions:

• The proposed ReLCNN model effectively predicts abnormal cardiac sounds using processed auscultation data.
• The MCHeart scheme and ReLCNN architecture offer a promising advancement in automated cardiac diagnostics.
• This approach enhances the accuracy and objectivity of heart sound analysis.