Heart sound recognition technology based on convolutional neural network

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Rising heart disease mortality rates necessitate improved diagnostic tools.
• Heart sound auscultation is a fundamental clinical skill for diagnosing cardiac conditions.
• Current diagnostic methods require enhancement for greater accuracy and accessibility.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) algorithm for accurate heart sound signal classification.
• To improve the efficiency and effectiveness of diagnosing heart disease through automated auscultation analysis.
• To provide a supplementary tool for clinicians in identifying cardiac abnormalities.

Main Methods:

• Collected clinical and literature-based heart sound data.
• Preprocessed heart sound signals into 5-second grayscale images.
• Trained and optimized a CNN model using the prepared datasets.
• Evaluated the CNN model's performance on a separate test set.

Main Results:

• The CNN achieved a training accuracy of 95.17% and a loss of 0.23.
• Testing yielded an accuracy of 94.80%, sensitivity of 94.29%, and specificity of 95.54%.
• The algorithm demonstrated superior accuracy and sensitivity compared to existing methods, with an AUC of 0.943.

Conclusions:

• The proposed CNN algorithm effectively classifies heart sound signals.
• This method shows significant potential for assisting in heart sound auscultation and early heart disease diagnosis.
• The findings suggest a valuable application of AI in improving cardiovascular diagnostics.