A cardiac audio classification method based on image expression of multidimensional features.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Heart sound auscultation is vital for early cardiovascular disease diagnosis.
• Current automatic heart sound classification methods often rely on traditional classifiers and segmentation features, not fully utilizing deep learning potential.

Purpose of the Study:

• To propose a novel cardiac audio classification method (CACIEMDF) that leverages deep networks by transforming multidimensional heart sound features into image representations.
• To improve the accuracy and efficiency of automatic heart sound classification for cardiovascular disease diagnosis.

Main Methods:

• A 102-dimensional feature vector was engineered from time, frequency, and statistical domains of heart sound data.
• Principal Component Analysis (PCA) and convex hull algorithm were used to create a 2D feature projection space, generating 2D coordinate representations for each feature component.
• Feature component values and their category divergence were mapped to the RGB channels of a color image, enhanced with Gaussian model dyeing, and then classified using a deep network (ResNet50).

Main Results:

• The proposed CACIEMDF method, when combined with ResNet50, achieved high classification accuracy on fused public heart sound datasets.
• Specifically, accuracies of 95.68% for two-class classification and 94.53% for five-class classification tasks were obtained.

Conclusions:

• The CACIEMDF method effectively converts complex multidimensional heart sound features into image data suitable for deep network analysis.
• This image-based deep learning approach significantly enhances the accuracy of automatic heart sound classification, offering a promising tool for cardiovascular disease diagnosis.