Classification of Heart Sounds Using Chaogram Transform and Deep Convolutional Neural Network Transfer Learning

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Heart sounds offer critical diagnostic information for cardiac conditions.
• Current heart sound classification methods are explored in telemedicine, digital signal processing, and machine learning for rapid cardiac pathology identification.

Purpose of the Study:

• To introduce chaogram, a new transform for converting heart sound signals into color images.
• To enable the application of deep convolutional neural networks and transfer learning to phonocardiogram (PCG) signal analysis.

Main Methods:

• The chaogram transform projects the phase space representation of PCG signals onto three coordinate planes, creating color images.
• The InceptionV3 model was utilized for classification on the PhysioNet dataset.

Main Results:

• The chaogram method successfully converted heart sound signals into images suitable for deep learning models.
• The InceptionV3 model achieved a classification score of 88.06% on the imbalanced PhysioNet dataset, using the average of sensitivity and specificity.

Conclusions:

• Chaogram is a viable method for transforming PCG signals into image data for machine learning.
• This approach facilitates the use of advanced deep learning techniques, like InceptionV3, for improved cardiac pathology detection.