CNN-Based Heart Sound Classification with an Imbalance-Compensating Weighted Loss Function.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Heart sound auscultation is crucial for early heart disease diagnosis.
• Automated heart sound classification using deep neural networks is an emerging field.
• Class imbalance in heart sound datasets poses a significant challenge.

Purpose of the Study:

• To propose and evaluate a deep Convolutional Neural Network (CNN) model for classifying normal and abnormal heart sounds.
• To utilize two-dimensional Mel-scale features, specifically Mel frequency cepstral coefficients (MFCCs) and Log Mel spectrum, as input for the CNN model.
• To address the class imbalance issue using weighted loss functions during model training.

Main Methods:

• Development of a deep CNN model for heart sound classification.
• Input features included two-dimensional Mel-scale representations: MFCCs and Log Mel spectrum.
• Training employed two weighted loss functions to mitigate class imbalance.
• Model validation was performed on the PhysioNet/CinC 2016 heart sound database.

Main Results:

• The proposed CNN model achieved an Unweighted Average Recall (UAR) of 89.6% when using Log Mel spectrum as features.
• The model demonstrated high accuracy with a sensitivity of 89.5% and specificity of 89.7%.
• The weighted loss functions effectively addressed the class imbalance problem.

Conclusions:

• A CNN-based model for automated heart sound classification has been successfully developed.
• The model shows significant potential for assisting in heart auscultation and screening for heart pathologies.
• This approach offers a cost-effective tool for clinical applications in early heart disease detection.