Supervised threshold-based heart sound classification algorithm.

Area of Science:

• Cardiology
• Machine Learning
• Signal Processing

Background:

• Deep classification networks are vital for heart sound analysis.
• Data augmentation via segmentation is common but poses challenges in instance categorization.
• Accurate classification of segmented heart sounds is crucial for reliable diagnostics.

Purpose of the Study:

• To develop a mathematical framework for determining heart sound instance categories from segment predictions.
• To enhance the performance of deep classification networks in heart sound analysis.
• To improve data augmentation strategies for training heart sound classifiers.

Main Methods:

• A mathematical formula was established to link instance classification performance with segment prediction results using a supervised threshold.
• The optimal threshold was determined by maximizing training instance prediction accuracy via a gradient-based method.
• A continuous function approximation was used to transform the discrete accuracy function, facilitating threshold optimization.

Main Results:

• The proposed algorithm improved mean accuracy (MAcc) by approximately 4% compared to the baseline.
• The method demonstrated superior performance, surpassing the champion results of the PhysioNet/Computing in Cardiology Challenge 2016.
• Cross-validation results showed significant improvements in sensitivity, specificity, and MAcc.

Conclusions:

• A robust methodology was developed for categorizing predicted heart sound instances from segment data.
• The approach effectively addresses challenges in data augmentation for deep learning models in cardiology.
• The study significantly advances the classification performance of automated heart sound analysis systems.