Phonocardiogram classification using deep neural networks and weighted probability comparisons.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiac auscultation is a conventional but user-dependent method for heart disease assessment.
• Existing computational phonocardiogram (PCG) analysis methods often lack sufficient sensitivity and specificity or require specialized hardware.
• The need for objective and repeatable heart sound analysis is critical for accurate diagnosis.

Purpose of the Study:

• To develop and validate a novel computational approach for classifying normal and abnormal heart sounds using phonocardiograms (PCG).
• To improve the sensitivity and specificity of automated heart sound classification compared to existing methods.
• To create a system suitable for clinical applications by addressing limitations of previous computational techniques.

Main Methods:

• Utilized deep neural networks (DNNs) to compute individual cardiac cycle probabilities from PCG data.
• Implemented a classification strategy based on weighted probability comparisons.
• Trained and tested the system on an extensive, balanced dataset of 18,179 normal and abnormal cardiac cycles.

Main Results:

• Achieved high diagnostic performance with a sensitivity (Se) of 91.3% and specificity (Sp) of 93.8%.
• The system demonstrated robustness due to training on a balanced dataset.
• A controllable decision factor allows for managing the trade-off between sensitivity and specificity.

Conclusions:

• The proposed deep learning approach offers a significant advancement in automated heart sound classification from PCG.
• The system's high accuracy and controllable performance metrics make it a promising tool for clinical heart disease assessment.
• This method overcomes limitations of previous approaches, paving the way for more reliable and accessible cardiac diagnostics.