Classification of real-world pathological phonocardiograms through multi-instance learning.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Heart auscultation is a fundamental diagnostic technique for cardiovascular disease.
• High human error rates and limited access to expertise hinder accurate auscultation.
• Developing automated tools for heart sound analysis is crucial for improving diagnosis.

Purpose of the Study:

• To develop an accurate computational strategy for diagnosing heart sounds.
• To address the challenges posed by data heterogeneity and low signal-to-noise ratios.
• To create a reliable tool for detecting heart sound abnormalities.

Main Methods:

• Automatic heart phase segmentation.
• Mel-frequency cepstral representation filters from speech synthesis.
• A multi-branch, multi-instance artificial neural network with convolutional layers.

Main Results:

• The developed artificial intelligence architecture achieved high performance.
• Demonstrated an Area Under the Curve (AUC) of 0.87.
• Achieved a sensitivity of 0.97 for heart sound classification.

Conclusions:

• The machine learning tool shows significant potential for accurate heart sound classification.
• This tool can serve as an effective screening method in various clinical settings.
• The approach is particularly promising for telemedicine applications.