Time-Frequency Fragment Selection for Disease Detection from Imbalanced Phonocardiogram Data.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Cardiovascular disease (CVD) is a major global health concern.
• Phonocardiogram (PCG) analysis is a non-invasive method for CVD detection.
• Machine learning (ML) methods for PCG analysis require large, balanced datasets, which are often unavailable.

Purpose of the Study:

• To address data scarcity and imbalance in Phonocardiogram (PCG) datasets for cardiovascular disease (CVD) detection.
• To improve the performance of machine learning models in classifying heart sounds.
• To provide reliable assistance for heart auscultation and screening of heart pathologies.

Main Methods:

• Proposed a variable-hop fragment selection method combined with a pre-trained Convolutional Neural Network (CNN) model.
• Developed a framework to overcome limitations of scarce and imbalanced PCG data.
• Evaluated the method on the PhysioNet/CinC Challenge 2016 dataset.

Main Results:

• Achieved a 7.12% improvement in unweighted average recall (UAR) compared to state-of-the-art methods.
• Reported an overall UAR of 92.46% on the imbalanced PhysioNet/CinC Challenge 2016 dataset.
• Demonstrated superior performance in assessing imbalanced PCG datasets.

Conclusions:

• The proposed framework effectively mitigates issues of data scarcity and imbalance in PCG analysis.
• The method offers reliable assistance for heart auscultation and potential for screening heart pathologies.
• The improved performance highlights clinical relevance for data-constrained applications.