Murmur identification and outcome prediction in phonocardiograms using deep features based on Stockwell transform

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiac auscultation for heart murmurs requires specialized expertise.
• Phonocardiograms (PCGs) offer an alternative data source for murmur detection.
• Developing automated methods for PCG analysis is crucial for broader clinical application.

Purpose of the Study:

• To predict clinical outcomes (normal/abnormal) and detect heart murmurs using PCGs.
• To develop and evaluate a semi-supervised model for enhanced PCG classification.
• To leverage time-frequency deep features for improved diagnostic performance.

Main Methods:

• Investigated PCG signals in the time-frequency domain using the Stockwell transform to create time-frequency maps (TFMs).
• Employed AlexNet to extract deep features from TFMs, followed by feature reduction.
• Evaluated feature effectiveness using three classifiers on the CinC/Physionet challenge 2022 dataset.

Main Results:

• Achieved 93% accuracy, 91% sensitivity, and 91% F1-score for heart murmur detection (Task I).
• Demonstrated a clinical outcome cost of 5290 for predicting patient outcomes (Task II), outperforming benchmark methods.
• The proposed semi-supervised model effectively utilizes deep time-frequency features for PCG analysis.

Conclusions:

• The developed semi-supervised model shows significant promise for automated heart murmur detection and clinical outcome prediction.
• Time-frequency deep features derived from PCGs are effective for improving diagnostic accuracy.
• This approach offers a potential advancement over traditional cardiac auscultation methods.