Feature extraction for murmur detection based on support vector regression of time-frequency representations.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Cardiac mechanical activity analysis relies on phonocardiographic (PCG) recordings.
• Understanding the underlying dynamics of PCG signals is crucial for diagnosing cardiac conditions.
• Existing time-frequency representation (TFR) analysis methods may not fully capture the nonlinear dynamics of PCG signals.

Purpose of the Study:

• To develop and validate a nonlinear approach for PCG signal analysis using time-frequency representations (TFRs).
• To model TFRs of cardiac sounds using Support Vector Regression (SVR).
• To classify normal and pathologic (murmur) PCG recordings based on extracted TFR features.

Main Methods:

• Utilized Support Vector Regression (SVR), a nonlinear statistical learning method, for modeling TFRs.
• Calculated dissimilarity measures between regressions using dot product.
• Employed a k-nearest neighbors (k-nn) classifier for the final classification task.

Main Results:

• The proposed nonlinear SVR-based methodology effectively models TFRs of PCG signals.
• Feature extraction from modeled TFRs enabled discrimination between normal and pathologic heart sounds.
• Achieved a high validation performance of 97.85% in classifying PCG recordings.

Conclusions:

• The nonlinear SVR approach offers a robust method for analyzing cardiac TFRs.
• This methodology shows significant potential for accurate automated diagnosis of cardiac conditions from PCG signals.
• The high classification accuracy validates the effectiveness of the proposed TFR modeling and feature extraction technique.