Improved Segmentation with Dynamic Threshold Adjustment for Phonocardiography Recordings

Area of Science:

• Biomedical Signal Processing
• Cardiology
• Machine Learning in Healthcare

Background:

• Phonocardiography (PCG) recordings are crucial for diagnosing heart conditions.
• Accurate segmentation of heart sounds (S1, S2) is essential for reliable PCG analysis.
• Existing methods may face challenges in segmenting complex or noisy heart sound signals.

Purpose of the Study:

• To develop and evaluate a novel algorithm for segmenting phonocardiography (PCG) recordings.
• To improve the accuracy of heart sound detection, including S1 and S2 sounds.
• To assess the algorithm's performance on both normal and abnormal heart sound datasets.

Main Methods:

• A feature vector is generated using joint wavelet transform and mel scaled energy spectrum of PCG signals.
• A peak detection algorithm identifies key features within the processed signal.
• Heart sounds are labeled via circular convolution with a template, guided by detected peaks.
• An error detection and correction stage refines the labeling process.

Main Results:

• The algorithm achieved 99.51% recall and 97.28% precision for detecting S1 and S2 sounds in normal heart sounds.
• For abnormal heart sounds, the algorithm demonstrated 97.59% recall and 92.53% precision.
• Performance was validated on a dataset of 80 total PCG records.

Conclusions:

• The proposed algorithm offers a robust and accurate method for PCG signal segmentation.
• The joint use of wavelet transform and mel scaled energy spectrum enhances feature representation for heart sound analysis.
• The algorithm shows high efficacy in detecting and labeling both normal and abnormal heart sounds.