Deep learning models for segmenting phonocardiogram signals: a comparative study.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiac auscultation relies on analyzing mechanical vibrations (phonocardiogram - PCG) from the heart.
• PCG signals contain various frequencies generated by cardiac structure movements during blood circulation.
• Accurate segmentation of PCG signals is crucial for diagnosing cardiac conditions.

Purpose of the Study:

• To apply and compare deep learning models for segmenting specific regions within PCG signals.
• To evaluate the accuracy of Gated Recurrent Neural Network (GRU), Bidirectional-GRU, and Bi-directional Long-Term Memory (BILSTM) models.
• To assess the performance of these models on diverse PCG datasets.

Main Methods:

• PCG signals were pre-processed using digital filtering and empirical mode decomposition.
• Deep learning models (GRU, Bidirectional-GRU, BILSTM) were independently applied for segmentation.
• Segmentation focused on four key regions: S1, systolic, S2, and diastolic periods.
• Models were tested on PhysioNet, MIT-HSDB, and CirCor DigiScope Phonocardiogram datasets.

Main Results:

• The proposed approach achieved high segmentation accuracy: 97.2% on PhysioNet and 96.98% on MIT-HSDB.
• This study marks the first investigation into CirCor DigiScope dataset segmentation, achieving 92.5% accuracy.
• Comparative analysis demonstrated the efficiency and reliability of the deep learning models.

Conclusions:

• Deep learning models, including GRU and BILSTM variants, show significant promise for accurate PCG signal segmentation.
• The developed approach offers a reliable software tool for enhancing cardiac auscultation analysis in clinical settings.
• Further research can explore these models for automated cardiac diagnosis and monitoring.