[Recognition of S1 and S2 heart sounds with two-stream convolutional neural networks]

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Auscultation of heart sounds is crucial for diagnosing cardiac conditions.
• Precise segmentation of heart sounds, specifically identifying the first (S1) and second (S2) heart sounds, is essential for accurate diagnosis.
• Existing methods may rely on cardiac cycle duration, which this study aims to bypass.

Purpose of the Study:

• To develop and evaluate a method for classifying S1 and S2 heart sounds based on their spectral properties.
• To improve the accuracy and efficiency of heart sound segmentation for clinical diagnosis.
• To avoid manual feature extraction in heart sound analysis.

Main Methods:

• Heart sounds (S1 and S2) were transformed into spectra using short-time Fourier transform.
• A two-stream convolutional neural network was employed for classification.
• The method focused on spectral properties, excluding cardiac cycle duration.

Main Results:

• The classification accuracy for S1 and S2 sounds reached 91.135% on the test dataset.
• Achieved a maximum sensitivity of 91.156% and specificity of 92.074%.
• The proposed method effectively avoids artificial feature extraction.

Conclusions:

• The developed method accurately distinguishes S1 and S2 heart sounds using spectral analysis and a convolutional neural network.
• This approach offers a computationally efficient and effective solution for real-time heart sound analysis.
• The findings support the potential for improved automated cardiac diagnosis through precise heart sound segmentation.