Abnormal heart sound recognition using SVM and LSTM models in real-time mode.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiovascular diseases are the leading global cause of death.
• Heart sound auscultation is crucial for diagnosis, but subjective.
• Automatic digital auscultation systems offer objective analysis.

Purpose of the Study:

• To develop a real-time heart sound recognition system.
• To classify phonocardiograms as normal or abnormal.
• To identify specific types of heart abnormalities.

Main Methods:

• Digital signal processing techniques including Fast Fourier Transform, filtering, and Dual-Tree Complex Wavelet Transform.
• Machine learning algorithms: Support Vector Machine (SVM) for classification and Long-Short Term Memory (LSTM) neural networks for abnormality type recognition.
• Utilized three datasets: PhysioNet (1395 files), GitHub (800 files), and PASCAL (100 files), segmented into cardiac cycles.

Main Results:

• Normal/abnormal classification achieved high accuracy (up to 98.1%) with SVM.
• LSTM models demonstrated excellent accuracy (up to 99.5%) in classifying multiple heart sound abnormalities.
• An efficient automatic segmentation method achieved low error rates and rapid computational time (0.04s per cycle).

Conclusions:

• The proposed system effectively classifies normal and abnormal heart sounds in real-time.
• The integration of signal processing and machine learning enhances diagnostic accuracy for cardiovascular diseases.
• This technology has the potential to reduce mortality rates and improve patient quality of life.