Transfer Learning Models for Detecting Six Categories of Phonocardiogram Recordings

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Auscultation is a fundamental yet variable method for cardiovascular disease detection.
• Misdiagnoses can occur even with experienced physicians, necessitating accurate computational tools.
• Artificial intelligence (AI) offers potential for efficient and accurate cardiovascular disease diagnosis.

Purpose of the Study:

• To propose an automatic multiple classification method for cardiovascular disease detection using heart sound signals.
• To develop an AI tool to assist in cardiovascular auscultation, particularly for developing countries.
• To evaluate the robustness of the method in noisy environments.

Main Methods:

• Heart sound signals were converted into 3D spectrograms using continuous wavelet transform (CWT).
• Six classes of heart sound data were used, combining an open database with self-collected pulmonary hypertension data.
• Ten transfer learning networks and other models (LSTM, CNN) were compared after background deformation for noise robustness.

Main Results:

• Four transfer learning networks (ResNet101, DenseNet201, DarkNet19, GoogleNet) achieved high accuracy (0.98) in detecting multiple heart diseases.
• The proposed method demonstrated robustness in noisy environments.
• Performance was validated using 10-fold cross-validation on both original and augmented heart sound data.

Conclusions:

• The developed AI classification method achieved high accuracy, even with noisy heart sound signals.
• This suggests the method's potential as a valuable auxiliary diagnostic tool for cardiovascular diseases.
• The AI approach can enhance the reliability of cardiovascular disease detection.