Deep learning based classification of unsegmented phonocardiogram spectrograms leveraging transfer learning

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiovascular diseases (CVDs) are a leading global cause of mortality.
• Early detection of heart abnormalities is crucial for effective CVD management.
• Phonocardiogram (PCG) signals offer valuable diagnostic information but present analysis challenges due to data variability.

Purpose of the Study:

• To develop and evaluate a deep learning-based computer-aided system for analyzing PCG signals.
• To improve the accuracy and timeliness of detecting heart abnormalities.
• To address the challenges posed by heterogeneous PCG datasets (PhysioNet and PASCAL).

Main Methods:

• Utilized short-time Fourier transform (STFT) to generate spectrograms from PCG signals.
• Developed and tested various convolutional neural network (CNN) models on PhysioNet, PASCAL, and combined datasets.
• Applied transfer learning techniques to enhance model performance on the PASCAL dataset.

Main Results:

• The CNN model achieved high performance on the PhysioNet dataset (e.g., 95.75% accuracy).
• Performance varied across datasets, with the PASCAL dataset yielding 75.25% accuracy.
• A combined dataset approach reached 92.7% accuracy, and transfer learning improved precision to 96.98% on noisy data.

Conclusions:

• A custom, lightweight CNN model effectively analyzes PCG spectrograms for CVD detection.
• The proposed deep learning approach demonstrates high classification accuracy and precision.
• This method shows potential for efficient and reliable screening of cardiovascular diseases using PCG signals.