Recurrent vs Non-Recurrent Convolutional Neural Networks for Heart Sound Classification

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are prevalent in heart sound analysis.
• Recurrent Neural Networks (RNNs) like Gated Recurrent Networks (GRNs) and Long Short-Term Memory (LSTM) are often combined with CNNs for complex tasks.
• Evaluating simpler models against complex hybrid architectures is crucial for efficient diagnostic tools.

Purpose of the Study:

• To compare the performance of a conventional CNN against various CNN-RNN hybrid architectures for abnormal-normal heart sound classification.
• To assess the accuracy and sensitivity of different parallel and cascaded integration methods.
• To determine the optimal model complexity for reliable heart sound signal classification.

Main Methods:

• Utilized the Physionet dataset for heart sound recordings.
• Implemented and evaluated a conventional CNN architecture.
• Developed and tested parallel and cascaded integration architectures of CNN with GRN and LSTM.
• Measured accuracy and sensitivity for each model configuration.

Main Results:

• The parallel LSTM-CNN architecture achieved the highest accuracy (98.0%) with 87.2% sensitivity.
• A conventional CNN demonstrated strong performance with 97.3% accuracy and 95.9% sensitivity.
• The conventional CNN required significantly less computational complexity compared to hybrid models.

Conclusions:

• Conventional CNNs are suitable for abnormal-normal heart sound classification.
• Simpler CNN models offer a balance of high performance and reduced complexity.
• The findings suggest that conventional CNNs can be independently utilized for effective heart sound analysis.