Heart sound classification based on improved MFCC features and convolutional recurrent neural networks.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Heart sound classification is crucial for early cardiovascular disorder detection, particularly in primary care settings.
• Existing methods often rely on conventional features and shallow classifiers, which struggle with complex cardiac acoustic environments.
• These limitations can lead to degraded performance in heart sound analysis.

Purpose of the Study:

• To propose a novel heart sound classification method using improved Mel-frequency cepstrum coefficient (MFCC) features and convolutional recurrent neural networks (CRNNs).
• To enhance the characterization of dynamic features in heart sound signals.
• To improve the accuracy and robustness of cardiovascular disorder detection.

Main Methods:

• Calculated Mel-frequency cepstrum coefficients (MFCCs) without segmenting the heart sound signal.
• Developed an improved MFCC feature extraction scheme to capture dynamic characteristics of consecutive heart sounds.
• Utilized a deep convolutional recurrent neural network (CRNN) for feature learning and classification.

Main Results:

• Achieved a classification accuracy of 98% for the two-class problem (pathological vs. non-pathological).
• The proposed deep learning framework effectively combines local feature extraction (CNN) and long-term dependency modeling (RNN).
• Demonstrated superior performance compared to state-of-the-art algorithms on the 2016 PhysioNet/CinC Challenge database.

Conclusions:

• The proposed MFCC-based CRNN method offers a robust and accurate approach for heart sound classification.
• This method shows significant potential for improving early detection of cardiovascular disorders in clinical practice.
• The integration of advanced deep learning techniques enhances the analysis of complex cardiac acoustic signals.