Heart sound classification based on improved mel-frequency spectral coefficients and deep residual learning

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Accurate heart sound classification is vital for early cardiovascular disease diagnosis.
• Existing methods often rely on conventional features and shallow classifiers, limiting performance.
• There is a need for more robust and accurate heart sound analysis techniques.

Purpose of the Study:

• To propose a novel heart sound classification method.
• To leverage improved mel-frequency cepstral coefficient (MFCC) features and deep residual learning.
• To enhance the accuracy and effectiveness of cardiovascular disease detection through heart sound analysis.

Main Methods:

• Preprocessing of heart sound signals.
• Computation of improved mel-frequency cepstral coefficient (MFCC) features.
• Application of deep residual learning networks for feature extraction and classification.
• Analysis of network parameters and connection strategies.

Main Results:

• The proposed method achieved a high accuracy of 94.43% on the evaluated dataset.
• Deep residual learning effectively extracted pathological information from heart sound signals.
• Improved MFCC features enhanced the discriminative power for classification.

Conclusions:

• The novel approach combining improved MFCC features and deep residual learning shows significant promise for accurate heart sound classification.
• This method offers a potential advancement in the early diagnosis of cardiovascular diseases.
• Further research into network architectures and feature optimization could yield even greater diagnostic capabilities.