Robust Heart Sound Analysis With MFCC and Light Weight Convolutional Neural Network

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Traditional heart sound analysis relies on manual feature engineering and clinical expertise.
• Automated methods are needed for efficient and accurate cardiovascular disorder classification.
• Existing techniques may lack generalizability and robustness.

Purpose of the Study:

• To develop and evaluate a Convolutional Neural Network (CNN)-based model for automated multiclass heart sound classification.
• To assess the model's performance using Mel-frequency cepstral coefficients (MFCCs) from real-world heart sound recordings.
• To establish a robust framework for automated auscultation and early cardiac screening.

Main Methods:

• Segmentation of real-world heart sound recordings.
• Extraction of Mel-frequency cepstral coefficients (MFCCs) as features.
• Implementation of a CNN model for multiclass classification of heart sounds (murmur, extrasystole, extrahls, artifact, normal).

Main Results:

• The MFCC-CNN model achieved 98.7% training accuracy and 91% validation accuracy.
• High precision and recall were observed for the normal and murmur heart sound classes.
• A weighted F1-score of 0.91 indicates strong overall classification performance.

Conclusions:

• The proposed MFCC-CNN framework demonstrates robustness and generalizability for heart sound analysis.
• The model is suitable for automated auscultation, offering potential for early cardiac screening.
• This approach reduces reliance on manual feature engineering and clinical expertise.