Automatic heart sound classification from segmented/unsegmented phonocardiogram signals using time and frequency...

Area of Science:

• Cardiology
• Biomedical Signal Processing
• Machine Learning

Background:

• Heart abnormality detection using phonocardiogram (PCG) signals is a critical area in cardiovascular research.
• Automatic classification of PCG signals aims to improve diagnostic accuracy and efficiency.

Purpose of the Study:

• To investigate the effectiveness of various time and frequency domain features for automatic heart sound classification.
• To compare the performance of different classification algorithms on both segmented and unsegmented PCG signals.
• To identify optimal feature subsets and classification methods for enhanced heart abnormality detection.

Main Methods:

• In-depth analysis of time and frequency domain features for PCG signals.
• Experimental determination of effective feature subsets for classification.
• Evaluation of classifiers including SVM, kNN, Decision Tree, Ensemble, ANN, and LSTM on segmented and unsegmented PCG data.
• Comparison of proposed feature subsets with established methods.

Main Results:

• Long Short-Term Memory (LSTM) networks achieved 91.39% AUC with Mel-Frequency Cepstral Coefficients (MFCCs) on unsegmented data, but showed inconsistency with other classifiers.
• Time-frequency features from segmented PCG data yielded AUC scores over 70% across all classifiers.
• Conventional machine learning techniques demonstrated consistent performance with AUC scores exceeding 80% on segmented data.

Conclusions:

• Time and frequency domain features are crucial for accurate heart sound classification.
• Segmented PCG signals combined with appropriate feature extraction yield superior and consistent classification performance.
• Conventional machine learning techniques offer reliable results for heart abnormality detection using segmented PCG signals.