A study of time-frequency features for CNN-based automatic heart sound classification for pathology detection

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Automatic detection of structural heart abnormalities is crucial for pediatric screening.
• Convolutional Neural Networks (CNNs) show promise using time-frequency representations of phonocardiogram (PCG) signals.
• Current CNN approaches often rely on standard hand-crafted features.

Purpose of the Study:

• To evaluate alternative time-frequency representations for CNN-based heart abnormality detection.
• To compare commonly used features (MFCC, Mel-Spectrogram) with domain-knowledge-influenced sub-band envelopes.
• To assess the impact of windowing techniques (synchronous vs. asynchronous) on detection performance.

Main Methods:

• Utilized two high-quality PCG databases for testing.
• Implemented CNN models with different time-frequency representations: Mel-Frequency Cepstral Coefficients (MFCC), Mel-Spectrogram, and sub-band envelopes.
• Compared period synchronous and asynchronous windowing methods for signal segmentation.

Main Results:

• Sub-band envelopes demonstrated superior performance compared to MFCC and Mel-Spectrogram features.
• Period synchronous windowing yielded better results than asynchronous windowing.
• The proposed approach enhances the accuracy of automatic structural heart abnormality risk detection.

Conclusions:

• Sub-band envelopes represent a more effective feature for CNN-based PCG analysis in pediatric heart disease screening.
• Period synchronous windowing is recommended for improved performance in PCG signal processing.
• This study offers advancements for non-invasive pediatric cardiac screening tools.