Novel features from autocorrelation and spectrum to classify Phonocardiogram quality

Area of Science:

• Biomedical Signal Processing
• Cardiology
• Medical Informatics

Background:

• Phonocardiogram (PCG) auscultation is crucial for preliminary medical screening.
• PCG recordings in uncontrolled environments are often noisy, hindering accurate analysis.
• Ensuring high-quality PCG data is essential for reliable diagnostic interpretation.

Purpose of the Study:

• To develop novel features for identifying and accepting diagnosable quality PCG recordings.
• To improve the robustness of PCG quality assessment across different devices and protocols.
• To enhance the reliability of automated analysis of heart sounds.

Main Methods:

• Derivation of novel features from the spectral domain of PCG signals.
• Utilizing autocorrelation waveforms for signal analysis.
• Developing training models using a large, diverse, medical-grade PCG dataset.
• Validation on an in-house dataset collected with a smartphone-based digital stethoscope.

Main Results:

• The proposed features effectively identify PCG recording quality.
• The methodology demonstrates robustness against variations in devices and data collection.
• An accuracy score of approximately 75% was achieved on the in-house dataset.
• The approach successfully filters out non-diagnosable PCG recordings.

Conclusions:

• Novel spectral and autocorrelation features enhance PCG quality assessment.
• The developed method ensures consistent data quality for further analysis.
• This approach supports reliable preliminary medical screening using PCG.
• The findings are applicable to low-cost digital stethoscopes and various recording conditions.