Complex audio signal data compression and reconstruction: A benchmark data pre-processing approach for machine...

Area of Science:

• Medical signal processing
• Biomedical engineering
• Digital signal processing

Background:

• Medical auscultation generates complex audio signals crucial for diagnosis.
• Efficiently compressing these signals is vital for machine learning classification and data management.
• Existing methods may struggle to preserve diagnostic integrity during compression.

Purpose of the Study:

• To develop and assess novel methods for compressing and reconstructing medical auscultation audio signals.
• To maintain diagnostic accuracy while reducing signal dimensionality for machine classification.
• To evaluate the effectiveness of various signal processing techniques for this purpose.

Main Methods:

• Utilized the ICBHI Respiratory Challenge 2017 Database for analysis.
• Evaluated compression frameworks including Discrete Fourier Transform, time-frequency transforms, dictionary learning, and Singular Value Decomposition.
• Assessed reconstruction quality using Mean Squared Error (MSE).

Main Results:

• The multi-resolution wavelet transform (MRWT) framework achieved the lowest average MSE of 0.037.
• The proposed time-frequency framework incorporating MRWT demonstrated 80% accuracy in differentiating chronic obstructive pulmonary disease from healthy respiratory sounds.
• MRWT proved superior among the assessed compression methods.

Conclusions:

• The study advances signal processing techniques for medical auscultation.
• The MRWT approach offers a promising balance between compression efficiency and reconstruction accuracy.
• These findings provide insights into preserving diagnostic information in compressed audio signals.