Compressed Sensing of Acoustic Cardiopulmonary Signals Using a CNN-based Reconstruction Method

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Cardiopulmonary sounds are vital for diagnosing respiratory and cardiovascular conditions.
• Traditional compressive sensing methods face challenges with the complexity and variability of cardiopulmonary sounds.
• Efficient data compression is crucial for real-time monitoring and telemedicine applications.

Purpose of the Study:

• To develop a novel approach for compressive sensing and reconstruction of cardiopulmonary sounds.
• To overcome limitations of traditional compressive sensing by using a CNN-U-Net architecture.
• To enable efficient data compression for low-cost edge devices in healthcare.

Main Methods:

• A Convolutional Neural Network (CNN) based on the U-Net architecture was trained for signal reconstruction.
• The CNN was trained on pseudo-randomly undersampled respiratory sounds (SPRSound dataset) and Phonocardiogram (PCG) signals (CirCor Digiscope PCG dataset).
• The method bypasses explicit sparsity enforcement, training the network directly on undersampled data.

Main Results:

• The proposed method achieved a compression ratio of up to 30 for cardiopulmonary sounds.
• Reconstruction quality was comparable to previous methods, but with significantly higher compression ratios (three times higher for respiratory sounds).
• The algorithm demonstrated high signal integrity after reconstruction for both respiratory and PCG signals.

Conclusions:

• The U-Net CNN approach offers an effective solution for compressing cardiopulmonary sounds, overcoming traditional limitations.
• This method enables efficient, low-power data compression suitable for implementation on edge devices.
• The technology supports enhanced real-time monitoring, telemedicine, and point-of-care diagnostics for cardiopulmonary conditions.