Multitask learning-based phonocardiogram denoising model for preserving valvular heart disease characteristics

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Phonocardiogram (PCG) signals are crucial for diagnosing valvular heart disease (VHD).
• Clinical PCG recordings are often contaminated by various noise sources, hindering accurate diagnosis.
• Existing denoising methods may struggle to preserve diagnostically relevant heart murmur features.

Purpose of the Study:

• To develop a multitask learning (MTL) based denoising model for reconstructing phonocardiogram (PCG) signals.
• To preserve heart murmur characteristics in PCG signals under diverse synthetic and real-world noise conditions.
• To jointly denoise PCG signals and classify valvular heart disease (VHD).

Main Methods:

• A multitask learning (MTL) model was developed to jointly denoise PCG data and classify VHD.
• Synthesized noisy PCG signals by mixing clean PCGs with synthetic and real-world noise sources.
• Spectrogram images were used as input for the model.
• Evaluated performance using signal-to-noise ratios (SNRs) from -5 to 10 dB with 5x5 nested cross-validation.
• Denoising performance measured by scale-invariant signal-to-distortion ratio (SI-SDR); VHD classification by accuracy.

Main Results:

• The MTL model achieved a mean SI-SDR of 14.05 dB at -5 dB SNR, outperforming a single-task U-Net baseline.
• Improved SI-SDR by up to 9.69 dB for real-world noise (hospital ambient, lung sounds) compared to previous methods.
• Achieved over 98.5% accuracy for VHD classification at 5 and 10 dB SNRs, indicating preserved murmur features.
• Demonstrated robust PCG reconstruction and improved VHD classification on noisy signals.

Conclusions:

• The developed MTL-based model provides consistent and robust PCG signal reconstruction across various noise conditions.
• The model effectively preserves critical heart murmur characteristics necessary for accurate VHD diagnosis.
• This approach enhances the reliability of PCG analysis in noisy clinical environments.