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

This study developed a multitask learning (MTL)-based denoising model to reconstruct phonocardiogram (PCG) signals while preserving heart murmur characteristics under both synthetic and diverse real-world clinical noise. The model was trained to jointly denoise PCG data and classify valvular heart disease (VHD). The original PCG dataset comprised recordings lasting up to 3 s, including normal heart sounds and four VHD classes, with 200 samples per class. To construct inputs, we synthesized contaminated PCGs by mixing synthetic and real-world clinical noise sources and converted them into spectrogram images. The model was evaluated at signal-to-noise ratios (SNRs) of -5, 0, 5, and 10 dB using 5 × 5 nested cross-validation. Denoising performance was measured by the scale-invariant signal-to-distortion ratio (SI-SDR), and VHD classification by accuracy. Our model outperformed a single-task U-Net denoising baseline; even at a highly contaminated SNR of -5 dB, it achieved an absolute mean SI-SDR of 14.05 dB. For real-world noise, including hospital ambient noise and lung sounds, our model improved SI-SDR by up to 9.69 dB over previous work. For VHD classification, the model exceeded 98.5% accuracy across most conditions at SNRs of 5 and 10 dB, suggesting that heart murmur-specific features were preserved during denoising. Overall, this MTL-based model demonstrated consistent and robust PCG reconstruction across diverse noise conditions, improving VHD classification on contaminated PCG signals.