Noise-Robust Heart Sound Analysis for Wearable Monitoring Using Cascaded Dual-Channel ANC and Wavelet Denoising

Phonocardiogram (PCG) signals are widely used for automated cardiovascular abnormality screening and wearable health monitoring. However, during real-world acquisition, PCG signals are often corrupted by environmental noise, which masks critical transient structures such as S1 and S2 and degrades automatic analysis and classification performance. To address the challenge of jointly achieving effective noise suppression and transient structure fidelity under complex noise conditions, this paper proposes a cascaded PCG frontend framework that integrates dual-channel subband adaptive noise cancellation with subband-selective wavelet denoising. The framework first suppresses environmental noise highly correlated with the reference channel at the subband level, and then applies selective wavelet-domain thresholding to further attenuate residual noise while preserving heart sound transients.A self-trained lightweight heart sound classification model is used for evaluation. Under low-SNR conditions, the proposed cascaded approach achieves classification performance gains of 9.3%-12.7% over noisy inputs, significantly outperforming single-stage wavelet denoising (6.3%-9.1%) and ANC (4.4%-5.0%). Stable positive gains are also maintained at mediumto-high SNRs, whereas single-stage wavelet denoising exhibits performance degradation. These results demonstrate that the proposed front-end jointly suppresses noise and preserves key transient structures at the signal level, relieving low-memory classifiers from learning additional noise robustness and enabling efficient collaboration for embedded and wearable heart sound analysis systems.