Diagnostic-Wise Hybrid Loss for ECG Signals Reconstruction Based on Wavelet Decomposition

Electrocardiogram (ECG) denoising requires preserving key diagnostic features while removing noise in both time and frequency domains. Current loss functions struggle to balance global reconstruction with local feature preservation, often causing oversmoothing or frequency distortions. To address these limitations, we propose DiagHuber, an innovative domain knowledge-integrated loss function that uniquely combines multiscale diagnostic localization with adaptive error tolerance. The diagnostic module implements discrete wavelet decomposition with adaptive coefficient weighting to enhance clinically critical ECG features while attenuating multiscale noise components. Simultaneously, the Huber loss component adaptively adjusts error sensitivity to balance outlier rejection with waveform fidelity. For practical implementation, we develop Mamba-UNet (MAUnet), an attention-enhanced U-Net incorporating Mamba blocks based on state-space models, which achieves real-time processing capabilities (<0.5 s per 30-min ECG recording). Extensive experimental evaluations on datasets contaminated with both synthetic and real-world noise demonstrate DiagHuber’s superior performance, achieving a mean signal-to-noise ratio (SNR) of 16.17 dB on test data and a 15.67% improvement over conventional mean squared error (mse) loss. Cross-architecture validation, time-frequency analysis, and 12-lead generalization test further confirm the proposed method’s robustness to simultaneously suppress broadband noise while meticulously preserving morphological features.