Comparative Study of ECG Denoising Methods for Wearable Applications

Abstract

Reliable electrocardiogram (ECG) monitoring in wearable and space environments requires effective denoising of signals corrupted by non-stationary electromyogram (EMG) interference. This paper presents a comparative evaluation of model-based and DL-based denoising techniques for upper-arm ECG recordings acquired under real conditions. The model-based methods include three empirical mode decomposition (EMD) variants and a discrete wavelet transform (DWT) approach, while the deep learning (DL) side is represented by a stacked denoising autoencoder (SDAE) and a physics-informed neural network (PINN). All methods are evaluated on real acquisitions under both relaxed and voluntary muscle contraction conditions, using root mean squared error (RMSE), Pearson correlation, and peak-to-peak signal-to-noise ratio (PPSNR) as performance metrics. Results reveal a fundamental trade-off: DL methods achieve superior morphological reconstruction, while DWT provides the strongest noise suppression, highlighting complementary strengths for wearable cardiac monitoring applications.