Dual-FreqDAE: Frequency-aware dual-branch transformer autoencoder for ECG denoising in mixed noise conditions

Electrocardiogram (ECG) denoising is essential for reliable cardiac analysis, especially in real-world settings where signals are often corrupted by multiple types of noise simultaneously. However, existing approaches often fail to restore high-quality signals under such complex noise conditions. In this paper, we propose Dual-FreqDAE , a deep learning-based denoising model that integrates both time-domain and frequency-domain representations through a dual-branch architecture. The combined features are processed by a Transformer encoder with Fourier Analysis Networks, enabling the model to capture multi-scale dependencies while preserving morphological characteristics of ECG signals. Extensive experiments demonstrate that Dual-FreqDAE consistently outperforms traditional filtering methods and state-of-the-art denoising autoencoders. Importantly, the explicit incorporation of frequency-domain features allows the model to recover signal components across a wide frequency range more effectively than conventional deep learning methods. Beyond waveform restoration, our model also achieves high accuracy in downstream tasks such as R-peak detection and cardiac abnormality classification, underscoring its potential practical applicability.