ECG-ARD: Autoregressive ECG Denoising With Dual-Domain Fusion and a Vector-Quantized Bottleneck

Electrocardiogram (ECG) recordings are frequently corrupted by nonstationary artifacts such as baseline wander, muscle activity, and electrode motion, which can distort waveform morphology, reduce signal quality, and compromise downstream analysis. Although recent deep learning approaches have improved signal-level fidelity, practical gaps remain in cross-dataset robustness under realistic noise mixtures, leakage-aware evaluation protocols, and computational efficiency for long-duration monitoring. We propose the ECG Autoregressive Denoiser (ECG-ARD), a patch-level conditional autoregressive framework that reconstructs clean ECG signals sequentially while conditioning on the fully observed noisy segment. ECG-ARD uses a hierarchical bidirectional hybrid Mamba-CNN encoder to aggregate long-range context, a multi-scale fusion module with channel, time-domain, and magnitude-only frequency-domain attention to selectively integrate cross-scale features into the conditioning representation, and a residual vector-quantized bottleneck to regularize the conditioning latent while preserving a continuous information pathway. A causal hybrid Mamba-CNN decoder then generates clean patches autoregressively from the clean-patch history. We evaluate ECG-ARD under a leakage-aware, deployment-oriented protocol that includes record-wise splits, NSTDB channel separation to reduce reuse of noise realizations between model development and testing, and noisy-only normalization to avoid oracle statistics. Across in-distribution benchmarks (MITDB and QTDB) and out-of-distribution testing on SimEMG, ECG-ARD reduces distortion metrics while maintaining high waveform similarity, with moderate inference latency per 10-second segment relative to diffusion-based baselines.