Electrocardiograms (ECGs) are critical for cardiovascular disease diagnosis, but their accuracy is often compromised by electrode motion (EM) artifacts—large, nonstationary distortions caused by patient movement and electrode-skin interface shifts. These artifacts overlap in frequency with genuine cardiac signals, rendering traditional filtering methods ineffective and increasing the risk of false alarms and misdiagnosis, particularly in wearable and ambulatory ECG applications. To address this, we propose the Feature-Shuffle Multi-Head Attention Autoencoder (FMHA-AE), a novel architecture integrating multi-head self-attention (MHSA) and a feature-shuffle mechanism to enhance ECG denoising. MHSA captures long-range temporal and spatial dependencies, while feature shuffling improves representation robustness and generalization. Experimental results show that FMHA-AE achieves an average signal-to-noise ratio (SNR) improvement of 25.34 dB and a percentage root mean square difference (PRD) of 10.29%, outperforming conventional wavelet-based and deep learning baselines. These results confirm the model’s ability to retain critical ECG morphology while effectively removing noise. FMHA-AE demonstrates strong potential for real-time ECG monitoring in mobile and clinical environments. This work contributes an efficient deep learning approach for noise-robust ECG analysis, supporting accurate cardiovascular assessment under motion-prone conditions.