A Novel Clustering Feature-Based BiLSTM-GAN Method for Enhancing ECG Signal

Abstract

Electrocardiogram (ECG) signals are frequently corrupted by diverse noises that degrade the accuracy of downstream arrhythmia detection and hamper reliable tele-cardiology. This paper proposes a clustering feature-driven Bidirectional-LSTM Generative Adversarial Network (CF-BiLSTM-GAN) that learns to recover clean heartbeat morphology from single-channel, noisy recordings. First, an adaptive k-means module is employed on the latent space of a pre-trained denoising auto-encoder to discover representative beat-level clusters; these clusters are treated as high-level priors that encode patient-specific QRS-T shape constraints. Second, a BiLSTM generator conditioned on the cluster centroids is trained within a GAN framework to reconstruct the temporal manifold of the ECG while preserving subtle atrial/ventricular signatures. A spectral-normalized CNN discriminator enforces local smoothness and global periodicity, suppressing both high-frequency artifacts and baseline wander simultaneously. Experiments on MIT-BIH and PhysioNet/CinC datasets demonstrate that CF-BiLSTM-GAN outperforms state-of-the-art wavelet, dictionary-learning, and vanilla DL methods, achieving lower MSE, higher cross-correlation and improved beat classification F1 after denoising. The proposed approach offers a practical, data-driven solution for enhancing ECG fidelity in ambulatory monitoring and edge devices.