Paced ECG classification with pacing pulse-preserving mechanism and hybrid data augmentation

Abstract Accurate identification of pacemaker spikes in the ECG is crucial for assessing pacemaker function and avoiding misdiagnosis in arrhythmia analysis. This study introduces a novel denoising approach that preserves the pacing pulses. We utilized a private clinical ECG dataset along with public databases PTB-XL and PacedECGdb. A Generative Adversarial Network (GAN) was employed on the private dataset to synthesize paced signals, effectively mitigating class imbalance. An improved ResNet-18 model was trained and tested in cross-dataset scenarios. When trained on the private data and tested on PTB-XL, the model achieved an accuracy of 96.37%, an F1-score of 0.96, and an AUC of 0.99. Conversely, when trained on PTB-XL and PacedECGdb and tested on the private dataset, it attained an accuracy of 94.47%, an F1-score of 0.88, and an AUC of 0.96. The proposed method outperforms existing state-of-the-art techniques, demonstrating robust generalization and offering a reliable solution for automated paced ECG analysis in clinical practice.