Robust R-peak detection in noisy ECG using deep residual U-Net for enhanced cardiac rhythm analysis

Accurate R-peak detection in electrocardiogram (ECG) signals represents a fundamental prerequisite for cardiac rhythm analysis and arrhythmia diagnosis. Traditional signal processing approaches demonstrate limited robustness when confronted with noise artifacts, baseline drift, and morphological variations commonly encountered in clinical recordings. This study introduces a novel deep residual U-Net (ResU-Net) architecture specifically designed for robust R-peak detection in noisy ECG signals, addressing the critical challenge of maintaining high detection accuracy across diverse noise conditions and patient populations. We developed an innovative deep learning framework combining the strengths of residual networks and U-Net architecture for precise R-peak localization. The proposed ResU-Net model incorporates skip connections, multi-scale feature extraction, and attention mechanisms to enhance feature representation. The architecture was trained and evaluated using the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) arrhythmia database, INCART database, and QT database, with comprehensive noise stress testing performed using the MIT-BIH noise stress test database. Data augmentation strategies including synthetic noise injection and morphological variations were employed to enhance model generalization. The ResU-Net architecture achieved superior performance compared to traditional signal processing methods (Pan-Tompkins, Hamilton-Tompkins algorithms) and state-of-the-art deep learning approaches (1D convolutional neural network (CNN), long short-term memory (LSTM)-CNN, ResNet-18) with sensitivity of 99.76%, positive predictive value of 99.82%, and F1-score of 99.79% on the MIT-BIH arrhythmia database. Cross-database validation demonstrated exceptional generalization capabilities with sensitivity values of 99.45% (INCART) and 99.38% (QT database). Under severe noise conditions (-6 dB signal-to-noise ratio), the model maintained sensitivity above 98.2%, significantly outperforming conventional algorithms and state-of-the-art deep learning approaches. The proposed deep ResU-Net architecture establishes a new benchmark for robust R-peak detection in noisy ECG signals. The integration of residual connections with U-Net's encoder-decoder structure enables superior feature learning and noise resilience, providing a reliable foundation for automated cardiac rhythm analysis in clinical applications.