Robust R-peak detection in ECG signals using mamba-enhanced 1D diffusion model

The R-peak detector is a critical measurement tooll for analyzing electrocardiogram (ECG) signals, as it directly impacts downstream tasks like Heart Rate Variability (HRV) assessment and Cardiovascular Disease (CVD) diagnosis. Current detectors perform well with high-quality signals, but struggle with low-quality, noisy, and imbalanced data collected from mobile sensors, such as Holter monitors, and frequently lack robustness across diverse datasets. To alleviate this problem, we propose DiffM-R, a novel R-peak detector based on a mamba-enhanced 1D diffusion model. The DiffM-R extends the conventional diffusion model into a 1D pattern specifically designed for ECG signals, and further combined with the Mamba structure, a sequence model based on State Space Models (SSM), to enhance the model’s inference performance. Importantly, DiffM-R integrates the diffusion model into the Unet architecture, leveraging its strong semantic representation capabilities, while Mamba’s long-range dependency modeling enhances stability. Experiments are conducted on the China Physiological Signal Challenge database (CPSC-DB), which comprises low-quality Holter ECG signals with over one million beats. The evaluation on CPSC-DB achieves a 99.41% F1-score, 99.54% recall, and 99.29% precision using raw data without augmentation, marking the best R-peak detection performance. It also significantly reduces missed detections of abnormal heartbeats. On other benchmark datasets, including the Fantasia Database (Fantasia-DB), the MIT-BIH Arrhythmia Database (MITBIH-DB), and the Long-Term ST Database (LTST-DB), our model outperforms leading baseline methods. To the best of our knowledge, this work is the first diffusion-based method specifically proposed for ECG R-peak detection, providing a new measurement path for accurate ECG signal analysis, which holds significant importance in the diagnosis of CVDs.