A reproducible benchmark of QRS detection algorithms across diverse ECG datasets and noise conditions

Accurate R-peak detection in electrocardiograms is critical for heart rate monitoring, heart rate variability analysis, and cardiac condition diagnosis. However, reliable detection remains challenging in real-world scenarios due to noise, artifacts, and signal variability. A key limitation in current research is the lack of reproducibility and comparability, as algorithms are often tested on varying datasets, hindering direct performance comparisons. To address this, we benchmark 17 R-peak detection algorithms, encompassing traditional signal processing, machine learning, and deep learning approaches, within a unified evaluation framework using five open-access ECG datasets from the PhysioNet platform. These databases represent diverse conditions, including long-term monitoring, arrhythmias, and noisy environments, enabling a standardized evaluation. Our results reveal that under a strict cross-dataset generalization setting, in which ML and DL models were trained on a single dataset without any target-domain adaptation, traditional signal processing methods provided more consistent overall performance. This highlights a trade-off between peak performance on familiar data and generalizable performance under distribution shift, whose extent for data-driven methods may depend substantially on training diversity. To support reproducibility and future benchmarking, we provide a fully open evaluation framework including all implementations, dataset references, and evaluation pipelines. These findings offer guidance for researchers and clinicians selecting R-peak detection algorithms for diverse clinical and practical scenarios.