Performance analysis of QRS detectors on real-life continuous rhythm monitoring surface electrocardiograms

Background The surface electrocardiogram (ECG) is routinely used in clinical practice and increasingly used for continuous rhythm monitoring (CRM) at home, driven by growing public interest in monitoring personal health. This facilitates earlier arrhythmia detection but also generates large volumes of data, requiring reliable automated analysis. Despite decades of progress in developing and improving algorithms, reliable QRS detection remains challenging due to temporal changes in beat morphology, inter-patient variability, artifacts and noise. Moreover, most QRS detection algorithms have been evaluated using the same publicly available database, MIT-BIH, containing clean and high-quality ECGs acquired in a controlled environment. These datasets do not reflect the noise, artifacts and variability that is often found in CRM recordings and thereby limits clinical generalizability. As a result, good performances are often reported in this database, but it remains challenging to accurately detect QRS complexes on real-life CRM data. Purpose To evaluate the performance of commonly used QRS detection algorithms on real-life CRM data, representing a clinical setting with substantial noise and variability. Methods The dataset comprised of 1,320 30-second ECG segments (lead II) obtained from continuous telemetry recordings after cardiac surgery. The segments contained combinations of normal sinus rhythm (SR), premature atrial complexes (PAC), premature ventricular complexes (PVC), atrial fibrillation (AF) and noise. Signal quality varied from good to poor, primarily due to movement artifacts, see Figure 1. True QRS locations were independently annotated by two researchers in all ECG segments. In case of disagreement, segments were discussed with a third reviewer. Nine commonly used QRS detection algorithms were evaluated. Algorithm performance was assessed using positive predictive value (PPV) and sensitivity. Results From all 1,320 ECGs segments, 53,137 QRS complexes were identified. The lowest PPV was found using GQRS (0.49), while the most well-known Pan-Tompkins algorithm demonstrated the lowest sensitivity (0.52). The best overall performance was obtained using the XQRS detection algorithm, with the highest PPV (0.97) and sensitivity (0.99), as shown in the top panel of Figure 2. Underlying rhythm and segment complexity had a considerable impact on algorithm performance. Although XQRS maintained consistently high performance across all rhythms, the two moving average and Pan-Tompkins algorithms, surprisingly, showed lower sensitivity in SR segments compared to AF, see bottom panel of Figure 2. Conclusion QRS detection performance varied substantially on real-life CRM data acquired in a real clinical environment containing noise and artifacts. The XQRS algorithm performed best with high PPV and sensitivity, and should therefore be preferred to detect QRS complexes in real-life CRM data.