Evaluation of real-time QRS detection algorithms in variable contexts

A method is presented to evaluate the detection performance of real-time QRS detection algorithms to propose a strategy for the adaptive selection of QRS detectors, in variable signal contexts. Signal contexts are defined as different combinations of QRS morphologies and clinical noise. Four QRS detectors are compared in these contexts by means of a multivariate analysis. This evaluation strategy is general and can be easily extended to a larger number of detectors. A set of morphology contexts, corresponding to eight QRS morphologies (normal, PVC, premature atrial beat, paced beat, LBBB, fusion, RBBB, junctional premature beat), was extracted from 17 standard ECG records. For each morphology context, the set of extracted beats, ranging from 30 to 23000, was resampled to generate 50 realisations of 20 concatenated beats. These realisations were then used as input to the QRS detectors, without noise, and with three different types of additive clinical noise (electrode motion artifact, muscle artifact, baseline wander) at three signal-to-noise ratios (5 dB, −5 dB, −15 dB). Performance was assessed by the number of errors, which reflected both false alarms and missed beats. The results show that the evaluated detectors are indeed complementary. For example, the Pan-Tompkins detector is the best in most contexts but the Okada detector generates fewer errors in the presence of electrode motion artifact. These results will be particularly useful to the development of a real-time system that will be able to choose the best QRS detector according to the current context.