Evaluation of real-time QRS detection algorithms in variable contexts

Overview & Synopsis

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, under variable signal contexts. Signal contexts are defined as different combinations of QRS morphologies and clinical noise. Four QRS detectors are compared under 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 8 QRS morphologies (Normal, PVC, premature atrial beat, paced beat, LBBB, fusion, RBBB, junctional premature beat), has been extracted from 17 standard ECG records. For each morphology context, the set of extracted beats, ranging from 30 to 23000, are resampled to generate 50 realizations of 20 concatenated beats. These realizations are then used as input to the QRS detectors, without noise, and with 3 different types of additive clinical noise (electrode motion artefact, muscle artefact, baseline wander) at 3 signal-to-noise ratios (5dB, -5dB, -15dB). Performance is assessed by the number of errors, which reflects both false alarms and missed beats. The results show that the evaluated detectors are indeed complementary. For example, the Pan and Tompkins's detector is the best in most contexts but the Okada's detector generates less errors in presence of electrode motion artefact. 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.