QRS delineation algorithms comparison and model fine tuning for automatic clinical classification

QRS complex extraction and wave detection have been paid intense efforts in scientific publications in the last two decades. This work elaborates on different QRS delineation algorithms for classification and for diagnostic indexing. A subset of 150 cases were randomly selected from a full database including over 3500 consecutive ECGs. Three QRS detection methods were implemented and later benchmarked against the information provided by a GE MAC 5000 ECG system, and also against a Gold-Standard manually and carefully developed by clinicians. All implemented methods were applied to the complete ECG signal and to a consolidated single ECG beat template. Better performance was obtained using beat template signals, due to denoising effects. The absolute error of the QRS duration was chosen as a figure of merit. Results showed that all developed methods outperformed information provided by the ECG device, when compared to Gold-Standard: 7,90 ± 6,83 ms of QRS duration for Physionet Method, 8,31 ± 3,07 ms for Chouhan Method, 6,27 ± 4,77 ms for an add-hoc two stages developed method, and 8,63 ± 5,89 ms for the GE device. Individual methods very much rely on one single measurement that does not easily match clinician's criteria. A two stage strategy, with first a initial candidates pre-selection, overcoming ECG local singularities, following by a fist and second momentum analysis, provided a better fit to Gold-Standard.