Evaluation of QRS morphological classifiers in the presence of noise

This paper analyzes the performance of four similarity measures (distances: d1, d2, and dinfinity, as well as correlation coefficient), when they are employed for morphological classification of QRS complexes by means of linear cluster formation. An important characteristic that any morphological classification method for QRS complexes should possess is the ability to perform waveform recognition despite the wide variety in which these could appear, as well as the diverse types of noise that could contaminate the signal. Evaluation of these classifiers constitutes an important problem for their selection. Evaluation was performed using electrocardiographic signals selected from the MIT-BIH database. These signals were contaminated with several noise types that are found in the environment where electrocardiograms are usually registered and processed, and the different noise waveforms were combined in an appropriate way to simulate practical situations, including some with severe noise contamination. Results are expressed in terms of probabilities of correct classification for different signal to noise ratios, allowing a comparison between the different distance measures in terms of their effectiveness.