Improvement of QRS boundary recognition by means of unsupervised learning

Most of the ECG wave boundaries detection algorithms are based on the matching of an one-dimensional detection function against a standard template computed from an expert controlled reference data set. In this paper, we propose to enhance the method by first stratifying the shapes of the detection functions in the vicinity of the waveform boundaries into K shape specific classes Cj (i=1,K) by means of a Kohonen self-organizing neural network. We then compute a matching template for each category Cj and we extend the standard wave delineation algorithm to take account of these new templates. The method has been assessed on the CSE databases DS1 and DS3 for the determination of the onset of QRS.