QRS Detection using Morphological and Rhythm information

An approach has been developed using artificial neural networks to detect QRS complexes within an ambulatory ECG signal. The method employs the use of an artificial neural network classifier to recognise the morphology of a QRS complex based on amplitude and derivative features. The feature vectors are derived from a representative annotated ECG trace and are used in the formulation of the ANN'S training set. The outputs, or p.d.f. estimates generated by the neural network are then used to determined if a "QRS-like spike" has occurred. These spike detections then undergo further post-processing which, biases these detections such that the spike detection "nearest" the anticipated location of the next QRS is confirmed as a QRS complex. This anticipation of the QRS complex location is based on the estimation of the next RR interval using past RR intervals of previously confirmed QRS complexes. Such post-processing has the effect of greatly reducing the number of false positive detections, particularly in noisy ECG traces. Vector Quantisation Introduction QRS detection is a well known problem of which many approachs to its solution have been proposed; [ 1][2] to name a few. Its represents the first stage of ECG analysis and its use in ECG monitoring equipment is now not uncommon. In this paper, we employ a decision theoretic pattern recognition approach to QRS detection incorporating both morphological and rhythm information in the process. Outline of Approach The outline of the QRS detection approach investigated in this paper is functionally represented in the figure 1. This approach assumes the ECG signals acquired undergo basic signal conditioning prior to digitisation; ie. the use of an anti-aliasing fiiter at least. It also encourages as much signal conditioning be done to clean-up the ECG signal before feature extraction; some such methodologies were reviewed in [ 11. Pre-Processing Digitisation CO nd itio ning