On the use of auto-regressive modeling for arrhythmia detection

This paper investigates the use of an auto-regressive modeling method for the classification of heartbeats into two categories: Normal beats (N) and Ventricular ectopic beats (VEB). The method is based on an auto-regressive modeling (AR) of QRS complexes. Each heartbeat is characterized by its AR coefficients. Then, K-nearest neighbor (K-NN) classifier uses the AR coefficients to discriminate between N beats and VEB. In addition, the use of AR modeling prediction error en as a discriminating feature is investigated. Results show that the prediction error power (σ2p) enhances significantly the classification accuracy. The proposed classifier is compared to a classifier based on the use of RR timing information. Finally, the two classifiers are combined together where the classification result is given by the agreement of the two classifiers. The proposed AR modeling approach performs better than the RR interval-based classifier and their combination enhances the classification accuracy.