ECG heartbeat classification by means of variable rational projection

Abstract

In this paper we are concerned with ECG heartbeat classification. To this order we develop an adaptive transformation method. The transformation we use is a variable orthogonal projection that makes use of the large variability of rational functions. As a result the orthogonal system is adjusted to the signal. Consequently, not only the coefficients of the projection but also the parameters controlling the system carry significant information about the signal. An additional advantage of the adaptivity is that the dimension of the representation can be substantially reduced. We note that the key issue in the use of an adaptive transformation is the optimization according to the specific task, which is ECG heartbeat classification in our case. It is a non-linear problem and is discussed in our paper in detail. Then we construct a feature vector as a combination of dynamic and morphological descriptors. The morphological part is divided into patient-depending and individual heartbeat-depending features. In every step we provide reasoning that justifies our way to proceed. We used support vector machine algorithm for classifying the heartbeats into 5 and 16 classes. The comparison tests were performed on the MIT-BIH Arrhythmia Database. They proved that our method is better than the previous ones.