Časoprostorová analýza vektorkardiografických záznamů

Abstract

Vectorcardiography is a method of recording the electrical activity of the heart from three orthogonal leads. These leads correspond to the projections of the electrical vector of the heart into directions X,Y and Z. So the records can be presented in the form three dimensional loops describing individual phases of the heartbeat.

In this thesis, there are introduced and discussed new features based on the morphology of QRS loops, describing ventricular depolarization. The most significant features are the velocity, curvature, minimal surface area, length and time of the QRS loops. The features are divided into octants for improving classification performance.

New feature coming from the octant theory, called Laufberger’s number was introduced. Laufberger’s number corresponds to a sequence of octants of the QRS loop and describes a pathway of the loop. The method for analyzing Laufberger’s numbers was also introduced and discussed.

All the presented features are described and analyzed with respect to the classification of myocardial infarction. The maximal sensitivity 89.9% and specificity 90.4% were achieved by using octants based QRS features and machine learning model based on random forests.

It has been shown that commonly used records-oriented cross validation causes overfitting problem and artificially increases classification performance. In this work, patient-oriented cross validation is prefered.

The important part of the thesis also describes the nonlinear transformation of the electrocardiographic leads into VCG leads. The method is based on the artificial neural network and achieves significantly higher accuracy than standard linear transformations.

A new method for individualization of the model based VCG generator is introduced. The individualization is based on the particle swarm optimization and provides high accurate models.