Date of Award

Spring 1993

Document Type

Thesis - Restricted

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering

First Advisor

Ropella, K. M.

Second Advisor

Mykelbust, J. B.

Third Advisor

Merrill, S. J.

Abstract

The problem of automatically analyzing time series associated with cardiac rhythms to infer physiologic state is a difficult one. In general, such time series contain some information characteristic of cardiac state along with a great deal of superfluous information. Moreover, how the information of interest is manifested in time series varies both within and among individuals. Consequently, it is not sufficient to extract an attribute of a signal which only partially describes variability. Instead, the whole signal must be examined. Standard decompositions of signal variability such as auto- and cross- spectra reveal that variance is characteristically distributed in frequency for different cardiac arrythmias. To obtain a normalized measure of covariance, a cross-spectrum must be normalized by its constituent autospectra. Averaging the resulting coherence function provides a normalized value which automatically separates electrogams according to rhythm. Unfortunately, estimation of spectral covariance requires two simultaneous time series. Consequently, coherence is generally region-inspecific. Moreover, coherence is made worthless on the body surface by the high correlation among measured potentials. Finally, although coherence is effective, it presumes to model time series as the outputs of linear systems. Time series corresponding to nonfibrillatory rhythms generally exhibit discrete complexes, which are brief assymetric bursts disrupting an otherwise stable baseline. Burst phenomena and time-irreversibility are two well-known indicators of the presence of nonlinear systems. Perhaps the simplest nonlinear systems are bilinear systems. Indeed, such systems are capable of producing time-irreversible burst phenomena. Moreover, the corresponding bi coherence function in the frequency domain allows covariability to be examined within a single time series. As such, the problems of nonlinearity and spatial localization have been addressed. Furthermore, since only a single time series is required, autobicoherence can be calculated from a surface ECG lead. Consequently, the self-similarity of surface ECG allows ventricular rhythms to be distinguished. The gist of this thesis is that average autobicoherence can discriminate rhythms in the same way that coherence does. Although it is more difficult to compute, only a single time series is required for rhythm discrimination. In addition, spatial localization is achieved and the nonlinear properties of the time series are acknowledged and utilized. Finally, many of the benefits of coherence can now be exploited using surface ECG. In short, the usefulness of coherence analysis of cardiac signals has been substantially enhanced. How autobicoherence analyses of surface ECG compare to more conventional discrimination methods remains to be seen, but the results of this study indicate that the future of autobicoherence looks bright.

Share

COinS

Restricted Access Item

Having trouble?