Abstract:
Cardiovascular diseases are a leading cause of death worldwide. To diagnose, assess, and
monitor these diseases clinically, electrical heart signals as they appear on the heart surface
are recorded and analyzed. These electric signals vary under different disease and drug states.
Objective computational methods for comparing and predicting cardiac signals for assisting
clinical assessments are desirable. In recent years, many automated signal decomposition
methods have been shown to be effective at representing a variety of signals from the human
body. This thesis explores the potential of Gaussian functions to decompose, represent,
analyze, and predict electrical signals from the heart.
A finite number Gaussian basis functions are fit to single beat surface electrocardiograms
(ECG) or heart electrograms (EGM) under a wide range of disease and drug states. The
proposed fitting techniques are automated and require less signal pre-processing when
compared to traditional methods. The fitted Gaussians robustly capture both heart surface
EGM and body surface ECG electrical potentials.
The Gaussian basis function parameters are linked to clinically important physiological
timing intervals such as the intervals between the Q and T waves and the J and T waves,
and the interval between the peak of T wave and its end. These relationships are explored
via data driven regression models built using recorded ECGs from various drug states. The
models enabled predictions of post-drug ECG morphology from pre-drug ECG morphology
using Gaussian basis function parameters.
An effective data driven model for predicting heart surface EGM potentials from body
surface ECG potentials is constructed using Gaussian function decomposition as a feature
extraction technique. The data driven model aims to bypass the need for invasive, risky,
and costly clinical procedures for recording heart surface potentials. The proposed model
is trained on a large number of computer generated Gaussian impulse signals. The trained
model is effective at predicting experimentally recorded heart surface potentials.
This work demonstrates that Gaussian basis signal decomposition can be foundational to
powerful and clinically relevant techniques for assessing, analyzing, and predicting electrical
signals from the heart.