Analysis of Ventricular Depolarisation and Repolarisation Using Registration and Machine Learning

Abstract

Our understanding of cardiac diseases has greatly advanced since the advent of electrocardiography (ECG). With the increasing influx of available data in recent times, significant research efforts have been put forth to automate the study and detection of cardiac conditions. Naturally, the focus has progressed toward studying dynamic changes in ventricular depolarisation and repolarisation across serial recordings - as complex beat-to-beat changes in morphology manifest over time. Manual extraction of diagnostic and prognostic markers is a laborious task. Hence, automated and accurate methods are required to extract markers for the study of ventricular lability and detection of common diseases, such as myocardial ischemia and myocardial infarction. The aim of this thesis is to improve automated marker extraction and detection of diseases for the study of ventricular depolarisation and repolarisation lability in ECG. As such, two novel template adaptation methods capable of capturing complex beat-to-beat morphological changes are proposed for three-dimensional and two-dimensional data, respectively. The proposed three-dimensional template adaptation method provides an inhomogeneous method for transforming template vectorcardiogram (VCG) by exploiting registrationinspired parametrisation and an efficient kernel ridge regression formulation. Analysis across simulated data and clinical myocardial infarction data demonstrates state-of-the-art results. The two-dimensional template adaptation method draws from traditional registrationbased techniques and treats the ECG as a two-dimensional point set problem. Validation against previously employed simulated data and a gold-standard annotated clinical database demonstrate the highest level of performance. Subsequently, frameworks employing the proposed template adaptation techniques are developed for the automated detection of ischemic beats and myocardial infarction. Furthermore, a small study analysing ventricular repolarisation variability (VRV) in non-ischemic cardiomyopathy (CM) is considered, utilising markers of cardiac lability proposed in the development of the three-dimensional template adaptation system. In summary, this thesis highlights the necessity for custom template adaptation methods for the accurate measurement of beat-to-beat variability in cardiac data. Two novel stateof- the-art methods are proposed and extended to study myocardial ischemia, myocardial infarction and non-ischemic CM.