Predicting cardiac functional and structural abnormalities from electrogram morphology using supervised machine learning

Abstract

The extracellular contact electrogram, which is the signature of the interaction of electrical activation and architecture of the local myocardium, is recorded clinically in contact with myocardium. The morphology of the signal could show relationships between the local electrogram and conduction abnormalities that influence the electrophysiology. In this thesis, I sought to address the hypothesis that the local electro-architecture, which is responsible for identifiable features of local action potentials, can be predicted from specific characteristics of electrogram recordings using supervised machine learning algorithms.

In addressing this hypothesis, I utilised in vitro multicellular preparations for obtaining unipolar electrogram data. The recordings were collected under a variety of experimental conditions, in order to investigate the effects of functional abnormalities, such as ion channel blockade and gap junction uncoupling, as well as structural determinants, such as increasing amounts of fibroblasts co-cultured with cardiac myocytes. A signal processing and feature extraction process was developed and applied on electrograms. The relationships between the abnormalities, which were introduced to experimental models, and specific electrogram characteristics were then investigated. Electrograms were then used inversely for the development of prediction models. To demonstrate the translational potential of these tools, they were tested on tissue slices derived from human end-stage heart failure hearts. It was found that EGM morphology was significantly modified due to the different heart failure phenotypes. These differences in morphology allowed accurate predictions. Paced data were also obtained from patients with a history of persistent AF.

The functional and structural determinants of unipolar electrogram morphology, which are also responsible for a variety of cardiac arrhythmias, can be predicted accurately using supervised machine learning. By better understanding the role of electro-architecture on electrogram morphology and utilising machine learning, we are provided with new insights that could contribute to a progress in diagnostics and treatment of cardiac diseases.