Bellfield, R (2024) Development of Machine Learning Models with Applications in Cardiovascular Research. Diploma thesis, Liverpool John Moores University.

Text 2024bellfieldphd.pdf - Published Version Restricted to Repository staff only until 11 June 2025. Available under License Creative Commons Attribution Non-commercial. Download (11MB)

Abstract

Cardiovascular disease (CVD) is one of the leading burdens on modern healthcare globally in terms of mortality, loss of health and healthcare costs. CVD covers all conditions that affect the heart and circulatory systems. Artificial intelligence (AI) and machine learning (ML) are being increasingly leveraged to help improve diagnosis, prognosis, treatment, and management of CVDs. This thesis aims to develop ML approaches that can generate novel, meaningful insights into several aspects of cardiovascular research. First, in Chapter 3 we use convolutional neural networks (CNN) to quantify the effect ECG data format has on ML predictive performance, through the clinical task of detecting myocardial infarction, providing the first results in determining the optimal ECG data format for ML modelling. The remaining analysis leverages the unsupervised, probabilistic ML technique generative topographic mapping (GTM). The analysis aims to generate 2-dimensional representations of data and propose different approaches that can identify large macro-clusters within the reduced dimension. Doing this gives an understanding of which patients/participants within a data set are clinically similar, along with interpretable visualisations that explain the rationale behind each cluster. Chapter 4 contains the first outline of this methodology, developed on a noncardiovascular dataset, to demonstrate the generalisability of such a methodology. Through this approach, we propose a novel freedom of expression index that provides an understanding of the level of restrictions placed on the population of a country. This index is defined by macro clusters generated through aggregating the normalised information contained in the GTM reference vector outputs. Chapter 5 applies this methodology to generate clinically relevant AF phenotypes for specific patient cohorts, from the general and the critical care populations. We propose a new methodological approach to achieve this that implements hierarchical clustering, again on the GTM reference vector outputs, to generate the phenotypes. Finally, Chapters 6 and 7 investigate the athlete’s heart, defined as the physiological changes that the heart undergoes due to exercising for prolonged periods. Chapter 6 contains an in-depth scoping review, evaluating the current ML applications in athlete’s heart and identifying the gaps for future research. Chapter 7 investigates features automatically extracted from ECG recordings from elite footballers, cyclists, rugby league players, and ultra runners to further the understanding of healthy athlete’s hearts. The methodology in Chapter 5 was further developed here to define a novel approach that uses magnification factors to define neighbourhoods in the 2-dimensional data representation, to carry out constrained hierarchical clustering on the reference vectors.

Item Type:	Thesis (Diploma)
Uncontrolled Keywords:	Artificial Intelligence; Machine Learning; Generative Topographic Mapping; Cardiovascular Research; Atrial Fibrillation; Athlete's Heart
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions:	Computer Science & Mathematics
SWORD Depositor:	A Symplectic
Date Deposited:	11 Jun 2024 09:59
Last Modified:	11 Jun 2024 10:00
DOI or ID number:	10.24377/LJMU.t.00023452
Supervisors:	Olier, I, Ortega-Martorell, S, Lip, G and Oxborough, D
URI:	https://researchonline.ljmu.ac.uk/id/eprint/23452

View Item