Abstract
The rising prevalence of type 2 diabetes (DM2) and cardiovascular disease necessitates for an improved understanding of these conditions. This thesis explores the application of machine learning applications in two areas. Epigenetic modifications associated with DM2 and segmentation of the left ventricle in echocardiography, with the aim of providing information that could improve clinical decision-making.
Elastic Net models with varying regularisation were employed to identify CpGs with differential methylation between DM2 and control liver tissue samples. The models identified CpG sites that distinguished between DM2 and non-DM2 samples and performed well in internal cross-validation. However, this accuracy was not replicated in an external dataset, where previously identified methylation patterns were absent. These findings show both the potential of machine learning in epigenetic research and the challenges of model generalisability across different datasets.
Deep learning models using nnU-Net were configured to segment the left ventricle (LV) and the left ventricle walls (LVW) from three echocardiogram views, the parasternal long axis, the parasternal short axis, and the apical four-chamber. These models performed consistently well in segmenting the LV, but all exhibited reduced performance in segmenting the LVW. This work expanded to measuring the left ventricular posterior wall (LVPW) from model-generated segmentation masks. Two novel measurement methods were developed for the task ("Geometric" and "Line Analysis" methods). The measurement methods proved suboptimal, with discrepancies between the model-generated mask measurements and the ground truth clinical measurements. These results were driven by the model’s tendency to overestimate the LVPW size and the dependence of the measurement methods on the mask geometry. These results provide important information in the challenges associated with automating left ventricular quantification and highlight areas for future measurement improvement. This thesis demonstrates the potential of machine learning techniques in clinical research, while acknowledging the complexities involved. Future work will address the limitations encountered in model performance and further explore the integration of epigenetic and imaging data to improve disease diagnosis and management.