Statistical methods and machine learning for the analysis of high-dimensional biomedical data in precision medicine

Abstract

Realising the promise of precision medicine relies on the continuous advancement of data technologies and the analytical capabilities essential for their processing. High-throughput biomedical technologies have consistently advanced in their depth and accessibility, holding promising potential for the discovery of novel biological relationships with patient clinical outcomes. Despite offering deep insights, these technologies come with several

analytical challenges that hinder their direct clinical translational impact. Addressing challenges such as high-dimensionality, accounting for heterogeneity, spatial analysis and method evaluation necessitates the development of innovative data science approaches and frameworks. To this end, this thesis contributes to the interdisciplinary field of data science and precision medicine by: (1) creating a conceptually different way to quantify cell type proportions, harmonising manual gating with unsupervised clustering techniques, to discover associations with patient outcomes – overcoming vast cell type heterogeneity contained in single-cell cytometry data; (2) suggesting several approaches for heterogeneity-aware modelling to enable more effective disease risk prediction in large patient cohorts; and (3) establishing an evaluation system to validate the clinical relevance of methods predicting spatial gene expression from histology images, providing direction for the future development of methods enhancing histology images for disease detection and treatment. The contributions outlined in this thesis aim to provide frameworks to advance precision medicine and also establish a foundation for further development of effective data science tools in biomedical research.