University of Minnesota
Methods For Supervised Machine Learning and Polygenic Risk Scores
Abstract
dc:description.abstractImprovements in technology enable the collection and analysis of increasingly large datasets. This permits statistical questions related to diseases and characteristics that are complex and high-dimensional in nature - necessitating investment in methods that are capable of learning or extracting useful information from high-dimensional datasets. These application areas include genetics, genomics, and brain imaging. The over-arching theme of my dissertation is extractingknowledge from high-dimensional data. In the second chapter, we consider a polygenic risk score (PRS) approach as a method to reduce the dimensionality of genomic data. We conducted a PRS analysis using genomewide association study (GWAS) summary statistics from substance-use traits, neuroticism, and risk-taking behavior. We found that Neuroblastoma patients had a lower relative genetic predisposition to cigarette smoking, but a significantly higher predisposition toward neuroticismcompared to controls. A popular method for dimension reduction is Principal Component Analysis (PCA), and its supervised counterpart (SPCA). Next, we propose a variable selection method using a nonparametric reproducing kernel Hilbert space (RKHS) process combined with kernel-based supervised principal component analysis (KSPCA). Previous methods using non-linear kernels lack mechanisms to achieve sparsity. We apply our method in a variety of data contexts to show superior performance compared to traditional PCA and SPCA methods. In Chapter 4 we introduce a decomposition process that includes both supervision and adjustment for confounding when learning about the outcome-exposure relationship. Statistical methods that make use of RKHS may provide an effective means to reduce the effects of confounding. We propose to include multiple principal components through iterative deflation of the data matrix and the inclusion of an L₁ penalty to induce sparsity. In Chapter 5 we extend SPCA and Statistically Inspired Modification to Partial Least Squares (SIMPLS) methods to account for variable groupings. Previous L₁ penalized modifications do not account for group structures within the variables. Groups of variables may have limited individual impact on a response, but considered together, they may explain a significant proportion of variation in the response. We propose a modification to both SPCA and SIMPLS that induces group-wise sparsity and also a second L₁ penalty to induce sparsity within groups. We demonstrate the utility of our method in a variety of simulated and real-world datasets.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Anderson, Michael
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Anderson, Michael. (2024). Methods For Supervised Machine Learning and Polygenic Risk Scores. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/282292.