Statistical Analysis and Modeling for Biomedical Applications

This dissertation discusses approaches to two different applied statistical challenges arising from the fields of genomics and biomedical research. The first takes advantage of the richness of whole genome sequencing data, which can uncover both regions of chromosomal aberration and highly specific information on point mutations. We propose a method to reconstruct parts of a tumor's history of chromosomal aberration using only data from a single time-point. We provide an application of the method, which was the first of its kind, to data from eight patients with squamous cell skin cancer, in which we were able to find that knockout of the tumor suppressor gene TP53 occur early in that cancer type. While the first chapter highlights what's possible with a deep analysis of data from a single patient, the second chapter of this dissertation looks at the opposite situation, aggregating data from several patients to identify gene expression signals for disease phenotypes. In this chapter, we provide a method for hierarchical multilabel classification from several separate classifiers for each node in the hierarchy. The first calls produced by our method improve upon the state-of-the-art, resulting in better performance in the early part of the precision-recall curve. We apply the method to disease classifiers constructed from public microarray data, and whose relationships to each other are given in a known medical hierarchy