Systemic Networks for High-Dimensional Exposures and Health Outcomes

Developing methods for identifying associations in high-dimensional data and evaluating the methods to detect these in both statistical simulations and real data applications is an area of growing importance in many domains of biomedical science. Network analysis is becoming increasingly recognized as a vital tool for analyzing high-dimensional biomedical data in order to: 1) understand the complex interaction of factors in a single dataset, 2) enable integration of heterogeneous datasets in order to elucidate the impact of factors from one dataset on the other, and 3) predict outcomes based on our understanding of complex structures of variables within datasets. In this thesis, I developed statistical methods to address these issues and tested them through simulations of generalized cases and applied them to real data from New Hampshire Birth Cohort Study. Thus, the goals of the thesis were to: (1) develop a unsupervised network method that better identifies patterns of dependency and association of features in the data and apply this method to concentrations of 24 trace elements measured in the human placentas, (2) detect the impact of placental trace elements on the interacting network of cord-blood metabolites by developing a data integration method, and (3) predict birthweight using complex structures of elemental and metabolomic data by comparing prediction methods with different grouping strategies. The methods developed here can be applied to analyze complex mechanism of other similarly structured biomedical data (e,g, environmental exposures or metabolomes)