Network analysis and visualization for electronic health records data
Electronic health records (EHRs) provide unprecedented amounts of data, facilitating entirely new kinds of observational research studies. One valuable and challenging area of research enabled by these databases is characterizing and exploring clinical phenotypes as represented by billing codes, asking questions such as how phenotypes are associated with biomarkers or each other. Analyses involving EHR derived phenotypes have seen great success but typically operate under assumptions of independence between the phenotypes themselves, a considerable simplification of the intricate relational structure that drives the human phenome. By acknowledging and incorporating the phenome's network structure into analyses, we can both enhance existing techniques and develop new methods for exploring this high-dimensional space. In this dissertation, we illustrate this with three products. PhewasME is an application that augments Phenome-Wide Association Studies by displaying typically provided tables and plots alongside with the underlying patient-phenotype network. Next, AssociationSubgraphs is a new algorithm and visualization for examining high-dimensional association patterns, such as the 1,815-dimensional phenome encoded by PheCodes. Finally, we demonstrate the use of network statistics to characterize topological conservation between two comorbidity networks derived from independent EHRs.