Electronic Health Record-Derived Phenotyping Models to Improve Genomic Research in Stroke

Stroke is a highly heterogeneous and complex disease that is a leading cause of death in the United States. The landscape of risk factors for stroke is vast, and its large genetic burden has yet to be fully discovered. We hypothesize that the small number of stroke variants recovered so far is due to 1) the vast phenotypic heterogeneity of stroke and 2) binary labeling of stroke genome-wide association study (GWAS) participants as cases or controls. Specifically, genome-wide association studies accumulate hundreds of thousands to millions of participants to acquire adequate signal for variant discovery. This requires time-consuming manual curation of cases and controls often involving large-scale collaborations. Genetic biobanks connected to electronic health records (EHR) can facilitate these studies by using data routinely captured during clinical care like billing diagnosis codes. These data, however, do not define adjudicated cases and controls, with many patients falling somewhere in between. There is an opportunity to use machine learning to add nuance to these definitions. We hypothesize that an expanded definition of disease by incorporating correlated diseases and risk factors from EHR data will improve GWAS power. We also hypothesize that granularly subtyping stroke using unsupervised learning methods can provide insight into stroke etiology and heterogeneity. In Chapter 1, we described the motivation for building upon current phenotyping methods for subtyping and genome-wide association studies to improve GWAS power. In Chapter 2, using patients from Columbia-New York Presbyterian (NYP) Hospital, we built and evaluated machine learning models to identify patients with acute ischemic stroke based on 75 different case-control and classifier combinations. In chapter 3, we compared two data-driven and unsupervised methods, non-negative matrix factorization (NMF) and Hierarchical Poisson Factorization, to subtype stroke patients and determined whether any of the subtypes correlate to stroke severity. In chapter 4, we estimated the heritability of acute ischemic stroke by treating the patient probabilities assigned by the machine learning phenotyping models for acute ischemic stroke in chapter 2 as a quantitative trait and mapping the probabilities to Columbia-NYP EHR-generated pedigrees. We also applied our machine learning phenotyping algorithm method, which we call QTPhenProxy, to venous thromboembolism on Columbia eMERGE Consortium patients and ran a genome-wide association study using the model probabilities as a quantitative trait. Finally, we applied QTPhenProxy to subjects in the UK Biobank for stroke and 14 other diseases and ran genome-wide association studies for each disease. We found that our machine-learned models performed well in identifying acute ischemic stroke patients in the Columbia-NYP EHR and in the UK Biobank. We also found some NMF-derived subtypes that were significantly correlated with stroke severity. We were underpowered in the eMERGE venous thromboembolism cohort GWAS and did not recover any known or new variants. Finally, we found that QTPhenProxy improved the power of GWAS of stroke and several subtypes in the UK Biobank, recovered known variants, and discovered a new variant that replicates in a previous stroke GWAS. Our results for QTPhenProxy demonstrate the promise of incorporating large but messy sets of data, such as the electronic health record, to improve signal in genome-wide association studies