High-Throughput Analysis of Electronic Health Records to Generate Reproducible and Interpretable Biological Insights
Abstract
Electronic health records (EHRs) offer unprecedented opportunities to study human health at scale, yet their use in biomedical research is challenged by heterogeneous data capture, incomplete phenotyping, and threats to reproducibility and interpretability. This thesis develops high-throughput computational frameworks to address these challenges and generate clinically meaningful insights from real-world EHR data, guided by three core principles: reproducibility, scalability, and interpretability. The work focuses on two application domains: medication safety during pregnancy and interpretation of rare genetic variants. First, this thesis applies a target trial emulation framework to evaluate the effects of continuing selective serotonin reuptake inhibitors (SSRIs) during pregnancy using linked maternal–infant EHR data. By explicitly aligning observational analyses with a hypothetical randomized trial, this approach improves causal interpretability and reduces bias. Results show modest associations between prenatal SSRI exposure and indicators of impaired neonatal adaptation, such as lower Apgar scores, with no evidence of increased risk for major congenital malformations or life-threatening neonatal outcomes. Second, this work advances the Phenotype Risk Score (PheRS) framework by introducing age-aware modeling that accounts for the timing of phenotypic manifestation. Using EHR-linked genetic data from the eMERGE network, these models outperform the original PheRS in distinguishing pathogenic from benign variants. An extension leveraging the absence of disease further enables detection of potentially protective variants. Finally, this thesis addresses reproducibility and scalability through the development of the phers R package, a modular, open-source implementation of PheRS-based analyses. By standardizing phenotype mappings, modeling strategies, and positive control validation, the package enables transparent and transferable rare variant analyses across diseases and cohorts. Together, these contributions demonstrate how principled study design, biologically informed modeling, and reproducible software can transform EHR data into a reliable resource for translational biomedical research.