Leveraging enhanced phenotyping and causal inference approaches to investigate the effect of maternal medication exposures on child outcomes

PROJECT SUMMARY Nine out of ten women in the US take at least one medication during pregnancy. However, due to ethical and practical constraints, pregnant women are typically excluded from clinical trials, resulting in a lack of data on medication safety during pregnancy and potential impact on the child. Epidemiological studies are increasingly using real-world data (RWD), including electronic medical records (EMRs) to study treatment effects during pregnancy, but these studies have relied primarily on structured data. This can miss additional context and information that can be gained from unstructured data sources (i.e. clinical notes), potentially resulting in limited analytic precision and potential bias. To investigate how maternal medication history affects child health outcomes, we will use both structured and unstructured data from a large repository of EMRs of mothers and their children. Specifically, in Aim 1, we will build a registry of mother-child pairs containing high-quality phenotype data, which we will extract from EMRs using large language models (LLMs). We will use a paired mother-child data mart that will include all individuals who were pregnant between 2012 and the end of the study, based on the clinical records at UCM, as well as the clinical records of their children. We will utilize LLMs to enhance our ability to extract additional medication exposures and phenotypes from clinical notes. In Aim 2, we will identify robust correlations between maternal medication exposures and early life health outcomes. We will perform a broad binary statistical association of maternal medication exposures to outcomes using a medication-wide association study (MedWAS), or modified PheWAS, approach. We will then perform rules-based refinement of select exposures and outcomes, which will include the application and development of exposure ascertainment and phenotype algorithms. In Aim 3, we will use causal inference models to investigate the effects of maternal medication exposures on early life health outcomes. For a subset of associations identified in Aim 2, as well as associations established in the literature, we will utilize causal inference machine learning approaches to estimate both binary and continuous (e.g., dose-dependent), individualized treatment effects. Overall, this work will provide insights into the safety and risks of medication exposures during pregnancy, which may inform data-driven treatment strategies for pregnant patients. At the same time, this work will provide me abundant opportunities to grow my computational research skills while benefitting from the support of diverse collaborators and mentors spanning clinical, informatics, and genomics expertise.