Natural Language Processing in Pharmacoepidemiology: Lessons from the Multi-source Observational Safety study for Advanced Information Classification using NLP (MOSAIC-NLP)

Background: When using real-world data (RWD) from electronic health records (EHRs), much important information on confounders and outcomes is contained in the clinical notes. In the MOSAIC-NLP study funded by the U.S. FDA through the Sentinel Innovation Center, we demonstrated the feasibility of applying natural language processing (NLP) to a data set including 17+ million notes from over 100 healthcare systems to extract key information on outcomes and potential confounders.

Objectives: To describe the design and lessons learned from the pharmacoepidemiology study, MOSAIC-NLP, which is using linked EHR-claims data to assess the risk of neuropsychiatric adverse events in a cohort of patients with asthma initiating montelukast. Researchers interested in methodological considerations related to incorporating structured data from EHR linked to claims and semi/unstructured EHR data extracted from clinical notes using NLP techniques will benefit from this session.

Description: Reports of neuropsychiatric events associated with the use of the asthma drug montelukast led to FDA requiring a Boxed Warning in the montelukast labeling. Notably, while previous studies using claims data or structured EHR data did not find evidence of this association, experts suggested that the potentially transitory nature of the drug-induced neuropsychiatric effects, such as depression, anxiety, sleep abnormalities, and agitation, might not be captured in the structured medical records, but might be recorded in the clinical notes. Further, confounding variables, such as disease severity and smoking status, would similarly be captured in clinical notes only. This use case was therefore chosen to demonstrate the feasibility, value, and transportability of applying NLP to clinical notes to abstract information for these types of outcomes and the relevant confounders. In this session, we will discuss methodological considerations of a pharmacoepidemiology study in training the NLP models (Dena Jaffe) and the implications, strengths, and learnings of the resultant addition of semi/unstructured data (Bridget Balkaran). We will also discuss briefly discuss other areas where NLP could be leveraged to efficiently enrich clinical information for pharmacoepidemiology studies (Rishi Desai) and the importance and applicability of this work for future drug safety investigations (Sarah Dutcher, Jummai Apata). The session will conclude with an interactive panel discussion and question and answer session on lessons learned.