Using computable knowledge mined from the literature to elucidate confounders for EHR-based pharmacovigilance

Area of Science:

• Pharmacovigilance
• Biomedical Informatics
• Causal Inference

Background:

• Drug safety studies often use observational data, which is susceptible to confounding bias.
• Identifying confounders requires deep knowledge of complex biological systems, posing a challenge for reliable causal inference.
• Literature-derived computable knowledge may offer a solution for identifying potential confounders.

Purpose of the Study:

• To test if incorporating literature-derived confounders improves causal inference from observational data in drug safety research.
• To introduce and evaluate methods for searching biomedical literature for confounder candidates.

Main Methods:

• Developed and compared two methods: semantic vector-based and string-based confounder search using SemMedDB.
• Queried SemMedDB for confounders by linking drug indications (exposure) to adverse events (outcomes).
• Integrated literature-derived confounders into statistical and causal models using NLP-processed clinical notes and evaluated against known drug-adverse event relationships.

Main Results:

• Semantic vector-based search demonstrated superiority in reducing confounding bias compared to string-based search.
• The impact of the quantity of literature-derived confounders on bias reduction was inconclusive.
• Naive association measures (chi-squared, reporting odds ratio) were used for comparison.

Conclusions:

• Semantic search methods are promising for enhancing confounder identification in drug safety.
• Further research is needed to optimize the use of literature-derived confounders, especially regarding quantity.
• Consideration of targeted learning estimation and expert adjudication is recommended for complex covariate scenarios.