Comparative ranking of marginal confounding impact of natural language processing-derived versus structured features in pharmacoepidemiology

Open Webpage

Joseph M. Plasek, Richard Wyss, Janick Weberpals, Jie Yang, Thomas DeRamus, Theodore N. Tsacogianis, Kerry Ngan, Lily G. Bessette, Kueiyu Joshua Lin, Li Zhou

Published: 2025, Last Modified: 15 Oct 2025Comput. Biol. Medicine 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Unstructured clinical notes contain potential confounder information.•Natural language processing can auto-generate features from notes.•We ranked the marginal confounding impact of features on estimated causal effects.•Prespecified knowledge and claims data dominated the top 25 ranked features.•39 % of the top 100 ranked features were natural language processing generated.