Natural language processing for scalable feature engineering and ultra-high-dimensional confounding adjustment in healthcare database studies

Open Webpage

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

Published: 2025, Last Modified: 15 Oct 2025J. Biomed. Informatics 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: BackgroundTo improve confounding control in healthcare database studies, data-driven algorithms may empirically identify and adjust for large numbers of pre-exposure variables that indirectly capture information on unmeasured confounding factors (‘proxy’ confounders). Current approaches for high-dimensional proxy adjustment do not leverage free-text notes from electronic health records (EHRs). Unsupervised natural language processing (NLP) technology can scale to generate large numbers of structured features from unstructured notes.ObjectiveTo assess the impact of supplementing claims data analyses with large numbers of NLP generated features for high-dimensional proxy adjustment.MethodsWe linked Medicare claims with EHR data to generate three cohorts comparing different classes of medications on the 6-month risk of cardiovascular outcomes. We used various NLP methods to generate structured features from free-text EHR notes and used least absolute shrinkage and selection operator (LASSO) regression to fit several propensity score (PS) models that included different covariate sets as candidate predictors. Covariate sets included features generated from claims data only, and claims data plus NLP-generated EHR features.ResultsIncluding both claims codes and NLP-generated EHR features as candidate predictors improved overall covariate balance with standardized differences being < 0.1 for all variables. While overall balance improved, the impact on estimated treatment effects was more nuanced with adjustment for NLP-generated features moving effect estimates further in the expected direction in two of the empirical studies but had no impact on the third study.ConclusionSupplementing administrative claims with large numbers of NLP-generated features for ultra-high-dimensional proxy confounder adjustment improved overall covariate balance and may provide a modest benefit in terms of capturing confounder information.