Natural language processing for scalable feature engineering and ultra-high-dimensional confounding adjustment in...

Area of Science:

• Health Informatics
• Epidemiology
• Data Science

Background:

• Healthcare database studies often struggle with confounding, particularly unmeasured factors.
• Current methods for high-dimensional proxy adjustment do not utilize free-text electronic health record (EHR) notes.
• Natural Language Processing (NLP) can extract structured features from unstructured EHR text.

Purpose of the Study:

• To evaluate the impact of adding NLP-generated features to claims data for high-dimensional proxy adjustment.
• To assess improvements in confounding control in healthcare database research.

Main Methods:

• Linked Medicare claims with EHR data to create cohorts for medication comparisons.
• Employed NLP techniques to generate structured features from free-text EHR notes.
• Utilized LASSO regression for propensity score (PS) model fitting with different covariate sets (claims-only vs. claims + NLP-EHR features).

Main Results:

• Including claims codes and NLP-generated EHR features enhanced covariate balance (standardized differences <0.1).
• Adjustment for NLP features moved effect estimates in the expected direction in two of three studies, showing nuanced impact.
• Overall covariate balance improved with the addition of NLP-derived features.

Conclusions:

• Supplementing administrative claims with NLP-generated features improves ultra-high-dimensional proxy confounder adjustment.
• This approach offers a modest benefit in capturing crucial confounder information for observational studies.
• Enhanced confounding control through NLP integration can refine healthcare database analyses.