Scalable Feature Engineering from Electronic Free Text Notes to Supplement Confounding Adjustment of Claims-Based...

Area of Science:

• Health Informatics
• Pharmacoepidemiology
• Computational Linguistics

Background:

• Electronic health records (EHRs) contain valuable free-text notes (FTNs).
• Natural language processing (NLP) can extract data from FTNs for research.
• Current NLP applications in pharmacoepidemiology face scalability challenges.

Purpose of the Study:

• To assess the utility of unsupervised NLP-generated features from FTNs.
• To improve prediction of drug exposure and clinical outcomes in pharmacoepidemiologic studies.
• To compare NLP-enhanced models against claims-based analyses for confounding adjustment.

Main Methods:

• Linked Medicare claims with EHR data for three cohort studies.
• Employed unsupervised NLP with a "bag-of-words" approach on FTNs.
• Compared machine learning models using researcher-specified variables, claims codes, and NLP features.

Main Results:

• NLP-generated features significantly improved prediction of treatment choice when combined with claims data.
• Models incorporating NLP features (Set4) outperformed those using only claims codes (Set2) or NLP features (Set3).
• Predicting clinical outcome risk showed minimal improvement beyond models using only researcher-specified variables (Set1).

Conclusions:

• Supplementing claims data with NLP features from EHR FTNs enhances prediction of prescribing choices.
• NLP features offer limited improvement for clinical risk prediction in this context.
• Findings inform strategies for leveraging EHR data to mitigate confounding in pharmacoepidemiologic research.