Using Structured Codes and Free-Text Notes to Measure Information Complementarity in Electronic Health Records:...

Area of Science:

• Health Informatics
• Clinical Data Science
• Natural Language Processing in Healthcare

Background:

• Electronic health records (EHRs) integrate structured (e.g., diagnostic codes) and unstructured (e.g., clinical notes) data.
• Unstructured clinical narratives are presumed to offer richer patient information, but lack large-scale validation.
• Existing validation methods for EHR data completeness are limited.

Purpose of the Study:

• To quantitatively compare information content between structured and unstructured EHR data.
• To validate the hypothesis that unstructured EHR data provides more extensive patient information.
• To assess data overlap across diverse patient populations and clinical contexts.

Main Methods:

• Analysis of structured and unstructured data from a large Dutch primary care EHR database (2021-2024).
• Utilized a Dutch-tailored extraction framework to identify clinical concepts from free-text notes.
• Employed cosine similarity of concept embeddings to measure semantic similarity between structured and unstructured data.
• Quantified concept overlap across domains and patient subpopulations.

Main Results:

• Only 13% of concepts from patient records and 7% from visits had similar structured counterparts.
• Conversely, 42% of structured concepts in records and 25% in visits were found in unstructured data.
• Condition, measurement, and drug concepts showed varying overlap; subpopulations exhibited distinct data reliance patterns.

Conclusions:

• Unstructured EHR data significantly contributes additional, valuable information beyond structured data.
• The study validates the feasibility of quantifying information differences and highlights the utility of unstructured data.
• The developed methodology is versatile for robust clinical research, with annotated matches shared publicly.