Contextualized race and ethnicity annotations for clinical text from MIMIC-III

Area of Science:

• Clinical Informatics
• Health Equity Research
• Natural Language Processing

Background:

• Accurate race and ethnicity (RE) data is essential for observational health research, including cohort characterization, quality assessment, and identifying health disparities.
• Electronic health records (EHRs) often contain incomplete, inaccurate, or insufficiently granular patient-level RE data.
• Natural language processing (NLP) offers a method to extract RE information from unstructured clinical text to supplement existing data.

Purpose of the Study:

• To introduce the Contextualized Race and Ethnicity Annotations for Clinical Text (C-REACT) Dataset.
• To provide a valuable resource for improving the completeness and accuracy of race and ethnicity data in health research.
• To support health systems in leveraging RE data for health equity initiatives.

Main Methods:

• Developed the C-REACT Dataset using 12,000 patients and 17,281 sentences from MIMIC-III clinical notes.
• Created two sets of reference standard annotations for RE data within the clinical text.
• Provided detailed annotation guidelines for consistent and reliable data labeling.

Main Results:

• The C-REACT Dataset contains granular RE information, including preferred language and country of origin.
• A second set of annotations includes RE labels as annotated by physicians.
• The dataset facilitates the extraction of comprehensive RE data from clinical narratives.

Conclusions:

• The C-REACT Dataset enhances the utility of patient-level RE data for health research.
• This resource can significantly aid health systems in addressing health disparities and promoting health equity.
• NLP-derived RE data can bridge gaps in structured EHR data, leading to more robust health analyses.