CARE-SD: classifier-based analysis for recognizing provider stigmatizing and doubt marker labels in electronic health...

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Public Health

Background:

• Electronic health records (EHRs) contain sensitive patient information.
• Identifying stigmatizing and biased language in EHRs is crucial for equitable care.
• Previous methods for detecting such language are limited.

Purpose of the Study:

• To develop and validate natural language processing (NLP) techniques for detecting stigmatizing and biased language in intensive care unit (ICU) EHRs.
• To create lexicons of stigmatizing labels, doubt markers, and scare quotes.
• To train supervised learning classifiers for automated identification of these linguistic features.

Main Methods:

• Developed lexicons using literature, Word2Vec, and GPT 3.5, refined by human evaluation.
• Searched 18 million sentences from the MIMIC-III dataset for linguistic bias features.
• Sampled, annotated by experts, and used 1000 sentences per feature for supervised learning classifiers.

Main Results:

• Created lexicons with 58 doubt marker and 127 stigmatizing label expressions.
• Achieved high performance for doubt marker and stigmatizing label classifiers (macro F1-scores of 0.84 and 0.79, respectively).
• Classifier accuracy closely matched human annotator agreement (0.87).

Conclusions:

• Demonstrated feasibility of supervised classifiers for identifying stigmatizing labels and doubt markers in medical text.
• Identified trends in stigmatizing language use within EHRs.
• Classifiers can be applied to reduce biased language in healthcare systems.