Using Natural Language Processing to Identify Stigmatizing Language in Labor and Birth Clinical Notes.

Area of Science:

• Computational linguistics
• Health informatics
• Medical sociology

Background:

• Racial and social biases contribute to disparities in pregnancy and birth outcomes.
• Analyzing stigmatizing language in electronic health records (EHRs) is an emerging method to identify bias.
• Automated methods are needed to efficiently detect biased language in clinical notes.

Purpose of the Study:

• To develop and evaluate automated natural language processing (NLP) methods for identifying stigmatizing language in labor and birth notes.
• To accurately detect two types of stigmatizing language: marginalizing language and power/privilege language.
• To assess the performance of machine learning algorithms in classifying these language categories.

Main Methods:

• Analysis of labor and birth notes from two hospitals for patients over 20 weeks' gestation.
• Application of text preprocessing techniques, including TF-IDF values.
• Testing of machine learning classifiers: Decision Trees, Random Forest, and Support Vector Machines.
• Utilizing InfoGain for feature importance evaluation to identify key linguistic markers.

Main Results:

• Decision Trees achieved the best classification for marginalizing language with an F-score of 0.73.
• Support Vector Machines demonstrated optimal performance for power/privilege language with an F-score of 0.91.
• The study confirmed the effectiveness of selected machine learning methods in classifying stigmatizing language.

Conclusions:

• Well-performing machine learning methods were identified for the automatic detection of stigmatizing language in clinical notes.
• This study is the first to use NLP performance metrics to evaluate machine learning for discerning stigmatizing language.
• Future research should focus on refining NLP methods and exploring advanced deep learning algorithms.